bench : make n_batch and n_ubatch configurable in Batched bench (#6500)

* bench: make n_batch and n_ubatch configurable

* bench: update doc for batched bench

.devops/full-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -28,6 +28,8 @@ COPY . .
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable CUDA
 ENV LLAMA_CUDA=1
+# Enable cURL
+ENV LLAMA_CURL=1
 
 RUN make
 

.devops/full-rocm.Dockerfile

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

.devops/full.Dockerfile

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

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

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

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

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

.devops/llama-cpp.srpm.spec

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

.devops/nix/package.nix

@@ -24,7 +24,7 @@
     useOpenCL
     useRocm
     useVulkan
-  ],
+  ] && blas.meta.available,
   useCuda ? config.cudaSupport,
   useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
   useMpi ? false, # Increases the runtime closure size by ~700M
@@ -67,10 +67,15 @@ let
     strings.optionalString (suffices != [ ])
       ", accelerated with ${strings.concatStringsSep ", " suffices}";
 
+  executableSuffix = effectiveStdenv.hostPlatform.extensions.executable;
+
   # TODO: package the Python in this repository in a Nix-like way.
   # It'd be nice to migrate to buildPythonPackage, as well as ensure this repo
   # is PEP 517-compatible, and ensure the correct .dist-info is generated.
   # https://peps.python.org/pep-0517/
+  #
+  # TODO: Package up each Python script or service appropriately, by making
+  # them into "entrypoints"
   llama-python = python3.withPackages (
     ps: [
       ps.numpy
@@ -159,11 +164,6 @@ effectiveStdenv.mkDerivation (
         --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\";"
-
-      # 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"
     '';
 
     # With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015,
@@ -244,8 +244,8 @@ effectiveStdenv.mkDerivation (
     # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
     # if they haven't been added yet.
     postInstall = ''
-      mv $out/bin/main $out/bin/llama
-      mv $out/bin/server $out/bin/llama-server
+      mv $out/bin/main${executableSuffix} $out/bin/llama${executableSuffix}
+      mv $out/bin/server${executableSuffix} $out/bin/llama-server${executableSuffix}
       mkdir -p $out/include
       cp $src/llama.h $out/include/
     '';

.devops/server-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential git
+    apt-get install -y build-essential git libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -22,11 +22,16 @@ COPY . .
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable CUDA
 ENV LLAMA_CUDA=1
+# Enable cURL
+ENV LLAMA_CURL=1
 
 RUN make
 
 FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
 COPY --from=build /app/server /server
 
 ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

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

.devops/server-rocm.Dockerfile

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

.devops/server-vulkan.Dockerfile

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

.devops/server.Dockerfile

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

.github/workflows/bench.yml

@@ -0,0 +1,294 @@
+# Benchmark
+name: Benchmark
+
+on:
+  workflow_dispatch:
+    inputs:
+      gpu-series:
+        description: 'Azure GPU series to run with'
+        required: true
+        type: choice
+        options:
+          - Standard_NC4as_T4_v3
+          - Standard_NC24ads_A100_v4
+          - Standard_NC80adis_H100_v5
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
+      duration:
+        description: 'Duration of the bench'
+        type: string
+        default: 10m
+
+  push:
+    branches:
+      - master
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+  pull_request_target:
+    types: [opened, synchronize, reopened]
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+  schedule:
+    -  cron: '04 2 * * *'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}-${{ github.event.inputs.sha }}
+  cancel-in-progress: true
+
+jobs:
+  bench-server-baseline:
+    runs-on: Standard_NC4as_T4_v3
+    env:
+      RUNNER_LABEL: Standard_NC4as_T4_v3 # FIXME Do not find a way to not duplicate it
+      N_USERS: 8
+      DURATION: 10m
+
+    strategy:
+      matrix:
+        model: [phi-2]
+        ftype: [q4_0, q8_0, f16]
+        include:
+          - model: phi-2
+            ftype: q4_0
+            pr_comment_enabled: "true"
+
+    if: ${{ github.event.inputs.gpu-series == 'Standard_NC4as_T4_v3' || github.event.schedule || github.event.pull_request || github.head_ref == 'master' || github.ref_name == 'master' || github.event.push.ref == 'refs/heads/master' }}
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Install python env
+        id: pipenv
+        run: |
+          cd examples/server/bench
+          python3 -m venv venv
+          source venv/bin/activate
+          pip install -r requirements.txt
+
+      - name: Prometheus
+        id: install_prometheus
+        run: |
+          wget --quiet https://github.com/prometheus/prometheus/releases/download/v2.51.0/prometheus-2.51.0.linux-amd64.tar.gz
+          tar xzf prometheus*.tar.gz --strip-components=1
+          ./prometheus --config.file=examples/server/bench/prometheus.yml &
+          while ! nc -z localhost 9090; do
+            sleep 0.1
+          done
+
+      - name: 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 ${{ matrix.model }}/ggml-model-${{ matrix.ftype }}.gguf \
+              --model-path-prefix /models \
+              --parallel ${{ env.N_USERS }} \
+              -ngl 33 \
+              --batch-size 2048 \
+              --ubatch-size	256 \
+              --ctx-size 16384 \
+              --n-prompts 1000 \
+              --max-prompt-tokens 1024 \
+              --max-tokens 2048
+
+          cat results.github.env >> $GITHUB_ENV
+
+          # Remove dataset as we do not want it in the artefact
+          rm ShareGPT_V3_unfiltered_cleaned_split.json
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
+          compression-level: 9
+          path: |
+            examples/server/bench/*.jpg
+            examples/server/bench/*.json
+            examples/server/bench/*.log
+
+      - name: Commit status
+        uses: Sibz/github-status-action@v1
+        with:
+          authToken: ${{secrets.GITHUB_TOKEN}}
+          sha: ${{ inputs.sha || github.event.pull_request.head.sha || github.sha }}
+          context: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
+          description: |
+            ${{ env.BENCH_RESULTS }}
+          state: 'success'
+
+      - name: Upload benchmark images
+        uses: devicons/public-upload-to-imgur@v2.2.2
+        continue-on-error: true # Important as it looks unstable: 503
+        id: imgur_step
+        with:
+          client_id: ${{secrets.IMGUR_CLIENT_ID}}
+          path: |
+            examples/server/bench/prompt_tokens_seconds.jpg
+            examples/server/bench/predicted_tokens_seconds.jpg
+            examples/server/bench/kv_cache_usage_ratio.jpg
+            examples/server/bench/requests_processing.jpg
+
+      - name: Extract mermaid
+        id: set_mermaid
+        run: |
+          set -eux
+
+          cd examples/server/bench
+          PROMPT_TOKENS_SECONDS=$(cat prompt_tokens_seconds.mermaid)
+          echo "PROMPT_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
+          echo "$PROMPT_TOKENS_SECONDS" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          PREDICTED_TOKENS_SECONDS=$(cat predicted_tokens_seconds.mermaid)
+          echo "PREDICTED_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
+          echo "$PREDICTED_TOKENS_SECONDS" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          KV_CACHE_USAGE_RATIO=$(cat kv_cache_usage_ratio.mermaid)
+          echo "KV_CACHE_USAGE_RATIO<<EOF" >> $GITHUB_ENV
+          echo "$KV_CACHE_USAGE_RATIO" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          REQUESTS_PROCESSING=$(cat requests_processing.mermaid)
+          echo "REQUESTS_PROCESSING<<EOF" >> $GITHUB_ENV
+          echo "$REQUESTS_PROCESSING" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+      - name: Extract image url
+        id: extract_image_url
+        continue-on-error: true
+        run: |
+          set -eux
+
+          echo "IMAGE_O=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[0] }}" >> $GITHUB_ENV
+          echo "IMAGE_1=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[1] }}" >> $GITHUB_ENV
+          echo "IMAGE_2=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[2] }}" >> $GITHUB_ENV
+          echo "IMAGE_3=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[3] }}" >> $GITHUB_ENV
+
+      - name: Comment PR
+        uses: mshick/add-pr-comment@v2
+        id: comment_pr
+        if: ${{ github.event.pull_request != '' && matrix.pr_comment_enabled == 'true' }}
+        with:
+          message-id: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
+          message: |
+            <p align="center">
+
+            📈 **llama.cpp server** for _${{ github.job }}_ on _${{ env.RUNNER_LABEL }}_ for `${{ matrix.model }}`-`${{ matrix.ftype }}`: **${{ env.BENCH_ITERATIONS}} iterations** 🚀
+
+            </p>
+
+            <details>
+
+            <summary>Expand details for performance related PR only</summary>
+
+            - Concurrent users: ${{ env.N_USERS }}, duration: ${{ github.event.inputs.duration || env.DURATION }}
+            - HTTP request          : avg=${{ env.HTTP_REQ_DURATION_AVG }}ms        p(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**
+            - ${{ env.BENCH_GRAPH_XLABEL }}
+
+
+            <p align="center">
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_O }}" alt="prompt_tokens_seconds" />
+
+            <details>
+
+            <summary>More</summary>
+
+            ```mermaid
+            ${{ env.PROMPT_TOKENS_SECONDS }}
+            ```
+
+            </details>
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_1 }}" alt="predicted_tokens_seconds"/>
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.PREDICTED_TOKENS_SECONDS }}
+            ```
+
+            </details>
+
+            </p>
+
+            <details>
+
+            <summary>Details</summary>
+
+            <p align="center">
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_2 }}" alt="kv_cache_usage_ratio" />
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.KV_CACHE_USAGE_RATIO }}
+            ```
+
+            </details>
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_3 }}" alt="requests_processing"/>
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.REQUESTS_PROCESSING }}
+            ```
+
+            </details>
+
+            </p>
+            </details>
+            </details>

.github/workflows/build.yml

@@ -16,7 +16,7 @@ on:
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 env:
@@ -31,7 +31,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -76,10 +76,10 @@ jobs:
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+          name: llama-bin-macos-arm64.zip
 
   macOS-latest-cmake-x64:
     runs-on: macos-latest
@@ -87,7 +87,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -132,10 +132,10 @@ jobs:
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+          name: llama-bin-macos-x64.zip
 
   ubuntu-focal-make:
     runs-on: ubuntu-20.04
@@ -146,7 +146,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -158,7 +158,7 @@ jobs:
         with:
           node-version: "20"
 
-      - uses: actions/setup-python@v4
+      - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
 
@@ -181,7 +181,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -203,7 +203,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -249,7 +249,7 @@ jobs:
 #    steps:
 #      - name: Clone
 #        id: checkout
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        id: depends
@@ -283,7 +283,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -311,7 +311,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -357,7 +357,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -398,7 +398,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -418,7 +418,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -449,7 +449,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -593,7 +593,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -723,10 +723,10 @@ jobs:
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
+          name: llama-bin-win-${{ matrix.build }}-x64.zip
 
   windows-latest-cmake-cuda:
     runs-on: windows-latest
@@ -739,7 +739,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -779,10 +779,10 @@ jobs:
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
         run: |
@@ -793,10 +793,10 @@ jobs:
 
       - name: Upload Cuda runtime
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+          path: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
   windows-latest-cmake-sycl:
     runs-on: windows-latest
@@ -812,7 +812,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -840,21 +840,21 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip .\build\bin\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+          name: llama-bin-win-sycl-x64.zip
 
   ios-xcode-build:
     runs-on: macos-latest
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
@@ -864,7 +864,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up JDK
         uses: actions/setup-java@v3
@@ -887,7 +887,7 @@ jobs:
 #    runs-on: macos-12
 #    steps:
 #    - name: Clone
-#      uses: actions/checkout@v3
+#      uses: actions/checkout@v4
 #
 #    - name: Build
 #      uses: cross-platform-actions/action@v0.19.0
@@ -918,7 +918,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -937,7 +937,7 @@ jobs:
 
       - name: Download artifacts
         id: download-artifact
-        uses: actions/download-artifact@v3
+        uses: actions/download-artifact@v4
 
       - name: Create release
         id: create_release
@@ -978,7 +978,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -1002,7 +1002,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -1026,7 +1026,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -1056,7 +1056,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -1072,7 +1072,7 @@ jobs:
 #          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
 #
 #      - name: Upload binaries
-#        uses: actions/upload-artifact@v1
+#        uses: actions/upload-artifact@v4
 #        with:
 #          name: llama-bin-${{ matrix.arch }}
 #          path: build/bin/${{ matrix.build }}
@@ -1095,7 +1095,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -1127,7 +1127,7 @@ jobs:
 #
 #      - name: Upload binaries
 #        if: matrix.blas == 'ON'
-#        uses: actions/upload-artifact@v1
+#        uses: actions/upload-artifact@v4
 #        with:
 #          name: llama-blas-bin-${{ matrix.arch }}
 #          path: build/bin/${{ matrix.build }}
@@ -1141,7 +1141,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |

.github/workflows/code-coverage.yml

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

.github/workflows/docker.yml

@@ -16,7 +16,7 @@ on:
       - master
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -46,7 +46,7 @@ jobs:
           - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v2

.github/workflows/editorconfig.yml

@@ -15,13 +15,13 @@ on:
       - master
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
   editorconfig:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - uses: editorconfig-checker/action-editorconfig-checker@main
       - run: editorconfig-checker

.github/workflows/gguf-publish.yml

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

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

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

.github/workflows/nix-ci.yml

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

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

@@ -17,7 +17,7 @@ on:
       - 'requirements/*.txt'
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -26,9 +26,9 @@ jobs:
     name: check-requirements
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Set up Python environment
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script

.github/workflows/python-lint.yml

@@ -3,7 +3,7 @@ name: flake8 Lint
 on: [push, pull_request]
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -12,9 +12,9 @@ jobs:
     name: Lint
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Set up Python environment
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: "3.11"
       - name: flake8 Lint

.github/workflows/server.yml

@@ -4,6 +4,10 @@ name: Server
 on:
   workflow_dispatch: # allows manual triggering
     inputs:
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
       slow_tests:
         description: 'Run slow tests'
         required: true
@@ -11,15 +15,15 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
-  pull_request:
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+  pull_request_target:
     types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
   schedule:
-    -  cron: '0 0 * * *'
+    -  cron: '2 4 * * *'
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -44,25 +48,45 @@ jobs:
       options: --cpus 4
 
     steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-        with:
-          fetch-depth: 0
-
       - name: Dependencies
         id: depends
         run: |
           apt-get update
           apt-get -y install \
             build-essential \
+            xxd \
             git \
             cmake \
             python3-pip \
+            curl \
             wget \
             language-pack-en \
             libcurl4-openssl-dev
 
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Verify server deps
+        id: verify_server_deps
+        run: |
+          git config --global --add safe.directory $(realpath .)
+          cd examples/server
+          git ls-files --others --modified
+          git status
+          ./deps.sh
+          git status
+          not_ignored_files="$(git ls-files --others --modified)"
+          echo "Modified files: ${not_ignored_files}"
+          if [ -n "${not_ignored_files}" ]; then
+            echo "Repository is dirty or server deps are not built as expected"
+            echo "${not_ignored_files}"
+            exit 1
+          fi
+
       - name: Build
         id: cmake_build
         run: |
@@ -102,7 +126,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 

.github/workflows/zig-build.yml

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

CMakeLists.txt

@@ -113,6 +113,9 @@ option(LLAMA_METAL                           "llama: use Metal"
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
 option(LLAMA_METAL_EMBED_LIBRARY             "llama: embed Metal library"                       OFF)
+set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
+                                             "llama: metal minimum macOS version")
+set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
@@ -250,6 +253,16 @@ if (LLAMA_METAL)
             set(XC_FLAGS -O3)
         endif()
 
+        # Append macOS metal versioning flags
+        if (LLAMA_METAL_MACOSX_VERSION_MIN)
+            message(STATUS "Adding -mmacosx-version-min=${LLAMA_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
+            list(APPEND XC_FLAGS -mmacosx-version-min=${LLAMA_METAL_MACOSX_VERSION_MIN})
+        endif()
+        if (LLAMA_METAL_STD)
+            message(STATUS "Adding -std=${LLAMA_METAL_STD} flag to metal compilation")
+            list(APPEND XC_FLAGS -std=${LLAMA_METAL_STD})
+        endif()
+
         add_custom_command(
             OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
             COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
@@ -1170,6 +1183,7 @@ add_library(llama
             llama.h
             unicode.h
             unicode.cpp
+            unicode-data.cpp
             )
 
 target_include_directories(llama PUBLIC .)

Makefile

@@ -556,6 +556,7 @@ ifdef LLAMA_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)
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 
@@ -678,7 +679,10 @@ ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
 unicode.o: unicode.cpp unicode.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o
+unicode-data.o: unicode-data.cpp unicode-data.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
 
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -863,6 +867,10 @@ passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)

Package.swift

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

README-sycl.md

@@ -3,7 +3,7 @@
 - [Background](#background)
 - [News](#news)
 - [OS](#os)
-- [Intel GPU](#intel-gpu)
+- [Supported Devices](#supported-devices)
 - [Docker](#docker)
 - [Linux](#linux)
 - [Windows](#windows)
@@ -14,17 +14,25 @@
 
 ## Background
 
-SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators—such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
+**SYCL** is a high-level parallel programming model designed to improve developers productivity writing code across various hardware accelerators such as CPUs, GPUs, and FPGAs. It is a single-source language designed for heterogeneous computing and based on standard C++17.
 
-oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
+**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
 
-Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
+- **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL - Math Kernel Library)*.
+- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
+- **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
-To avoid to re-invent the wheel, this code refer other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
+### Llama.cpp + SYCL
+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.
 
-The llama.cpp for SYCL is used to support Intel GPUs.
+The llama.cpp SYCL backend supports:
+- Intel GPUs.
+- Nvidia GPUs.
 
-For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
+*Upcoming support: AMD GPUs*.
+
+When targetting **Intel CPUs**, it is recommended to  use llama.cpp for [x86_64](README.md#intel-onemkl) approach.
 
 ## News
 
@@ -51,9 +59,16 @@ For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
 |Windows|Support|Windows 11|
 
 
-## Intel GPU
+## Supported devices
 
-### Verified
+### 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**
 
 |Intel GPU| Status | Verified Model|
 |-|-|-|
@@ -63,198 +78,229 @@ For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
 |Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
 |Intel iGPU| Support| iGPU in i5-1250P, i7-1260P, i7-1165G7|
 
-Note: If the EUs (Execution Unit) in iGPU is less than 80, the inference speed will be too slow to use.
+*Notes:*
 
-### Memory
+- 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`.
 
-The memory is a limitation to run LLM on GPUs.
+- 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.
 
-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`.
+- If the iGPU has less than 80  EUs *(Execution Unit)*, the inference speed will likely be too slow for practical use.
 
-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+.
+### 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)
 
-For dGPU, please make sure the device memory is enough. For llama-2-7b.Q4_0, recommend the device memory is 4GB+.
+- **Tested devices**
 
-## Nvidia GPU
-
-### Verified
-
-|Intel GPU| Status | Verified Model|
+|Nvidia GPU| Status | Verified Model|
 |-|-|-|
-|Ampere Series| Support| A100|
+|Ampere Series| Support| A100, A4000|
+|Ampere Series *(Mobile)*| Support| RTX 40 Series|
 
-### oneMKL for CUDA
+*Notes:*
+  - Support for Nvidia targets through oneAPI is currently limited to Linux platforms.
 
-The current oneMKL release does not contain the oneMKL cuBlas backend.
-As a result for Nvidia GPU's oneMKL must be built from source.
+  - 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.
 
-```
-git clone https://github.com/oneapi-src/oneMKL
-cd oneMKL
-mkdir build
-cd build
-cmake -G Ninja .. -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON
-ninja
-// Add paths as necessary
-```
 
 ## Docker
-
-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.
-
-
+The docker build option is currently limited to *intel GPU* targets.
+### Build image
 ```sh
-# For F16:
-#docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
-
-# Or, for F32:
-docker build -t llama-cpp-sycl -f .devops/main-intel.Dockerfile .
-
-# Note: you can also use the ".devops/server-intel.Dockerfile", which compiles the "server" example
+# Using FP16
+docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
 ```
 
-### Run
+*Notes*:
+
+To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="LLAMA_SYCL_F16=ON"` argument from the previous command.
+
+You can also use the `.devops/server-intel.Dockerfile`, which builds the *"server"* alternative.
+
+### Run container
 
 ```sh
-# Firstly, find all the DRI cards:
+# First, find all the DRI cards
 ls -la /dev/dri
-# Then, pick the card that you want to use.
-
-# For example with "/dev/dri/card1"
+# Then, pick the card that you want to use (here for e.g. /dev/dri/card1).
 docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-sycl -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
 ```
 
+*Notes:*
+- Docker has been tested successfully on native Linux. WSL support has not been verified yet.
+- You may need to install Intel GPU driver on the **host** machine *(Please refer to the [Linux configuration](#linux) for details)*.
+
 ## Linux
 
-### Setup Environment
+### I. Setup Environment
 
-1. Install Intel GPU driver.
+1. **Install GPU drivers**
 
-a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
+  - **Intel GPU**
 
-Note: for iGPU, please install the client GPU driver.
+Intel data center GPUs drivers installation guide and download page can be found here: [Get intel dGPU Drivers](https://dgpu-docs.intel.com/driver/installation.html#ubuntu-install-steps).
 
-b. Add user to group: video, render.
+*Note*: for client GPUs *(iGPU & Arc A-Series)*, please refer to the [client iGPU driver installation](https://dgpu-docs.intel.com/driver/client/overview.html).
+
+Once installed, add the user(s) to the `video` and `render` groups.
 
 ```sh
-sudo usermod -aG render username
-sudo usermod -aG video username
+sudo usermod -aG render $USER
+sudo usermod -aG video $USER
 ```
 
-Note: re-login to enable it.
+*Note*: logout/re-login for the changes to take effect.
 
-c. Check
+Verify installation through `clinfo`:
 
 ```sh
 sudo apt install clinfo
 sudo clinfo -l
 ```
 
-Output (example):
+Sample output:
 
-```
+```sh
 Platform #0: Intel(R) OpenCL Graphics
  `-- Device #0: Intel(R) Arc(TM) A770 Graphics
 
-
 Platform #0: Intel(R) OpenCL HD Graphics
  `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
 ```
 
-2. Install Intel® oneAPI Base toolkit.
+- **Nvidia GPU**
 
-a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
+In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#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 |
++-----------------------------------------+----------------------+----------------------+
+```
 
-Recommend to install to default folder: **/opt/intel/oneapi**.
 
-Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
+2. **Install Intel® oneAPI Base toolkit**
 
-b. Check
+- **Base installation**
+
+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
-source /opt/intel/oneapi/setvars.sh
+git clone https://github.com/oneapi-src/oneMKL
+cd oneMKL
+mkdir -p buildWithCublas && cd buildWithCublas
+cmake ../ -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas
+make
+```
 
+
+3. **Verify installation and environment**
+
+In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
+```sh
+source /opt/intel/oneapi/setvars.sh
 sycl-ls
 ```
 
-There should be one or more level-zero devices. Please confirm that at least one GPU is present, like **[ext_oneapi_level_zero:gpu:0]**.
+- **Intel GPU**
+
+When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`ext_oneapi_level_zero:gpu:0`] in the sample output below:
 
-Output (example):
 ```
 [opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
 [opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
 [opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
 [ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
-
 ```
 
-2. Build locally:
+- **Nvidia GPU**
 
-Note:
-- You can choose between **F16** and **F32** build. F16 is faster for long-prompt inference.
-- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
+Similarly, user 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]
+```
 
+### II. Build llama.cpp
+
+#### Intel GPU
 ```sh
-mkdir -p build
-cd build
+# Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
 
-# For FP16:
-#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+# Build LLAMA with MKL BLAS acceleration for intel GPU
+mkdir -p build && cd build
 
-# Or, for FP32:
+# Option 1: Use 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
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# For Nvidia GPUs
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# Build example/main only
-#cmake --build . --config Release --target main
-
-# Or, build all binary
-cmake --build . --config Release -v
-
-cd ..
 ```
 
-or
-
+#### Nvidia GPU
 ```sh
-./examples/sycl/build.sh
+# Export relevant ENV variables
+export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
+export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
+
+# Build LLAMA with Nvidia BLAS acceleration through SYCL
+mkdir -p build && cd build
+
+# Option 1: Use FP16 for better performance in long-prompt  inference
+cmake .. -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
 ```
 
-### Run
+### III. Run the inference
 
-1. Put model file to folder **models**
+1. Retrieve and prepare model
 
-You could download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
 2. Enable oneAPI running environment
 
-```
+```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List device ID
+3. List devices information
 
-Run without parameter:
+Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/ls-sycl-device
-
-# or running the "main" executable and look at the output log:
-
-./build/bin/main
 ```
-
-Check the ID in startup log, like:
-
+A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following:
 ```
 found 6 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
@@ -270,15 +316,15 @@ found 6 SYCL devices:
 
 |Attribute|Note|
 |-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
+|compute capability 1.3|Level-zero driver/runtime, recommended |
+|compute capability 3.0|OpenCL driver/runtime, slower than level-zero in most cases|
 
-4. Device selection and execution of llama.cpp
+4. Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device assigned by user.
-- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
+- Single device: Use one device target specified by the user.
+- Multiple devices: Automatically select the devices with the same largest Max compute-units.
 
 |Device selection|Parameter|
 |-|-|
@@ -303,74 +349,64 @@ or run by script:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
-or run by script:
+
+Otherwise, you can run the script:
 
 ```sh
 ./examples/sycl/run_llama2.sh
 ```
 
-Note:
+*Notes:*
 
-- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
+- 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:
 
-
-5. Verify the device ID in output
-
-Verify to see if the selected GPU is shown in the output, like:
-
-```
+```sh
 detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
 Or
-```
+```sh
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
-
 ## Windows
 
-### Setup Environment
+### I. Setup Environment
 
-1. Install Intel GPU driver.
+1. Install GPU driver
 
-Please install Intel GPU driver by official guide: [Install GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+Intel GPU drivers instructions guide and download page can be found here: [Get intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
 
-Note: **The driver is mandatory for compute function**.
+2. Install Visual Studio
 
-2. Install Visual Studio.
+If you already have a recent version of Microsoft Visual Studio, you can skip this step. Otherwise, please refer to the official download page for [Microsoft Visual Studio](https://visualstudio.microsoft.com/).
 
-Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact oneAPI environment enabling in Windows.
+3. Install Intel® oneAPI Base toolkit
 
-3. Install Intel® oneAPI Base toolkit.
+The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
 
-a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
+Please follow the instructions for downloading and installing the Toolkit for Windows, and preferably keep the default installation values unchanged, notably the installation path *(`C:\Program Files (x86)\Intel\oneAPI` by default)*.
 
-Recommend to install to default folder: **C:\Program Files (x86)\Intel\oneAPI**.
-
-Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
+Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable.
 
 b. Enable oneAPI running environment:
 
-- In Search, input 'oneAPI'.
+- Type "oneAPI" in the search bar, then open the `Intel oneAPI command prompt for Intel 64 for Visual Studio 2022` App.
 
-Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
-
-- In Run:
-
-In CMD:
+- On the command prompt, enable the runtime environment with the following:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-c. Check GPU
+c. Verify installation
 
-In oneAPI command line:
+In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
 sycl-ls
 ```
 
-There should be one or more level-zero devices. Please confirm that at least one GPU is present, like **[ext_oneapi_level_zero:gpu:0]**.
+There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
 
 Output (example):
 ```
@@ -380,7 +416,7 @@ Output (example):
 [ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Iris(R) Xe Graphics 1.3 [1.3.28044]
 ```
 
-4. Install cmake & make
+4. Install build tools
 
 a. Download & install cmake for Windows: https://cmake.org/download/
 
@@ -390,76 +426,53 @@ b. Download & install mingw-w64 make for Windows provided by w64devkit
 
 - Extract `w64devkit` on your pc.
 
-- Add the **bin** folder path in the Windows system PATH environment, like `C:\xxx\w64devkit\bin\`.
+- Add the **bin** folder path in the Windows system PATH environment (for e.g. `C:\xxx\w64devkit\bin\`).
 
-### Build locally:
+### II. Build llama.cpp
 
-In oneAPI command line window:
+On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
 mkdir -p build
 cd build
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
-::  for FP16
-::  faster for long-prompt inference
-::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
-::  for FP32
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
-
-
-::  build example/main only
-::  make main
-
-::  build all binary
-make -j
-cd ..
+make
 ```
 
-or
-
-```
+Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
+```sh
 .\examples\sycl\win-build-sycl.bat
 ```
 
-Note:
+*Notes:*
 
-- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
+- By default, calling `make` will build all target binary files. In case of a minimal experimental setup, the user can build the inference executable only through `make main`.
 
-### Run
+### III. Run the inference
 
-1. Put model file to folder **models**
+1. Retrieve and prepare model
 
-You could download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) as example.
+You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
 2. Enable oneAPI running environment
 
-- In Search, input 'oneAPI'.
-
-Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
-
-- In Run:
-
-In CMD:
+On the oneAPI command line window, run the following and step into the llama.cpp directory:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-3. List device ID
+3. List devices information
 
-Run without parameter:
+Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
 build\bin\ls-sycl-device.exe
-
-or
-
-build\bin\main.exe
 ```
 
-Check the ID in startup log, like:
-
+The output of this command in a system with 1 *intel CPU* and 1 *intel GPU* would look like the following:
 ```
 found 6 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
@@ -480,7 +493,7 @@ found 6 SYCL devices:
 |compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
 
 
-4. Device selection and execution of llama.cpp
+4. Launch inference
 
 There are two device selection modes:
 
@@ -505,7 +518,7 @@ build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be
 ```
 build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
-or run by script:
+Otherwise, run the following wrapper script:
 
 ```
 .\examples\sycl\win-run-llama2.bat
@@ -513,19 +526,14 @@ or run by script:
 
 Note:
 
-- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
+- 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:
 
-
-
-5. Verify the device ID in output
-
-Verify to see if the selected GPU is shown in the output, like:
-
-```
+```sh
 detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
 Or
-```
+```sh
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
@@ -535,64 +543,54 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 |Name|Value|Function|
 |-|-|-|
-|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, LLAMA_SYCL=ON is mandatory.|
-|LLAMA_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path. Faster for long-prompt inference. <br>For FP32, not set it.|
-|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
-|CMAKE_CXX_COMPILER|icpx (Linux), icx (Windows)|use icpx/icx for SYCL code path|
-
-#### Running
+|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.|
 
+#### Runtime
 
 |Name|Value|Function|
 |-|-|-|
 |GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
 |ZES_ENABLE_SYSMAN| 0 (default) or 1|Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer|
 
-## Known Issue
+## Known Issues
 
-- Hang during startup
+- Hanging during startup
 
-  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
+  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.
 
-  Solution: add **--no-mmap** or **--mmap 0**.
+  - **Solution**: add `--no-mmap` or `--mmap 0` flag to the `main` executable.
 
-- Split-mode: [row] is not supported
-
-  It's on developing.
+- `Split-mode:[row]` is not supported.
 
 ## Q&A
 
-Note: please add prefix **[SYCL]** in issue title, so that we will check it as soon as possible.
-
-
 - Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
 
-  Miss to enable oneAPI running environment.
+  - Potential cause: Unavailable oneAPI installation or not set ENV variables.
+  - Solution: Install *oneAPI base toolkit* and enable its ENV through: `source /opt/intel/oneapi/setvars.sh`.
 
-  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
+- General compiler error:
 
-- In Windows, no result, not error.
+  - Remove build folder or try a clean-build.
 
-  Miss to enable oneAPI running environment.
+- I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
 
-- Meet compile error.
+  Please double-check with `sudo sycl-ls`.
 
-  Remove folder **build** and try again.
-
-- I can **not** see **[ext_oneapi_level_zero:gpu:0]** afer install GPU driver in Linux.
-
-  Please run **sudo sycl-ls**.
-
-  If you see it in result, please add video/render group to your ID:
+  If it's present in the list, please add video/render group to your user then **logout/login** or restart your system:
 
   ```
-  sudo usermod -aG render username
-  sudo usermod -aG video username
+  sudo usermod -aG render $USER
+  sudo usermod -aG video $USER
   ```
+  Otherwise, please double-check the GPU driver installation steps.
 
-  Then **relogin**.
-
-  If you do not see it, please check the installation GPU steps again.
+### **GitHub contribution**:
+Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
 ## Todo
 

README.md

@@ -10,6 +10,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### 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
@@ -17,12 +18,13 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
+- **MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387**
+- Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
 - Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
-- Multi-GPU pipeline parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017
+- Multi-GPU pipeline parallelism support https://github.com/ggerganov/llama.cpp/pull/6017
 - Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
 - Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
 - Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
-- Support loading sharded model, using `gguf-split` CLI https://github.com/ggerganov/llama.cpp/pull/6187
 
 ----
 
@@ -114,7 +116,9 @@ Typically finetunes of the base models below are supported as well.
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
 - [x] [Gemma](https://ai.google.dev/gemma)
 - [x] [Mamba](https://github.com/state-spaces/mamba)
+- [x] [Xverse](https://huggingface.co/models?search=xverse)
 - [x] [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)
+- [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 
 **Multimodal models:**
 
@@ -138,6 +142,7 @@ Typically finetunes of the base models below are supported as well.
 - JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
 - Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
+- Rust (more features): [edgenai/llama_cpp-rs](https://github.com/edgenai/llama_cpp-rs)
 - Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
@@ -147,6 +152,7 @@ Typically finetunes of the base models below are supported as well.
 - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
 - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
+- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
 
 **UI:**
 
@@ -173,6 +179,10 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
 - [Msty](https://msty.app) (proprietary)
 - [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
+- [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file)(Apachev2.0 or later)
+- [Dot](https://github.com/alexpinel/Dot) (GPL)
+
+*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 ---
 

SECURITY.md

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

build.zig

@@ -116,6 +116,7 @@ pub fn build(b: *std.build.Builder) !void {
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
     const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
     const unicode = make.obj("unicode", "unicode.cpp");
+    const unicode_data = make.obj("unicode-data", "unicode-data.cpp");
     const llama = make.obj("llama", "llama.cpp");
     const buildinfo = make.obj("common", "common/build-info.cpp");
     const common = make.obj("common", "common/common.cpp");
@@ -127,14 +128,14 @@ pub fn build(b: *std.build.Builder) !void {
     const clip = make.obj("clip", "examples/llava/clip.cpp");
     const llava = make.obj("llava", "examples/llava/llava.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, json_schema_to_grammar, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, grammar_parser, json_schema_to_grammar, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -575,7 +575,7 @@ function gg_run_embd_bge_small {
     cd ${SRC}
 
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
-    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.model
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer.json
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer_config.json
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/special_tokens_map.json
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin

common/common.cpp

@@ -1062,8 +1062,8 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.ignore_eos = true;
         return true;
     }
-    if (arg == "--no-penalize-nl") {
-        sparams.penalize_nl = false;
+    if (arg == "--penalize-nl") {
+        sparams.penalize_nl = true;
         return true;
     }
     if (arg == "-l" || arg == "--logit-bias") {
@@ -1373,7 +1373,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -dt N, --defrag-thold N\n");
     printf("                        KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
-    printf("  --no-penalize-nl      do not penalize newline token\n");
+    printf("  --penalize-nl         penalize newline tokens\n");
     printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
     printf("  --all-logits          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
@@ -1928,11 +1928,6 @@ struct llama_model * llama_load_model_from_url(
         return NULL;
     }
 
-    if (!curl) {
-        fprintf(stderr, "%s: error initializing libcurl\n", __func__);
-        return NULL;
-    }
-
     if (!llama_download_file(curl, model_url, path_model)) {
         return NULL;
     }
@@ -2489,7 +2484,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
     fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
-    fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
+    fprintf(stream, "penalize_nl: %s # default: false\n", sparams.penalize_nl ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
     fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);

convert-hf-to-gguf.py

@@ -23,7 +23,7 @@ if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
-from convert import HfVocab
+from convert import LlamaHfVocab, permute
 
 
 ###### MODEL DEFINITIONS ######
@@ -230,7 +230,7 @@ class Model(ABC):
     def _set_vocab_gpt2(self):
         dir_model = self.dir_model
         hparams = self.hparams
-        tokens: list[bytearray] = []
+        tokens: list[str] = []
         toktypes: list[int] = []
 
         from transformers import AutoTokenizer
@@ -243,8 +243,7 @@ class Model(ABC):
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
-                pad_token = f"[PAD{i}]".encode('utf-8')
-                tokens.append(bytearray(pad_token))
+                tokens.append(f"[PAD{i}]")
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
@@ -266,7 +265,7 @@ class Model(ABC):
     def _set_vocab_qwen(self):
         dir_model = self.dir_model
         hparams = self.hparams
-        tokens: list[bytearray] = []
+        tokens: list[str] = []
         toktypes: list[int] = []
 
         from transformers import AutoTokenizer
@@ -291,8 +290,7 @@ class Model(ABC):
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
-                pad_token = f"[PAD{i}]".encode("utf-8")
-                tokens.append(bytearray(pad_token))
+                tokens.append(f"[PAD{i}]")
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
@@ -325,13 +323,12 @@ class Model(ABC):
         toktypes: list[int] = []
 
         if not tokenizer_path.is_file():
-            print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
-            sys.exit(1)
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
 
         tokenizer = SentencePieceProcessor(str(tokenizer_path))
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
-        for token_id in range(vocab_size):
+        for token_id in range(tokenizer.vocab_size()):
             piece = tokenizer.id_to_piece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.get_score(token_id)
@@ -356,9 +353,13 @@ class Model(ABC):
                 added_tokens_json = json.load(f)
 
                 for key in added_tokens_json:
-                    tokens.append(key.encode("utf-8"))
-                    scores.append(-1000.0)
-                    toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+                    key = key.encode("utf-8")
+                    if key not in tokens:
+                        tokens.append(key)
+                        scores.append(-1000.0)
+                        toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+
+        assert len(tokens) == vocab_size
 
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_token_list(tokens)
@@ -368,12 +369,8 @@ class Model(ABC):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-    def _set_vocab_hf(self):
-        path = self.dir_model
-        added_tokens_path = self.dir_model
-        vocab = HfVocab(
-            path, added_tokens_path if added_tokens_path.exists() else None
-        )
+    def _set_vocab_llama_hf(self):
+        vocab = LlamaHfVocab(self.dir_model)
         tokens = []
         scores = []
         toktypes = []
@@ -513,6 +510,17 @@ class BloomModel(Model):
 class MPTModel(Model):
     model_arch = gguf.MODEL_ARCH.MPT
 
+    def set_vocab(self):
+        try:
+            self._set_vocab_gpt2()
+        except Exception:
+            # Fallback for SEA-LION model
+            self._set_vocab_sentencepiece()
+            self.gguf_writer.add_add_bos_token(False)
+            self.gguf_writer.add_pad_token_id(3)
+            self.gguf_writer.add_eos_token_id(1)
+            self.gguf_writer.add_unk_token_id(0)
+
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layers"]
         self.gguf_writer.add_name(self.dir_model.name)
@@ -526,7 +534,10 @@ class MPTModel(Model):
         self.gguf_writer.add_layer_norm_eps(1e-5)
         if self.hparams["attn_config"]["clip_qkv"] is not None:
             self.gguf_writer.add_clamp_kqv(self.hparams["attn_config"]["clip_qkv"])
-        self.gguf_writer.add_max_alibi_bias(self.hparams["attn_config"]["alibi_bias_max"])
+        if self.hparams["attn_config"]["alibi"]:
+            self.gguf_writer.add_max_alibi_bias(self.hparams["attn_config"]["alibi_bias_max"])
+        else:
+            self.gguf_writer.add_max_alibi_bias(0.0)
 
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers"))
@@ -775,6 +786,148 @@ class BaichuanModel(Model):
         return weights[r * n_part:r * n_part + r, ...]
 
 
+@Model.register("XverseForCausalLM")
+class XverseModel(Model):
+    model_arch = gguf.MODEL_ARCH.XVERSE
+
+    def set_vocab(self):
+        assert (self.dir_model / "tokenizer.json").is_file()
+        dir_model = self.dir_model
+        hparams = self.hparams
+
+        tokens: list[bytearray] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model)
+        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+        assert max(tokenizer.vocab.values()) < vocab_size
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        for token_id in range(vocab_size):
+            token_text = reverse_vocab[token_id].encode('utf-8')
+            # replace "\x00" to string with length > 0
+            if token_text == b"\x00":
+                toktype = gguf.TokenType.BYTE  # special
+                token_text = f"<{token_text}>".encode('utf-8')
+            elif re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
+                toktype = gguf.TokenType.BYTE  # special
+            elif reverse_vocab[token_id] in added_vocab:
+                if tokenizer.added_tokens_decoder[token_id].special:
+                    toktype = gguf.TokenType.CONTROL
+                else:
+                    toktype = gguf.TokenType.USER_DEFINED
+            else:
+                toktype = gguf.TokenType.NORMAL
+
+            tokens.append(token_text)
+            toktypes.append(toktype)
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+        hf_repo = self.hparams.get("_name_or_path", "")
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            print("gguf: can not find ctx length parameter.")
+            sys.exit()
+
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_source_hf_repo(hf_repo)
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+
+    def write_tensors(self):
+        # Collect tensors from generator object
+        model_kv = dict(self.get_tensors())
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+
+        for name, data_torch in model_kv.items():
+            # we don't need these
+            if name.endswith(".rotary_emb.inv_freq"):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            # HF models permute some of the tensors, so we need to undo that
+            if name.endswith(("q_proj.weight")):
+                data_torch = self._reverse_hf_permute(data_torch, head_count, head_count)
+            if name.endswith(("k_proj.weight")):
+                data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+            self.gguf_writer.add_tensor(new_name, data)
+
+    def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
+        if n_kv_head is not None and n_head != n_kv_head:
+            n_head //= n_kv_head
+
+        return (
+            weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape)
+        )
+
+
 @Model.register("FalconForCausalLM", "RWForCausalLM")
 class FalconModel(Model):
     model_arch = gguf.MODEL_ARCH.FALCON
@@ -1054,12 +1207,120 @@ class StableLMModel(Model):
         self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
 
 
-@Model.register("MixtralForCausalLM")
-class MixtralModel(Model):
+@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
+class LlamaModel(Model):
     model_arch = gguf.MODEL_ARCH.LLAMA
 
     def set_vocab(self):
-        self._set_vocab_sentencepiece()
+        try:
+            self. _set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_llama_hf()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+
+    # Same as super class, but permuting q_proj, k_proj
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        n_head = self.hparams.get("num_attention_heads")
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        n_experts = self.hparams.get("num_local_experts")
+        experts = dict()
+        for name, data_torch in self.get_tensors():
+            # we don't need these
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.numpy()
+
+            if name.endswith("q_proj.weight"):
+                data = permute(data, n_head, n_head)
+            if name.endswith("k_proj.weight"):
+                data = permute(data, n_head, n_kv_head)
+
+            data = data.squeeze()
+
+            # process the experts separately
+            if name.find("block_sparse_moe.experts") != -1:
+                experts[name] = data
+                if len(experts) >= n_experts:
+                    # merge the experts into a single 3d tensor
+                    for bid in range(block_count):
+                        for wid in range(1, 4):
+                            full = True
+                            for xid in range(n_experts):
+                                ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.w{wid}.weight"
+                                if ename not in experts:
+                                    full = False
+                                    break
+                            if not full:
+                                continue
+
+                            datas = []
+                            for xid in range(n_experts):
+                                ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.w{wid}.weight"
+                                datas.append(experts[ename])
+                                del experts[ename]
+
+                            data = np.stack(datas, axis=0)
+                            data_dtype = data.dtype
+
+                            if self.ftype == 0 and data_dtype == np.float16:
+                                data = data.astype(np.float32)
+
+                            if self.ftype == 1 and data_dtype == np.float32:
+                                data = data.astype(np.float16)
+
+                            merged_name = f"layers.{bid}.feed_forward.experts.w{wid}.weight"
+
+                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
+                            if new_name is None:
+                                print(f"Can not map tensor {name!r}")
+                                sys.exit()
+
+                            print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
+
+                            self.gguf_writer.add_tensor(new_name, data)
+                continue
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # 1d tensors need to be converted to float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+        if len(experts) > 0:
+            raise ValueError(f"Unprocessed experts: {experts.keys()}")
 
 
 @Model.register("GrokForCausalLM")
@@ -1076,6 +1337,92 @@ class GrokModel(Model):
         super().set_gguf_parameters()
         self.gguf_writer.add_name("Grok")
 
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        n_experts = self.hparams.get("num_local_experts")
+        experts = dict()
+        for name, data_torch in self.get_tensors():
+            # we don't need these
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # process the experts separately
+            if name.find(".moe.") != -1:
+                experts[name] = data
+                if len(experts) >= n_experts:
+                    # merge the experts into a single 3d tensor
+                    for bid in range(block_count):
+                        for wid in ["linear", "linear_1", "linear_v"]:
+                            full = True
+                            for xid in range(n_experts):
+                                ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid}.weight"
+                                if ename not in experts:
+                                    full = False
+                                    break
+                            if not full:
+                                continue
+
+                            datas = []
+                            for xid in range(n_experts):
+                                ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid}.weight"
+                                datas.append(experts[ename])
+                                del experts[ename]
+
+                            data = np.stack(datas, axis=0)
+                            data_dtype = data.dtype
+
+                            if self.ftype == 0 and data_dtype == np.float16:
+                                data = data.astype(np.float32)
+
+                            if self.ftype == 1 and data_dtype == np.float32:
+                                data = data.astype(np.float16)
+
+                            merged_name = f"transformer.decoder_layer.{bid}.moe.{wid}.weight"
+
+                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
+                            if new_name is None:
+                                print(f"Can not map tensor {name!r}")
+                                sys.exit()
+
+                            print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
+
+                            self.gguf_writer.add_tensor(new_name, data)
+                continue
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
 
 @Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
@@ -1095,7 +1442,7 @@ class MiniCPMModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
 
     def set_vocab(self):
-        self._set_vocab_hf()
+        self._set_vocab_llama_hf()
 
     def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
         if n_kv_head is not None and n_head != n_kv_head:
@@ -1696,11 +2043,8 @@ class BertModel(Model):
             self.gguf_writer.add_pooling_type(pooling_type)
 
     def set_vocab(self):
-        path = self.dir_model
-        added_tokens_path = self.dir_model if self.dir_model.exists() else None
-
         # use huggingface vocab to get all tokens
-        vocab = HfVocab(path, added_tokens_path)
+        vocab = LlamaHfVocab(self.dir_model, ignore_nonllama=True)
         tokens, scores, toktypes = zip(*vocab.all_tokens())
         assert len(tokens) == vocab.vocab_size
         self.vocab_size = vocab.vocab_size

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 = tensors[name]
+        data_torch = tensors[name]
         if name.endswith(".self_attention.rotary_emb.inv_freq"):
             continue
-        old_dtype = data.dtype
+        old_dtype = data_torch.dtype
         # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
-        data = data.to(torch.float32).squeeze().numpy()
+        data = data_torch.to(torch.float32).squeeze().numpy()
         new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
         if new_name is None:
             print("Can not map tensor '" + name + "'")

convert.py

@@ -16,13 +16,14 @@ import re
 import signal
 import struct
 import sys
+import textwrap
 import time
 import zipfile
-from abc import ABCMeta, abstractmethod
+from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar
+from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable
 
 import numpy as np
 from sentencepiece import SentencePieceProcessor
@@ -43,6 +44,9 @@ ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
 
+ADDED_TOKENS_FILE = 'added_tokens.json'
+FAST_TOKENIZER_FILE = 'tokenizer.json'
+
 #
 # data types
 #
@@ -188,8 +192,10 @@ class Params:
             n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
 
         if n_layer < 1:
-            raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n"
-                            "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
+            msg = """\
+                failed to guess 'n_layer'. This model is unknown or unsupported.
+                Suggestion: provide 'config.json' of the model in the same directory containing model files."""
+            raise KeyError(textwrap.dedent(msg))
 
         n_head = n_embd // 128 # guessed
         n_mult = 256           # guessed
@@ -211,7 +217,8 @@ class Params:
 
     @staticmethod
     def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
-        config = json.load(open(config_path))
+        with open(config_path) as f:
+            config = json.load(f)
 
         rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
         rope_scaling = config.get("rope_scaling")
@@ -233,8 +240,10 @@ class Params:
         elif "max_position_embeddings" in config:
             n_ctx = config["max_position_embeddings"]
         else:
-            raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
-                            "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
+            msg = """\
+                failed to guess 'n_ctx'. This model is unknown or unsupported.
+                Suggestion: provide 'config.json' of the model in the same directory containing model files."""
+            raise KeyError(textwrap.dedent(msg))
 
         n_experts      = None
         n_experts_used = None
@@ -265,7 +274,8 @@ class Params:
     # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1}
     @staticmethod
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
-        config = json.load(open(config_path))
+        with open(config_path) as f:
+            config = json.load(f)
 
         n_experts      = None
         n_experts_used = None
@@ -331,47 +341,86 @@ class Params:
 # vocab
 #
 
-class BpeVocab:
+@runtime_checkable
+class BaseVocab(Protocol):
+    tokenizer_model: ClassVar[str]
+    name: ClassVar[str]
+
+
+class NoVocab(BaseVocab):
+    tokenizer_model = "no_vocab"
+    name = "no_vocab"
+
+    def __repr__(self) -> str:
+        return "<NoVocab for a model without integrated vocabulary>"
+
+
+@runtime_checkable
+class Vocab(BaseVocab, Protocol):
+    vocab_size: int
+    added_tokens_dict: dict[str, int]
+    added_tokens_list: list[str]
+    fname_tokenizer: Path
+
+    def __init__(self, base_path: Path): ...
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ...
+
+
+class BpeVocab(Vocab):
     tokenizer_model = "gpt2"
     name = "bpe"
 
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
-        self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
-        if isinstance(self.bpe_tokenizer.get('model'), dict):
-            self.vocab = self.bpe_tokenizer["model"]["vocab"]
-        else:
-            self.vocab = self.bpe_tokenizer
-        added_tokens: dict[str, int]
-        if fname_added_tokens is not None:
-            # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
-            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
-        else:
-            # Fall back to trying to find the added tokens in tokenizer.json
-            tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
-            if not tokenizer_json_file.is_file():
-                added_tokens = {}
-            else:
-                tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
-                added_tokens = dict(
-                    (item['content'], item['id'])
-                    for item in tokenizer_json.get('added_tokens', [])
-                    # Added tokens here can be duplicates of the main vocabulary.
-                    if item['content'] not in self.bpe_tokenizer)
+    def __init__(self, base_path: Path):
+        added_tokens: dict[str, int] = {}
 
-        vocab_size: int = len(self.vocab)
-        expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
-        actual_ids      = sorted(added_tokens.values())
+        if (fname_tokenizer := base_path / 'vocab.json').exists():
+            # "slow" tokenizer
+            with open(fname_tokenizer, encoding="utf-8") as f:
+                self.vocab = json.load(f)
+
+            try:
+                # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
+                with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f:
+                    added_tokens = json.load(f)
+            except FileNotFoundError:
+                pass
+        else:
+            # "fast" tokenizer
+            fname_tokenizer = base_path / FAST_TOKENIZER_FILE
+
+            # if this fails, FileNotFoundError propagates to caller
+            with open(fname_tokenizer, encoding="utf-8") as f:
+                tokenizer_json = json.load(f)
+
+            tokenizer_model: dict[str, Any] = tokenizer_json['model']
+            if (
+                tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False)
+                or tokenizer_json['decoder']['type'] != 'ByteLevel'
+            ):
+                raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer')
+
+            self.vocab = tokenizer_model["vocab"]
+
+            if (added := tokenizer_json.get('added_tokens')) is not None:
+                # Added tokens here can be duplicates of the main vocabulary.
+                added_tokens = {item['content']: item['id']
+                                for item in added
+                                if item['content'] not in self.vocab}
+
+        vocab_size   = len(self.vocab)
+        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
+        actual_ids   = sorted(added_tokens.values())
         if expected_ids != actual_ids:
             expected_end_id = vocab_size + len(actual_ids) - 1
-            raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
+            raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range "
+                             f"{vocab_size} - {expected_end_id}; got {actual_ids}")
 
         items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
         self.added_tokens_dict    = added_tokens
         self.added_tokens_list    = [text for (text, idx) in items]
-        self.vocab_size_base: int = vocab_size
-        self.vocab_size: int      = self.vocab_size_base + len(self.added_tokens_list)
+        self.vocab_size_base      = vocab_size
+        self.vocab_size           = self.vocab_size_base + len(self.added_tokens_list)
         self.fname_tokenizer      = fname_tokenizer
-        self.fname_added_tokens   = fname_added_tokens
 
     def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()}
@@ -392,19 +441,25 @@ class BpeVocab:
         return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-class SentencePieceVocab:
+class SentencePieceVocab(Vocab):
     tokenizer_model = "llama"
     name = "spm"
 
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
-        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
-        added_tokens: dict[str, int]
-        if fname_added_tokens is not None:
-            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
-        else:
-            added_tokens = {}
+    def __init__(self, base_path: Path):
+        added_tokens: dict[str, int] = {}
+        if (fname_tokenizer := base_path / 'tokenizer.model').exists():
+            # normal location
+            try:
+                with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f:
+                    added_tokens = json.load(f)
+            except FileNotFoundError:
+                pass
+        elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists():
+            # not found in alternate location either
+            raise FileNotFoundError('Cannot find tokenizer.model')
 
-        vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
+        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
+        vocab_size = self.sentencepiece_tokenizer.vocab_size()
 
         new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
         expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
@@ -414,18 +469,17 @@ class SentencePieceVocab:
             raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
 
         # Token pieces that were added to the base vocabulary.
-        self.added_tokens_dict = added_tokens
+        self.added_tokens_dict  = added_tokens
         self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
         self.vocab_size_base    = vocab_size
         self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
         self.fname_tokenizer    = fname_tokenizer
-        self.fname_added_tokens = fname_added_tokens
 
     def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         tokenizer = self.sentencepiece_tokenizer
         for i in range(tokenizer.vocab_size()):
             piece = tokenizer.id_to_piece(i)
-            text: bytes = piece.encode("utf-8")
+            text         = piece.encode("utf-8")
             score: float = tokenizer.get_score(i)
 
             toktype = gguf.TokenType.NORMAL
@@ -458,27 +512,42 @@ class SentencePieceVocab:
         return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-class HfVocab:
+class LlamaHfVocab(Vocab):
     tokenizer_model = "llama"
     name = "hfft"
 
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
+    def __init__(self, base_path: Path, 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')
+
         try:
             from transformers import AutoTokenizer
         except ImportError as e:
             raise ImportError(
-                "To use HfVocab, please install the `transformers` package. "
+                "To use LlamaHfVocab, please install the `transformers` package. "
                 "You can install it with `pip install transformers`."
             ) from e
 
-        print("fname_tokenizer:", fname_tokenizer)
         # Allow the tokenizer to default to slow or fast versions.
         # Explicitly set tokenizer to use local paths.
         self.tokenizer = AutoTokenizer.from_pretrained(
-            fname_tokenizer,
-            cache_dir=fname_tokenizer,
+            base_path,
+            cache_dir=base_path,
             local_files_only=True,
         )
+        assert self.tokenizer.is_fast  # assume tokenizer.json is used
 
         # Initialize lists and dictionaries for added tokens
         self.added_tokens_list = []
@@ -506,8 +575,7 @@ class HfVocab:
         self.vocab_size_base = self.tokenizer.vocab_size
         self.vocab_size      = self.vocab_size_base + len(self.added_tokens_list)
 
-        self.fname_tokenizer    = fname_tokenizer
-        self.fname_added_tokens = fname_added_tokens
+        self.fname_tokenizer = fname_tokenizer
 
     def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         reverse_vocab = {
@@ -559,18 +627,7 @@ class HfVocab:
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
-        return f"<HfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
-
-
-class NoVocab:
-    tokenizer_model = "no_vocab"
-    name = "no_vocab"
-
-    def __repr__(self) -> str:
-        return "<NoVocab for a model without integrated vocabulary>"
-
-
-Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab | NoVocab"
+        return f"<LlamaHfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
 #
@@ -588,7 +645,7 @@ def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
             .reshape(weights.shape))
 
 
-class Tensor(metaclass=ABCMeta):
+class Tensor(ABC):
     data_type: DataType
 
     @abstractmethod
@@ -610,7 +667,7 @@ def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
 
 
 class UnquantizedTensor(Tensor):
-    def __init__(self, ndarray: NDArray) -> None:
+    def __init__(self, ndarray: NDArray):
         assert isinstance(ndarray, np.ndarray)
         self.ndarray = ndarray
         self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
@@ -689,7 +746,7 @@ class ModelPlus:
     model: LazyModel
     paths: list[Path]  # Where this was read from.
     format: Literal['ggml', 'torch', 'safetensors', 'none']
-    vocab: Vocab | None  # For GGML models (which have vocab built in), the vocab.
+    vocab: BaseVocab | None  # For GGML models (which have vocab built in), the vocab.
 
 
 def merge_sharded(models: list[LazyModel]) -> LazyModel:
@@ -698,7 +755,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel:
     names = {name: None for model in models for name in model}
 
     def convert(name: str) -> LazyTensor:
-        lazy_tensors: list[LazyTensor] = [model[name] for model in models]
+        lazy_tensors = [model[name] for model in models]
         if len(lazy_tensors) == 1:
             # only one file; don't go through this procedure since there might
             # be quantized tensors
@@ -719,7 +776,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel:
 
         def load() -> UnquantizedTensor:
             ndarrays = [load_unquantized(tensor) for tensor in lazy_tensors]
-            concatenated: NDArray = np.concatenate(ndarrays, axis=axis)
+            concatenated = np.concatenate(ndarrays, axis=axis)
             return UnquantizedTensor(concatenated)
         description = 'concatenated[[' + '] | ['.join(lt.description for lt in lazy_tensors) + ']]'
         return LazyTensor(load, concatenated_shape, lazy_tensors[0].data_type, description)
@@ -771,6 +828,15 @@ def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     return LazyTensor(load, s, lazy_tensor.data_type, 'part ' + lazy_tensor.description)
 
 
+def pack_experts_lazy(lazy_tensors: list[LazyTensor]) -> LazyTensor:
+    def load() -> Tensor:
+        tensors = [lazy_tensor.load() for lazy_tensor in lazy_tensors]
+        return UnquantizedTensor(np.array([tensor.ndarray for tensor in tensors]))
+    s = lazy_tensors[0].shape.copy()
+    s.insert(0, len(lazy_tensors))
+    return LazyTensor(load, s, lazy_tensors[0].data_type, 'pack_experts ' + ' | '.join(lt.description for lt in lazy_tensors))
+
+
 # Functionality that simulates `torch.load` but where individual tensors are
 # only loaded into memory on demand, not all at once.
 # PyTorch can't do this natively as of time of writing:
@@ -807,10 +873,10 @@ class LazyUnpickler(pickle.Unpickler):
 
         def load(offset: int, elm_count: int) -> NDArray:
             dtype = data_type.dtype
-            fp = self.zip_file.open(info)
-            fp.seek(offset * dtype.itemsize)
-            size = elm_count * dtype.itemsize
-            data = fp.read(size)
+            with self.zip_file.open(info) as fp:
+                fp.seek(offset * dtype.itemsize)
+                size = elm_count * dtype.itemsize
+                data = fp.read(size)
             assert len(data) == size
             return np.frombuffer(data, dtype)
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
@@ -831,7 +897,7 @@ class LazyUnpickler(pickle.Unpickler):
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES: dict[tuple[str, str], Any] = {
+    CLASSES = {
         # getattr used here as a workaround for mypy not being smart enough to determine
         # the staticmethods have a __func__ attribute.
         ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
@@ -890,7 +956,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
 def must_read(fp: IO[bytes], length: int) -> bytes:
     ret = fp.read(length)
     if len(ret) < length:
-        raise Exception("unexpectedly reached end of file")
+        raise EOFError("unexpectedly reached end of file")
     return ret
 
 
@@ -948,13 +1014,14 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
             yield result
 
 
-def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None:
+def check_vocab_size(params: Params, vocab: BaseVocab, pad_vocab: bool = False) -> None:
     # Handle special case where the model's vocab size is not set
     if params.n_vocab == -1:
         raise ValueError(
-            f"The model's vocab size is set to -1 in params.json. Please update it manually.{f' Maybe {vocab.vocab_size}?' if hasattr(vocab, 'vocab_size') else ''}"
+            "The model's vocab size is set to -1 in params.json. Please update it manually."
+            + (f" Maybe {vocab.vocab_size}?" if isinstance(vocab, Vocab) else ""),
         )
-    if isinstance(vocab, NoVocab):
+    if not isinstance(vocab, Vocab):
         return  # model has no vocab
 
     # Check for a vocab size mismatch
@@ -979,11 +1046,11 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N
     if vocab.vocab_size < params.n_vocab:
         msg += " Add the --pad-vocab option and try again."
 
-    raise Exception(msg)
+    raise ValueError(msg)
 
 
 class OutputFile:
-    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
+    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
     def add_meta_arch(self, params: Params) -> None:
@@ -1034,8 +1101,6 @@ class OutputFile:
             self.gguf.add_file_type(params.ftype)
 
     def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
-        assert not isinstance(vocab, NoVocab)
-
         tokens = []
         scores = []
         toktypes = []
@@ -1135,7 +1200,7 @@ class OutputFile:
 
     @staticmethod
     def write_all(
-        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab,
+        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
         concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
     ) -> None:
@@ -1145,11 +1210,11 @@ class OutputFile:
 
         # meta data
         of.add_meta_arch(params)
-        if isinstance(vocab, NoVocab):
-            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
-        else:
+        if isinstance(vocab, Vocab):
             of.add_meta_vocab(vocab)
             of.add_meta_special_vocab(svocab)
+        else:  # NoVocab
+            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -1176,7 +1241,7 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
 
     name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
 
-    raise Exception(f"Unexpected combination of types: {name_to_type}")
+    raise ValueError(f"Unexpected combination of types: {name_to_type}")
 
 
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
@@ -1186,10 +1251,26 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
 
 def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
-    should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
+    should_skip = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
     tmp = model
 
+    # merge experts into one tensor
+    if params.n_experts and params.n_experts > 0:
+        for i_l in range(params.n_layer):
+            for w in range(1, 4):
+                experts = []
+                for e in range(params.n_experts):
+                    if f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight" in model:
+                        experts.append(model[f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight"])
+                        del tmp[f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight"]
+                    elif f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight" in model:
+                        experts.append(model[f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight"])
+                        del tmp[f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight"]
+                    else:
+                        raise ValueError(f"Expert tensor not found: layers.{i_l}.feed_forward.experts.{e}.w{w}.weight")
+                tmp[f"layers.{i_l}.feed_forward.experts.w{w}.weight"] = pack_experts_lazy(experts)
+
     # HF models permut or pack some of the tensors, so we need to undo that
     for i in itertools.count():
         if f"model.layers.{i}.self_attn.q_proj.weight" in model:
@@ -1213,8 +1294,7 @@ def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) ->
             if skip_unknown:
                 print(f"Unexpected tensor name: {name} - skipping")
                 continue
-            else:
-                raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
+            raise ValueError(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
 
         if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
@@ -1231,7 +1311,7 @@ def nth_multifile_path(path: Path, n: int) -> Path | None:
     the nth path in the model.
     '''
     # Support the following patterns:
-    patterns: list[tuple[str, str]] = [
+    patterns = [
         # - x.00.pth, x.01.pth, etc.
         (r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'),
         # - x-00001-of-00002.bin, x-00002-of-00002.bin, etc.
@@ -1277,9 +1357,9 @@ 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 Exception(f"Can't find model in directory {path}")
+            raise FileNotFoundError(f"Can't find model in directory {path}")
         if len(files) > 1:
-            raise Exception(f"Found multiple models in {path}, not sure which to pick: {files}")
+            raise ValueError(f"Found multiple models in {path}, not sure which to pick: {files}")
         path = files[0]
 
     paths = find_multifile_paths(path)
@@ -1293,36 +1373,14 @@ def load_some_model(path: Path) -> ModelPlus:
 
 
 class VocabFactory:
-    _FILES = {"spm": "tokenizer.model", "bpe": "vocab.json", "hfft": "tokenizer.json"}
+    _VOCAB_CLASSES: list[type[Vocab]] = [SentencePieceVocab, BpeVocab, LlamaHfVocab]
 
     def __init__(self, path: Path):
         self.path = path
-        self.file_paths = self._detect_files()
-        print(f"Found vocab files: {self.file_paths}")
 
-    def _detect_files(self) -> dict[str, Path | None]:
-        def locate(file: str) -> Path | None:
-            if (path := self.path / file).exists():
-                return path
-            if (path := self.path.parent / file).exists():
-                return path
-            return None
-
-        return {vt: locate(f) for vt, f in self._FILES.items()}
-
-    def _select_file(self, vocab_types: list[str]) -> tuple[str, Path]:
-        for vtype in vocab_types:
-            try:
-                path = self.file_paths[vtype]
-            except KeyError:
-                raise ValueError(f"Unsupported vocabulary type {vtype}") from None
-            if path is not None:
-                return vtype, path
-        raise FileNotFoundError(f"Could not find any of {[self._FILES[vt] for vt in vocab_types]}")
-
-    def _create_special_vocab(self, vocab: Vocab, model_parent_path: Path) -> gguf.SpecialVocab:
+    def _create_special_vocab(self, vocab: BaseVocab, model_parent_path: Path) -> gguf.SpecialVocab:
         load_merges = vocab.name == "bpe"
-        n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
+        n_vocab = vocab.vocab_size if isinstance(vocab, Vocab) else None
         return gguf.SpecialVocab(
             model_parent_path,
             load_merges=load_merges,
@@ -1331,27 +1389,29 @@ class VocabFactory:
         )
 
     def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
-        vocab_type, path = self._select_file(vocab_types)
-        print(f"Loading vocab file {path!r}, type {vocab_type!r}")
+        vocab_classes: dict[str, type[Vocab]] = {cls.name: cls for cls in self._VOCAB_CLASSES}
+        selected_vocabs: dict[str, type[Vocab]] = {}
+        for vtype in vocab_types:
+            try:
+                selected_vocabs[vtype] = vocab_classes[vtype]
+            except KeyError:
+                raise ValueError(f"Unsupported vocabulary type {vtype}") from None
 
-        added_tokens_path = path.parent / "added_tokens.json"
-        if vocab_type == "bpe":
-            return BpeVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
-        if vocab_type == "spm":
-            return SentencePieceVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
-        if vocab_type == "hfft":
-            return HfVocab(
-                path.parent, added_tokens_path if added_tokens_path.exists() else None
-            )
-        raise ValueError(vocab_type)
+        for vtype, cls in selected_vocabs.items():
+            try:
+                vocab = cls(self.path)
+                break
+            except FileNotFoundError:
+                pass  # ignore unavailable tokenizers
+        else:
+            raise FileNotFoundError(f"Could not find a tokenizer matching any of {vocab_types}")
 
-    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
-        vocab: Vocab
-        if len(vocab_types) == 1 and "no_vocab" in vocab_types:
+        print(f"Loaded vocab file {vocab.fname_tokenizer!r}, type {vocab.name!r}")
+        return vocab
+
+    def load_vocab(self, vocab_types: list[str] | None, model_parent_path: Path) -> tuple[BaseVocab, gguf.SpecialVocab]:
+        vocab: BaseVocab
+        if vocab_types is None:
             vocab = NoVocab()
         else:
             vocab = self._create_vocab_by_path(vocab_types)
@@ -1408,10 +1468,8 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
-    if args.no_vocab:
-        if args.vocab_only:
-            raise ValueError("no need to specify --vocab-only if using --no-vocab")
-        args.vocab_type = "no_vocab"
+    if args.no_vocab and args.vocab_only:
+        raise ValueError("--vocab-only does not make sense with --no-vocab")
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
@@ -1433,10 +1491,12 @@ def main(args_in: list[str] | None = None) -> None:
     params = Params.load(model_plus)
     if params.n_ctx == -1:
         if args.ctx is None:
-            raise Exception("The model doesn't have a context size, and you didn't specify one with --ctx\n"
-                            "Please specify one with --ctx:\n"
-                            " - LLaMA v1: --ctx 2048\n"
-                            " - LLaMA v2: --ctx 4096\n")
+            msg = """\
+                The model doesn't have a context size, and you didn't specify one with --ctx
+                Please specify one with --ctx:
+                 - LLaMA v1: --ctx 2048
+                 - LLaMA v2: --ctx 4096"""
+            parser.error(textwrap.dedent(msg))
         params.n_ctx = args.ctx
 
     if args.outtype:
@@ -1451,9 +1511,11 @@ def main(args_in: list[str] | None = None) -> None:
     model_parent_path = model_plus.paths[0].parent
     vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
     vocab_factory = VocabFactory(vocab_path)
-    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type.split(","), model_parent_path)
+    vocab_types = None if args.no_vocab else args.vocab_type.split(",")
+    vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
 
     if args.vocab_only:
+        assert isinstance(vocab, Vocab)
         if not args.outfile:
             raise ValueError("need --outfile if using --vocab-only")
         outfile = args.outfile

examples/batched-bench/README.md

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

examples/batched-bench/batched-bench.cpp

@@ -32,13 +32,15 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
         printf("  <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
-        printf("  example: %s ggml-model-f16.gguf 2048 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
+        printf("  example: %s ggml-model-f16.gguf 2048 2048 512 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
         return 1 ;
     }
 
     int n_kv_max     = 2048;
+    int n_batch      = 2048;
+    int n_ubatch     = 512;
     int is_pp_shared = 0;
     int n_gpu_layers = 0;
 
@@ -56,23 +58,31 @@ int main(int argc, char ** argv) {
     }
 
     if (argc >= 4) {
-        is_pp_shared = std::atoi(argv[3]);
+        n_batch = std::atoi(argv[3]);
     }
 
     if (argc >= 5) {
-        n_gpu_layers = std::atoi(argv[4]);
+        n_ubatch = std::atoi(argv[4]);
     }
 
     if (argc >= 6) {
-        n_pp = parse_list(argv[5]);
+        is_pp_shared = std::atoi(argv[5]);
     }
 
     if (argc >= 7) {
-        n_tg = parse_list(argv[6]);
+        n_gpu_layers = std::atoi(argv[6]);
     }
 
     if (argc >= 8) {
-        n_pl = parse_list(argv[7]);
+        n_pp = parse_list(argv[7]);
+    }
+
+    if (argc >= 9) {
+        n_tg = parse_list(argv[8]);
+    }
+
+    if (argc >= 10) {
+        n_pl = parse_list(argv[9]);
     }
 
     // init LLM
@@ -100,7 +110,8 @@ int main(int argc, char ** argv) {
 
     ctx_params.seed      = 1234;
     ctx_params.n_ctx     = n_kv_max;
-    ctx_params.n_batch   = 512;
+    ctx_params.n_batch   = n_batch;
+    ctx_params.n_ubatch  = n_ubatch;
 
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
@@ -158,7 +169,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, n_batch, n_ubatch, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");

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

@@ -880,7 +880,7 @@ int main(int argc, char ** argv) {
     TransformerWeights weights = {};
     {
         LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
-        FILE *file = fopen(params.fn_llama2c_model, "r");
+        FILE * file = fopen(params.fn_llama2c_model, "rb");
         if (!file) {
             LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
             return 1;

examples/embedding/embedding.cpp

@@ -178,25 +178,27 @@ int main(int argc, char ** argv) {
     float * out = emb + p * n_embd;
     batch_decode(ctx, batch, out, s, n_embd);
 
-    // print the first part of the embeddings
+    // print the first part of the embeddings or for a single prompt, the full embedding
     fprintf(stdout, "\n");
     for (int j = 0; j < n_prompts; j++) {
         fprintf(stdout, "embedding %d: ", j);
-        for (int i = 0; i < std::min(16, n_embd); i++) {
+        for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
             fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
         }
         fprintf(stdout, "\n");
     }
 
     // print cosine similarity matrix
-    fprintf(stdout, "\n");
-    printf("cosine similarity matrix:\n\n");
-    for (int i = 0; i < n_prompts; i++) {
-        for (int j = 0; j < n_prompts; j++) {
-            float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-            fprintf(stdout, "%6.2f ", sim);
-        }
+    if (n_prompts > 1) {
         fprintf(stdout, "\n");
+        printf("cosine similarity matrix:\n\n");
+        for (int i = 0; i < n_prompts; i++) {
+            for (int j = 0; j < n_prompts; j++) {
+                float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                fprintf(stdout, "%6.2f ", sim);
+            }
+            fprintf(stdout, "\n");
+        }
     }
 
     // clean up

examples/gbnf-validator/CMakeLists.txt

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

examples/gbnf-validator/gbnf-validator.cpp

@@ -0,0 +1,132 @@
+#define LLAMA_API_INTERNAL
+
+#include "grammar-parser.h"
+#include "ggml.h"
+#include "llama.h"
+#include "unicode.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+
+static bool llama_sample_grammar_string(struct llama_grammar * grammar, const std::string & input_str, size_t & error_pos, std::string & error_msg) {
+    auto decoded = decode_utf8(input_str, {});
+    const auto & code_points = decoded.first;
+
+    size_t pos = 0;
+    for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
+        auto prev_stacks = grammar->stacks;
+        grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
+        if (grammar->stacks.empty()) {
+            error_pos = pos;
+            error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'";
+            grammar->stacks = prev_stacks;
+            return false;
+        }
+        ++pos;
+    }
+
+    for (const auto & stack : grammar->stacks) {
+        if (stack.empty()) {
+            return true;
+        }
+    }
+
+    error_pos = pos;
+    error_msg = "Unexpected end of input";
+    return false;
+}
+
+static void print_error_message(const std::string & input_str, size_t error_pos, const std::string & error_msg) {
+    fprintf(stdout, "Input string is invalid according to the grammar.\n");
+    fprintf(stdout, "Error: %s at position %zu\n", error_msg.c_str(), error_pos);
+    fprintf(stdout, "\n");
+    fprintf(stdout, "Input string:\n");
+    fprintf(stdout, "%s", input_str.substr(0, error_pos).c_str());
+    if (error_pos < input_str.size()) {
+        fprintf(stdout, "\033[1;31m%c", input_str[error_pos]);
+        if (error_pos+1 < input_str.size()) {
+            fprintf(stdout, "\033[0;31m%s", input_str.substr(error_pos+1).c_str());
+        }
+        fprintf(stdout, "\033[0m\n");
+    }
+}
+
+int main(int argc, char** argv) {
+    if (argc != 3) {
+        fprintf(stdout, "Usage: %s <grammar_filename> <input_filename>\n", argv[0]);
+        return 1;
+    }
+
+    const std::string grammar_filename = argv[1];
+    const std::string input_filename = argv[2];
+
+    // Read the GBNF grammar file
+    FILE* grammar_file = fopen(grammar_filename.c_str(), "r");
+    if (!grammar_file) {
+        fprintf(stdout, "Failed to open grammar file: %s\n", grammar_filename.c_str());
+        return 1;
+    }
+
+    fseek(grammar_file, 0, SEEK_END);
+    size_t grammar_size = ftell(grammar_file);
+    fseek(grammar_file, 0, SEEK_SET);
+
+    std::string grammar_str(grammar_size, ' ');
+    fread(&grammar_str[0], 1, grammar_size, grammar_file);
+    fclose(grammar_file);
+
+    // Parse the GBNF grammar
+    auto parsed_grammar = grammar_parser::parse(grammar_str.c_str());
+
+    // will be empty (default) if there are parse errors
+    if (parsed_grammar.rules.empty()) {
+        fprintf(stdout, "%s: failed to parse grammar\n", __func__);
+        return 1;
+    }
+
+    // Ensure that there is a "root" node.
+    if (parsed_grammar.symbol_ids.find("root") == parsed_grammar.symbol_ids.end()) {
+        fprintf(stdout, "%s: grammar does not contain a 'root' symbol\n", __func__);
+        return 1;
+    }
+
+    std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+
+    // Create the LLAMA grammar
+    auto grammar = llama_grammar_init(
+            grammar_rules.data(),
+            grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+
+    // Read the input file
+    FILE* input_file = fopen(input_filename.c_str(), "r");
+    if (!input_file) {
+        fprintf(stdout, "Failed to open input file: %s\n", input_filename.c_str());
+        return 1;
+    }
+
+    fseek(input_file, 0, SEEK_END);
+    size_t input_size = ftell(input_file);
+    fseek(input_file, 0, SEEK_SET);
+
+    std::string input_str(input_size, ' ');
+    fread(&input_str[0], 1, input_size, input_file);
+    fclose(input_file);
+
+    // Validate the input string against the grammar
+    size_t error_pos;
+    std::string error_msg;
+    bool is_valid = llama_sample_grammar_string(grammar, input_str, error_pos, error_msg);
+
+    if (is_valid) {
+        fprintf(stdout, "Input string is valid according to the grammar.\n");
+    } else {
+        print_error_message(input_str, error_pos, error_msg);
+    }
+
+    // Clean up
+    llama_grammar_free(grammar);
+
+    return 0;
+}

examples/gguf-split/gguf-split.cpp

@@ -28,9 +28,11 @@ enum split_operation : uint8_t {
 
 struct split_params {
     split_operation operation = SPLIT_OP_SPLIT;
+    size_t n_bytes_split = 0;
     int n_split_tensors = 128;
     std::string input;
     std::string output;
+    bool dry_run = false;
 };
 
 static void split_print_usage(const char * executable) {
@@ -41,15 +43,36 @@ static void split_print_usage(const char * executable) {
     printf("Apply a GGUF operation on IN to OUT.");
     printf("\n");
     printf("options:\n");
-    printf("  -h, --help            show this help message and exit\n");
-    printf("  --version             show version and build info\n");
-    printf("  --split               split GGUF to multiple GGUF (default)\n");
-    printf("  --split-max-tensors   max tensors in each split: default(%d)\n", default_params.n_split_tensors);
-    printf("  --merge               merge multiple GGUF to a single GGUF\n");
+    printf("  -h, --help              show this help message and exit\n");
+    printf("  --version               show version and build info\n");
+    printf("  --split                 split GGUF to multiple GGUF (enabled by default)\n");
+    printf("  --merge                 merge multiple GGUF to a single GGUF\n");
+    printf("  --split-max-tensors     max tensors in each split (default: %d)\n", default_params.n_split_tensors);
+    printf("  --split-max-size N(M|G) max size per split\n");
+    printf("  --dry-run               only print out a split plan and exit, without writing any new files\n");
     printf("\n");
 }
 
-static bool split_params_parse_ex(int argc, const char ** argv, split_params & params) {
+// return convert string, for example "128M" or "4G" to number of bytes
+static size_t split_str_to_n_bytes(std::string str) {
+    size_t n_bytes = 0;
+    int n;
+    if (str.back() == 'M') {
+        sscanf(str.c_str(), "%d", &n);
+        n_bytes = n * 1024 * 1024; // megabytes
+    } else if (str.back() == 'G') {
+        sscanf(str.c_str(), "%d", &n);
+        n_bytes = n * 1024 * 1024 * 1024; // gigabytes
+    } else {
+        throw std::invalid_argument("error: supported units are M (megabytes) or G (gigabytes), but got: " + std::string(1, str.back()));
+    }
+    if (n <= 0) {
+        throw std::invalid_argument("error: size must be a positive value");
+    }
+    return n_bytes;
+}
+
+static void split_params_parse_ex(int argc, const char ** argv, split_params & params) {
     std::string arg;
     const std::string arg_prefix = "--";
     bool invalid_param = false;
@@ -62,6 +85,8 @@ static bool split_params_parse_ex(int argc, const char ** argv, split_params & p
         }
 
         bool arg_found = false;
+        bool is_op_set = false;
+        bool is_mode_set = false;
         if (arg == "-h" || arg == "--help") {
             split_print_usage(argv[0]);
             exit(0);
@@ -71,23 +96,46 @@ static bool split_params_parse_ex(int argc, const char ** argv, split_params & p
             fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
             exit(0);
         }
+        if (arg == "--dry-run") {
+            arg_found = true;
+            params.dry_run = true;
+        }
 
+        if (is_op_set) {
+            throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
+        }
         if (arg == "--merge") {
             arg_found = true;
+            is_op_set = true;
             params.operation = SPLIT_OP_MERGE;
         }
         if (arg == "--split") {
             arg_found = true;
+            is_op_set = true;
             params.operation = SPLIT_OP_SPLIT;
         }
+
+        if (is_mode_set) {
+            throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
+        }
         if (arg == "--split-max-tensors") {
             if (++arg_idx >= argc) {
                 invalid_param = true;
                 break;
             }
             arg_found = true;
+            is_mode_set = true;
             params.n_split_tensors = atoi(argv[arg_idx]);
         }
+        if (arg == "--split-max-size") {
+            if (++arg_idx >= argc) {
+                invalid_param = true;
+                break;
+            }
+            arg_found = true;
+            is_mode_set = true;
+            params.n_bytes_split = split_str_to_n_bytes(argv[arg_idx]);
+        }
 
         if (!arg_found) {
             throw std::invalid_argument("error: unknown argument: " + arg);
@@ -99,24 +147,17 @@ static bool split_params_parse_ex(int argc, const char ** argv, split_params & p
     }
 
     if (argc - arg_idx < 2) {
-        printf("%s: bad arguments\n", argv[0]);
-        split_print_usage(argv[0]);
-        return false;
+        throw std::invalid_argument("error: bad arguments");
     }
 
     params.input = argv[arg_idx++];
     params.output = argv[arg_idx++];
-
-    return true;
 }
 
 static bool split_params_parse(int argc, const char ** argv, split_params & params) {
     bool result = true;
     try {
-        if (!split_params_parse_ex(argc, argv, params)) {
-            split_print_usage(argv[0]);
-            exit(EXIT_FAILURE);
-        }
+        split_params_parse_ex(argc, argv, params);
     }
     catch (const std::invalid_argument & ex) {
         fprintf(stderr, "%s\n", ex.what());
@@ -140,15 +181,11 @@ struct split_strategy {
     struct ggml_context * ctx_meta = NULL;
     const int n_tensors;
 
-    const int n_split;
-    int i_split = 0;
+    // one ctx_out per one output file
+    std::vector<struct gguf_context *> ctx_outs;
 
-    int i_tensor = 0;
-
-    std::vector<uint8_t> read_data;
-
-    struct gguf_context * ctx_out;
-    std::ofstream fout;
+    // temporary buffer for reading in tensor data
+    std::vector<uint8_t> read_buf;
 
     split_strategy(const split_params & params,
             std::ifstream & f_input,
@@ -158,79 +195,141 @@ struct split_strategy {
         f_input(f_input),
         ctx_gguf(ctx_gguf),
         ctx_meta(ctx_meta),
-        n_tensors(gguf_get_n_tensors(ctx_gguf)),
-        n_split(std::ceil(1. * n_tensors / params.n_split_tensors)) {
+        n_tensors(gguf_get_n_tensors(ctx_gguf)) {
+
+        // because we need to know list of tensors for each file in advance, we will build all the ctx_out for all output splits
+        int i_split = -1;
+        struct gguf_context * ctx_out = NULL;
+        auto new_ctx_out = [&]() {
+            i_split++;
+            if (ctx_out != NULL) {
+                if (gguf_get_n_tensors(ctx_out) == 0) {
+                    fprintf(stderr, "error: one of splits have 0 tensors. Maybe size or tensors limit is too small\n");
+                    exit(EXIT_FAILURE);
+                }
+                ctx_outs.push_back(ctx_out);
+            }
+            ctx_out = gguf_init_empty();
+            // Save all metadata in first split only
+            if (i_split == 0) {
+                gguf_set_kv(ctx_out, ctx_gguf);
+            }
+            gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_NO, i_split);
+            gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_COUNT, 0); // placeholder
+            gguf_set_val_i32(ctx_out, LLM_KV_SPLIT_TENSORS_COUNT, n_tensors);
+        };
+
+        // initialize ctx_out for the first split
+        new_ctx_out();
+
+        // process tensors one by one
+        size_t curr_tensors_size = 0; // current size by counting only tensors size (without metadata)
+        for (int i = 0; i < n_tensors; ++i) {
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
+            // calculate the "imaginary" size = the current size + next tensor size
+            size_t n_bytes = GGML_PAD(ggml_nbytes(t), GGUF_DEFAULT_ALIGNMENT);
+            size_t next_tensors_size = curr_tensors_size + n_bytes;
+            if (should_split(i, next_tensors_size)) {
+                new_ctx_out();
+                curr_tensors_size = n_bytes;
+            } else {
+                curr_tensors_size = next_tensors_size;
+            }
+            gguf_add_tensor(ctx_out, t);
         }
 
-    bool should_split() const {
-        return i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
+        // push the last ctx_out
+        ctx_outs.push_back(ctx_out);
+
+        // set the correct n_split for all ctx_out
+        for (auto & ctx : ctx_outs) {
+            gguf_set_val_u16(ctx, LLM_KV_SPLIT_COUNT, ctx_outs.size());
+        }
     }
 
-    void split_start() {
-        ctx_out = gguf_init_empty();
-
-        // Save all metadata in first split only
-        if (i_split == 0) {
-            gguf_set_kv(ctx_out, ctx_gguf);
+    ~split_strategy() {
+        for (auto & ctx_out : ctx_outs) {
+            gguf_free(ctx_out);
         }
-        gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_NO, i_split);
-        gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_COUNT, n_split);
-        gguf_set_val_i32(ctx_out, LLM_KV_SPLIT_TENSORS_COUNT, n_tensors);
-
-        // populate the original tensors, so we get an initial metadata
-        for (int i = i_split * params.n_split_tensors; i < n_tensors && i < (i_split + 1) * params.n_split_tensors; ++i) {
-            struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
-            gguf_add_tensor(ctx_out, meta);
-        }
-
-        char split_path[PATH_MAX] = {0};
-        llama_split_path(split_path, sizeof(split_path), params.output.c_str(), i_split, n_split);
-
-        fprintf(stderr, "%s: %s ...", __func__, split_path);
-        fout = std::ofstream(split_path, std::ios::binary);
-        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
-
-        auto meta_size = gguf_get_meta_size(ctx_out);
-
-        // placeholder for the meta data
-        ::zeros(fout, meta_size);
-
-        i_split++;
     }
 
-    void next_tensor() {
-        const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
-        struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
-        auto n_bytes = ggml_nbytes(t);
-
-        if (read_data.size() < n_bytes) {
-            read_data.resize(n_bytes);
+    bool should_split(int i_tensor, size_t next_size) {
+        if (params.n_bytes_split > 0) {
+            // split by max size per file
+            return next_size > params.n_bytes_split;
+        } else {
+            // split by number of tensors per file
+            return i_tensor > 0 && i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
         }
-
-        auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
-        f_input.seekg(offset);
-        f_input.read((char *)read_data.data(), n_bytes);
-
-        t->data = read_data.data();
-
-        // write tensor data + padding
-        fout.write((const char *)t->data, n_bytes);
-        zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
-
-        i_tensor++;
     }
 
-    void split_end() {
-        // go back to beginning of file and write the updated metadata
-        fout.seekp(0);
-        std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
-        gguf_get_meta_data(ctx_out, data.data());
-        fout.write((const char *)data.data(), data.size());
+    void print_info() {
+        printf("n_split: %ld\n", ctx_outs.size());
+        int i_split = 0;
+        for (auto & ctx_out : ctx_outs) {
+            // re-calculate the real gguf size for each split (= metadata size + total size of all tensors)
+            size_t total_size = gguf_get_meta_size(ctx_out);
+            for (int i = 0; i < gguf_get_n_tensors(ctx_out); ++i) {
+                struct ggml_tensor * t = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_out, i));
+                total_size += ggml_nbytes(t);
+            }
+            total_size = total_size / 1024 / 1024; // convert to megabytes
+            printf("split %05d: n_tensors = %d, total_size = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
+            i_split++;
+        }
+    }
 
-        fout.close();
-        gguf_free(ctx_out);
+    void write() {
+        int i_split = 0;
+        int n_split = ctx_outs.size();
+        for (auto & ctx_out : ctx_outs) {
+            // construct file path
+            char split_path[PATH_MAX] = {0};
+            llama_split_path(split_path, sizeof(split_path), params.output.c_str(), i_split, n_split);
 
-        fprintf(stderr, "\033[3Ddone\n");
+            // open the output file
+            printf("Writing file %s ... ", split_path);
+            fflush(stdout);
+            std::ofstream fout = std::ofstream(split_path, std::ios::binary);
+            fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+            // write metadata
+            std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+            gguf_get_meta_data(ctx_out, data.data());
+            fout.write((const char *)data.data(), data.size());
+
+            // write tensors
+            for (int i = 0; i < gguf_get_n_tensors(ctx_out); ++i) {
+                // read tensor meta and prepare buffer
+                const char * t_name = gguf_get_tensor_name(ctx_out, i);
+                struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+                auto n_bytes = ggml_nbytes(t);
+                read_buf.resize(n_bytes);
+
+                // calculate offset
+                auto i_tensor_in = gguf_find_tensor(ctx_gguf, t_name); // idx of tensor in the input file
+                auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in);
+
+                // copy tensor from input to output file
+                copy_file_to_file(f_input, fout, offset, n_bytes);
+                zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+            }
+
+            printf("done\n");
+            // close the file
+            fout.close();
+            i_split++;
+        }
+    }
+
+    void copy_file_to_file(std::ifstream & f_in, std::ofstream & f_out, const size_t in_offset, const size_t len) {
+        // TODO: detect OS and use copy_file_range() here for better performance
+        if (read_buf.size() < len) {
+            read_buf.resize(len);
+        }
+        f_in.seekg(in_offset);
+        f_in.read((char *)read_buf.data(), len);
+        f_out.write((const char *)read_buf.data(), len);
     }
 };
 
@@ -254,32 +353,22 @@ static void gguf_split(const split_params & split_params) {
         exit(EXIT_FAILURE);
     }
 
+    // prepare the strategy
     split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
+    int n_split = strategy.ctx_outs.size();
+    strategy.print_info();
 
-    char first_split_path[PATH_MAX] = {0};
-    llama_split_path(first_split_path, sizeof(first_split_path),
-                     split_params.output.c_str(), strategy.i_split, strategy.n_split);
-    fprintf(stderr, "%s: %s -> %s (%d tensors per file)\n",
-            __func__, split_params.input.c_str(),
-            first_split_path,
-            split_params.n_split_tensors);
-
-    strategy.split_start();
-
-    while (strategy.i_tensor < strategy.n_tensors) {
-        strategy.next_tensor();
-        if (strategy.should_split()) {
-            strategy.split_end();
-            strategy.split_start();
-        }
+    if (!split_params.dry_run) {
+        // write all output splits
+        strategy.write();
     }
-    strategy.split_end();
 
+    // done, clean up
     gguf_free(ctx_gguf);
     f_input.close();
 
     fprintf(stderr, "%s: %d gguf split written with a total of %d tensors.\n",
-            __func__, strategy.n_split, strategy.n_tensors);
+            __func__, n_split, strategy.n_tensors);
 }
 
 static void gguf_merge(const split_params & split_params) {
@@ -448,10 +537,6 @@ static void gguf_merge(const split_params & split_params) {
 }
 
 int main(int argc, const char ** argv) {
-    if (argc < 3) {
-        split_print_usage(argv[0]);
-    }
-
     split_params params;
     split_params_parse(argc, argv, params);
 

examples/imatrix/imatrix.cpp

@@ -98,35 +98,38 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 
     const float * data = is_host ? (const float *) src1->data : m_src1_data.data();
 
+    // this has been adapted to the new format of storing merged experts in a single 3d tensor
+    // ref: https://github.com/ggerganov/llama.cpp/pull/6387
     if (t->op == GGML_OP_MUL_MAT_ID) {
         const int idx  = ((int32_t *) t->op_params)[0];
-        const int n_as = ((int32_t *) t->op_params)[1];
+        const ggml_tensor * ids = t->src[2];
+        const int n_as = src0->ne[2];
 
-        // the top-k selected expert ids are stored in the src0 tensor
-        // for simplicity, always copy src0 to host, because it is small
-        // take into account that src0 is not contiguous!
-        GGML_ASSERT(src0->ne[1] == src1->ne[1]);
-        GGML_ASSERT(n_as*ggml_nrows(src0)*sizeof(int) == GGML_PAD(ggml_nbytes(src0), n_as*sizeof(int)));
-        m_ids.resize(ggml_nbytes(src0)/sizeof(int));
-        ggml_backend_tensor_get(src0, m_ids.data(), 0, ggml_nbytes(src0));
+        // the top-k selected expert ids are stored in the ids tensor
+        // for simplicity, always copy ids to host, because it is small
+        // take into account that ids is not contiguous!
+        GGML_ASSERT(ids->ne[1] == src1->ne[1]);
+        GGML_ASSERT(n_as*ggml_nrows(ids)*sizeof(int) == GGML_PAD(ggml_nbytes(ids), n_as*sizeof(int)));
+        m_ids.resize(ggml_nbytes(ids)/sizeof(int));
+        ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
+
+        auto & e = m_stats[wname];
+
+        ++e.ncall;
+        // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
+        //       using the following line, we can correct for that if needed by replacing the line above with:
+        //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
 
         // loop over all possible experts, regardless if they are used or not in the batch
-        // this is necessary to guarantee equal number of "ncall" for each tensor
         for (int ex = 0; ex < n_as; ++ex) {
-            src0 = t->src[2 + ex];
-            wname = filter_tensor_name(src0->name);
-            auto& e = m_stats[wname];
+            size_t e_start = ex*src1->ne[0];
             if (e.values.empty()) {
-                e.values.resize(src1->ne[0], 0);
+                e.values.resize(src1->ne[0]*n_as, 0);
             }
-            else if (e.values.size() != (size_t)src1->ne[0]) {
-                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
+            else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
+                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
                 exit(1); //GGML_ASSERT(false);
             }
-            // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
-            //       using the following line, we can correct for that if needed
-            //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
-            ++e.ncall;
             if (m_params.verbosity > 1) {
                 printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
             }
@@ -136,7 +139,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
                 if (excur != ex) continue;
                 const float * x = data + row * src1->ne[0];
                 for (int j = 0; j < (int)src1->ne[0]; ++j) {
-                    e.values[j] += x[j]*x[j];
+                    e.values[e_start + j] += x[j]*x[j];
                 }
             }
             if (e.ncall > m_last_call) {
@@ -424,6 +427,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
+            // TODO: use batch.logits to save computations instead of relying on logits_all == true
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return false;

examples/llava/MobileVLM-README.md

@@ -6,7 +6,7 @@ for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com
 
 The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava.
 
-Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion  is a little different. Therefore, using MobiVLM as an example, the different conversion step will be shown.
+Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion is a little different. Therefore, using **MobileVLM-1.7B** as an example, the different conversion step will be shown.
 
 ## Usage
 Build with cmake or run `make llava-cli` to build it.
@@ -36,7 +36,7 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B
 ```
 
-3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` (for **V2** the arg is `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
+3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
 
 ```sh
 python ./examples/llava/convert-image-encoder-to-gguf \
@@ -78,7 +78,7 @@ cd examples/llava/android/build_64
 ### run on Android
 refer to `android/adb_run.sh`, modify resources' `name` and `path`
 
-## some result on Android with `Snapdragon 888` chip
+## Some result on Android with `Snapdragon 888` chip
 ### case 1
 **input**
 ```sh
@@ -109,7 +109,6 @@ llama_print_timings:       total time =   34731.93 ms
     --image /data/local/tmp/cat.jpeg \
     -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:"
 ```
-
 **output**
 ```sh
 encode_image_with_clip: image encoded in 21149.51 ms by CLIP (  146.87 ms per image patch)
@@ -121,12 +120,82 @@ llama_print_timings:        eval time =    1279.03 ms /    18 runs   (   71.06 m
 llama_print_timings:       total time =   34570.79 ms
 ```
 
+
+## Some result on Android with `Snapdragon 778G` chip
+### MobileVLM-1.7B case
+#### llava-cli release-b2005
+**input**
+```sh
+/data/local/tmp/llava-cli \
+    -m /data/local/tmp/ggml-model-q4_k.gguf \
+    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
+    -t 4 \
+    --image /data/local/tmp/many_llamas.jpeg \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat's that? ASSISTANT:"
+```
+**output**
+```sh
+encode_image_with_clip: image encoded in 18728.52 ms by CLIP (  130.06 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ A group of llamas are standing in a green pasture.
+
+llama_print_timings:        load time =   20357.33 ms
+llama_print_timings:      sample time =       2.96 ms /    14 runs   (    0.21 ms per token,  4734.53 tokens per second)
+llama_print_timings: prompt eval time =    8119.49 ms /   191 tokens (   42.51 ms per token,    23.52 tokens per second)
+llama_print_timings:        eval time =    1005.75 ms /    14 runs   (   71.84 ms per token,    13.92 tokens per second)
+llama_print_timings:       total time =   28038.34 ms /   205 tokens
+```
+#### llava-cli latest-version
+**input**
+
+Just the same as above.
+
+**output**(seems to be much slower)
+```sh
+encode_image_with_clip: image embedding created: 144 tokens
+
+encode_image_with_clip: image encoded in 288268.88 ms by CLIP ( 2001.87 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ It is a group of sheep standing together in a grass field.
+
+llama_print_timings:        load time =  818120.91 ms
+llama_print_timings:      sample time =       3.44 ms /    14 runs   (    0.25 ms per token,  4067.40 tokens per second)
+llama_print_timings: prompt eval time =  529274.69 ms /   191 tokens ( 2771.07 ms per token,     0.36 tokens per second)
+llama_print_timings:        eval time =   43894.02 ms /    13 runs   ( 3376.46 ms per token,     0.30 tokens per second)
+llama_print_timings:       total time =  865441.76 ms /   204 tokens
+```
+### MobileVLM_V2-1.7B case
+#### llava-cli release-2005b
+**input**
+
+Just the same as above.
+
+**output**
+```sh
+encode_image_with_clip: image encoded in 20609.61 ms by CLIP (  143.12 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ This image captures a lively scene of 20 llamas in motion on an expansive, grassy field. The llama is scattered across the landscape with some standing and others sitting down as if taking rest or observing their surroundings from different vantage points within this verdant setting.
+
+The background offers glimpses into a picturesque town nestled amidst hills under an overcast sky, adding depth to the scene while also emphasizing that distance between these llama and human-made structures like houses or roads in which they roam freely without any barriers around them. The image is framed by text at both right angles on white backgrounds against a contrasting blue backdrop with green foliage, further drawing attention to the llamas amidst their natural habitat while also inviting viewers into this picturesque landscape within town limits of Alta Llama
+
+llama_print_timings:        load time =   22406.77 ms
+llama_print_timings:      sample time =      49.26 ms /   186 runs   (    0.26 ms per token,  3776.27 tokens per second)
+llama_print_timings: prompt eval time =    9044.54 ms /   191 tokens (   47.35 ms per token,    21.12 tokens per second)
+llama_print_timings:        eval time =   14497.49 ms /   186 runs   (   77.94 ms per token,    12.83 tokens per second)
+llama_print_timings:       total time =   44411.01 ms /   377 tokens
+```
+
 ## Orin compile and run
 ### compile
 ```sh
 make LLAMA_CUDA=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
 ```
-
 ### run on Orin
 ### case 1
 **input**
@@ -175,8 +244,121 @@ llama_print_timings:        eval time =     166.65 ms /    11 runs   (   15.15 m
 llama_print_timings:       total time =    1365.47 ms /   243 tokens
 ```
 
-## Minor shortcomings
-The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quickly, we uniformly modified `clip_n_patches` function to a quarter. when counting the time consumption, the calculated time will be 4 times bigger than the real cost.
+## Running on Intel(R) Core(TM) i7-10750H
+### Operating system
+Ubuntu22.04
+### compile
+```sh
+make -j32
+```
+### MobileVLM-1.7B case
+**input**
+```sh
+-m /path/to/ggml-model-q4_k.gguf \
+    --mmproj /path/to/mmproj-model-f16.gguf \
+    --image /path/to/many_llamas.jpeg
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat's that? ASSISTANT:" \
+```
+**output**
+```sh
+encode_image_with_clip: image embedding created: 144 tokens
+
+encode_image_with_clip: image encoded in  2730.94 ms by CLIP (   18.96 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that?ASSISTANT:
+
+ A group of llamas are walking together in a field.
+
+llama_print_timings:        load time =    5506.60 ms
+llama_print_timings:      sample time =       0.44 ms /    13 runs   (    0.03 ms per token, 29545.45 tokens per second)
+llama_print_timings: prompt eval time =    2031.58 ms /   190 tokens (   10.69 ms per token,    93.52 tokens per second)
+llama_print_timings:        eval time =     438.92 ms /    12 runs   (   36.58 ms per token,    27.34 tokens per second)
+llama_print_timings:       total time =    5990.25 ms /   202 tokens
+```
+
+### MobileVLM_V2-1.7B case
+**input**
+
+Just the same as above.
+
+**ouput**
+```sh
+encode_image_with_clip: image embedding created: 144 tokens
+
+encode_image_with_clip: image encoded in  3223.89 ms by CLIP (   22.39 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that?ASSISTANT:
+
+ The image captures a tranquil scene in a park, where a group of approximately 20 llamas are gathered. The llamas, a mix of white and black, are standing in a line, their black and white patterns contrasting with the lush green grass of the park. The lamas are arranged in a line, suggesting a social order.
+
+The park itself is lush and green, with trees dotting the landscape in the background. A sign reading "Llamas Tico  Ana" is also visible in the image, possibly indicating the location or the breed of the llamas. The image seems to be taken from a distance, providing a wide view of the scene and the surrounding environment.
+
+The llamas' positions relative to each other, the sign, and the trees create a harmonious composition. The image does not contain any discernible text. The overall scene is one of peace and natural beauty, with the llamas in their natural habitat, surrounded by the vibrant colors and lush greenery of the park.
+
+llama_print_timings:        load time =    6642.61 ms
+llama_print_timings:      sample time =       8.15 ms /   223 runs   (    0.04 ms per token, 27358.61 tokens per second)
+llama_print_timings: prompt eval time =    2475.07 ms /   190 tokens (   13.03 ms per token,    76.77 tokens per second)
+llama_print_timings:        eval time =    8760.60 ms /   222 runs   (   39.46 ms per token,    25.34 tokens per second)
+llama_print_timings:       total time =   15513.95 ms /   412 tokens
+```
+
+## Run on Intel(R) Core(TM) Ultra7 115H
+### operation system
+Windows11
+### comiple
+```sh
+make -j32
+```
+### MobileVLM-1.7B case
+**input**
+```sh
+-m /path/to/ggml-model-q4_k.gguf \
+    --mmproj /path/to/tmp/mmproj-model-f16.gguf \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat's that? ASSISTANT:" \
+```
+**output**
+```sh
+encode_image_with_clip: image encoded in  4902.81 ms by CLIP (   34.05 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ The image features a group of brown and white llamas standing in a grassy field.
+
+llama_print_timings:        load time =    7441.06 ms
+llama_print_timings:      sample time =       0.72 ms /    19 runs   (    0.04 ms per token, 26279.39 tokens per second)
+llama_print_timings: prompt eval time =    2090.71 ms /   191 tokens (   10.95 ms per token,    91.36 tokens per second)
+llama_print_timings:        eval time =     512.35 ms /    18 runs   (   28.46 ms per token,    35.13 tokens per second)
+llama_print_timings:       total time =    7987.23 ms /   209 tokens
+```
+
+### MobileVLM_V2-1.7B case
+**input**
+
+Just the same as above.
+
+**output**
+```sh
+encode_image_with_clip: image encoded in  4682.44 ms by CLIP (   32.52 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ This image captures a lively scene of a group of 14 llamas in a grassy field. The llamas, with their distinctive black and white coats, are standing and walking in a line, seemingly engaged in a social activity. One
+ of them, possibly the first in the line, has its back turned, perhaps observing something in the distance.
+
+The llama in the front of the line stands out due to its black and white coloring, which is quite unusual for llama patterns. The llama in the front also seems to be more aware of its surroundings, as it faces the camera, giving a sense of engagement with the viewer.
+
+The image is taken from the side of the llama, providing a clear view of the llama in the front and its companions. The lameness in the llama in
+ front is not visible, indicating that it might not be the main focus of the photo.
+
+The background of the image features a grassy field, with a fence and a tree visible in the distance. The tree appears to be bare, suggesting that it might be during a time of year when most trees are dormant or have shed their leaves.
+
+
+llama_print_timings:        load time =    7015.35 ms
+llama_print_timings:      sample time =      10.61 ms /   256 runs   (    0.04 ms per token, 24119.09 tokens per second)
+llama_print_timings: prompt eval time =    2052.45 ms /   191 tokens (   10.75 ms per token,    93.06 tokens per second)
+llama_print_timings:        eval time =    7259.43 ms /   255 runs   (   28.47 ms per token,    35.13 tokens per second)
+llama_print_timings:       total time =   14371.19 ms /   446 tokens
+```
 
 ## TODO
 
@@ -191,5 +373,5 @@ The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quic
 
 ## contributor
 ```sh
-zhangjidong05, yangyang260, huyiming03, chenxiaotao03
+zhangjidong05, yangyang260, huyiming03, chenxiaotao03, ZiangWu-77
 ```

examples/llava/clip.cpp

@@ -835,9 +835,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
             // weight ne = [3, 3, 2048, 1]
             struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
-            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
             peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
             peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
             peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
             embeddings = peg_0;
         }
@@ -1755,7 +1756,7 @@ int clip_n_patches(const struct clip_ctx * ctx) {
 
     int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
 
-    if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
         n_patches /= 4;
     }
 

examples/main/README.md

@@ -296,7 +296,9 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Batch Size
 
--   `-b N, --batch-size N`: Set the batch size for prompt processing (default: 512). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
+-   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
+
+- `-ub N`, `--ubatch-size N`: physical maximum batch size. This is for pipeline parallelization. Default: `512`.
 
 ### Prompt Caching
 

examples/parallel/parallel.cpp

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

examples/perplexity/perplexity.cpp

@@ -380,6 +380,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
             const int batch_size  = std::min(end - batch_start, n_batch);
 
             //fprintf(stderr, "    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
+            // TODO: use llama_batch.logits instead of relying on logits_all == true
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 //fprintf(stderr, "%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
@@ -552,6 +553,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
+            int n_outputs = 0;
+
             batch.n_tokens = 0;
             for (int seq = 0; seq < n_seq_batch; seq++) {
                 int seq_start = batch_start + seq*n_ctx;
@@ -566,11 +569,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
                 for (int k = 0; k < batch_size; ++k) {
                     const int idx = seq*n_ctx + k;
-                    batch.token[idx] = tokens[seq_start + k];
-                    batch.pos[idx] = j*n_batch + k;
-                    batch.n_seq_id[idx] = 1;
-                    batch.seq_id[idx][0] = seq;
-                    batch.logits[idx] = batch.pos[idx] >= first ? 1 : 0;
+                    batch.token   [idx]    = tokens[seq_start + k];
+                    batch.pos     [idx]    = j*n_batch + k;
+                    batch.n_seq_id[idx]    = 1;
+                    batch.seq_id  [idx][0] = seq;
+                    batch.logits  [idx]    = batch.pos[idx] >= first ? 1 : 0;
+
+                    n_outputs += batch.logits[idx] != 0;
                 }
                 batch.n_tokens += batch_size;
 
@@ -583,9 +588,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
                 return {tokens, -1, logit_history, prob_history};
             }
 
-            if (num_batches > 1) {
+            if (num_batches > 1 && n_outputs > 0) {
                 const auto * batch_logits = llama_get_logits(ctx);
-                logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+                logits.insert(logits.end(), batch_logits, batch_logits + n_outputs * n_vocab);
             }
         }
 
@@ -604,14 +609,15 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         }
 
         for (int seq = 0; seq < n_seq_batch; seq++) {
-            const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx);
+            const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx + first);
+
             llama_token * tokens_data = tokens.data() + start + seq*n_ctx + first;
             if (!params.logits_file.empty()) {
-                process_logits(logits_stream, n_vocab, all_logits + first*n_vocab,
+                process_logits(logits_stream, n_vocab, all_logits,
                         tokens_data, n_ctx - 1 - first,
                         workers, log_probs, nll, nll2);
             } else {
-                process_logits(n_vocab, all_logits + first*n_vocab,
+                process_logits(n_vocab, all_logits,
                         tokens_data, n_ctx - 1 - first,
                         workers, nll, nll2,
                         logit_history.data() + start + seq*n_ctx + first,
@@ -652,6 +658,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 }
 
 static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<float> & batch_logits, int32_t n_batch, int32_t n_vocab) {
+    int prev_outputs = 0;
     for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
         const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
 
@@ -672,7 +679,14 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
             return false;
         }
 
-        memcpy(batch_logits.data() + i*n_vocab, llama_get_logits(ctx), n_tokens*n_vocab*sizeof(float));
+        int n_outputs = 0;
+        for (int i = 0; i < n_tokens; ++i) {
+            n_outputs += batch_view.logits[i] != 0;
+        }
+
+        memcpy(batch_logits.data() + prev_outputs*n_vocab, llama_get_logits(ctx), n_outputs*n_vocab*sizeof(float));
+
+        prev_outputs += n_outputs;
     }
 
     return true;
@@ -779,7 +793,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         size_t ending_logprob_count[4];
         double ending_logprob[4];
 
-        size_t i_batch;         // starting index in the llama_batch
+        size_t i_logits;        // starting index of logits in the llama_batch
         size_t common_prefix;   // max number of initial tokens that are the same in all sentences
         size_t required_tokens; // needed number of tokens to evaluate all 4 endings
         std::vector<llama_token> seq_tokens[4];
@@ -844,9 +858,10 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     const int max_tasks_per_batch = 32;
     const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+    llama_batch batch = llama_batch_init(n_ctx, 0, 4);
 
     std::vector<float> tok_logits(n_vocab);
+    // TODO: this could be made smaller; it's currently the worst-case size
     std::vector<float> batch_logits(n_vocab*n_ctx);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
@@ -857,16 +872,17 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0; // this tells us where in `llama_batch` we are currently
+        size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
         llama_batch_clear(batch);
 
         // batch as much tasks as possible into the available context
-        // each task has 4 unique seuqnce ids - one for each ending
+        // each task has 4 unique sequence ids - one for each ending
         // the common prefix is shared among the 4 sequences to save tokens
         // we extract logits only from the last common token and from all ending tokens of each sequence
         while (n_cur + (int) hs_data[i1].required_tokens <= n_ctx) {
             auto & hs_cur = hs_data[i1];
+            int n_logits = 0;
 
             const int s0 = 4*(i1 - i0);
             if (s0 + 4 > max_seq) {
@@ -874,18 +890,23 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             }
 
             for (size_t i = 0; i < hs_cur.common_prefix; ++i) {
-                llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
+                llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
             }
             batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+            n_logits += 1;
 
             for (int s = 0; s < 4; ++s) {
-                for (size_t i = hs_cur.common_prefix; i < hs_cur.seq_tokens[s].size(); ++i) {
-                    llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, true);
+                const size_t seq_tokens_size = hs_cur.seq_tokens[s].size();
+                // TODO: don't evaluate the last token of each sequence
+                for (size_t i = hs_cur.common_prefix; i < seq_tokens_size; ++i) {
+                    const bool needs_logits = i < seq_tokens_size - 1;
+                    llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    n_logits += needs_logits;
                 }
             }
 
-            hs_cur.i_batch = i_batch;
-            i_batch += hs_cur.required_tokens;
+            hs_cur.i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += hs_data[i1].required_tokens;
             if (++i1 == hs_task_count) {
@@ -911,12 +932,11 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         eval_pairs.clear();
         for (size_t i = i0; i < i1; ++i) {
             auto & hs_cur = hs_data[i];
-            size_t li = hs_cur.common_prefix;
+            size_t li = 1; // skip the last logit of the common prefix (computed separately below)
             for (int s = 0; s < 4; ++s) {
                 for (size_t j = hs_cur.common_prefix; j < hs_cur.seq_tokens[s].size() - 1; j++) {
-                    eval_pairs.emplace_back(hs_cur.i_batch + li++, hs_cur.seq_tokens[s][j + 1]);
+                    eval_pairs.emplace_back(hs_cur.i_logits + li++, hs_cur.seq_tokens[s][j + 1]);
                 }
-                ++li;
             }
         }
         // Then we do the actual calculation
@@ -928,7 +948,8 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         for (size_t i = i0; i < i1; ++i) {
             auto & hs_cur = hs_data[i];
 
-            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(hs_cur.i_batch + hs_cur.common_prefix - 1), n_vocab*sizeof(float));
+            // get the logits of the last token of the common prefix
+            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*hs_cur.i_logits, n_vocab*sizeof(float));
 
             const auto first_probs = softmax(tok_logits);
 
@@ -978,7 +999,7 @@ struct winogrande_entry {
     std::array<std::string, 2> choices;
     int answer;
 
-    size_t i_batch;
+    size_t i_logits;
     size_t common_prefix;
     size_t required_tokens;
     size_t n_base1; // number of tokens for context + choice 1
@@ -1104,6 +1125,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
             task.common_prefix++;
         }
 
+        // TODO: the last token of each of the sequences don't need to be evaluated
         task.required_tokens = task.common_prefix +
             task.seq_tokens[0].size() - task.common_prefix +
             task.seq_tokens[1].size() - task.common_prefix;
@@ -1121,9 +1143,10 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
     const int max_tasks_per_batch = 128;
     const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+    llama_batch batch = llama_batch_init(n_ctx, 0, 2);
 
     std::vector<float> tok_logits(n_vocab);
+    // TODO: this could be made smaller; it's currently the worst-case size
     std::vector<float> batch_logits(n_vocab*n_ctx);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
@@ -1137,29 +1160,33 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0;
+        size_t i_logits = 0;
 
         llama_batch_clear(batch);
 
         while (n_cur + (int) data[i1].required_tokens <= n_ctx) {
+            int n_logits = 0;
             const int s0 = 2*(i1 - i0);
             if (s0 + 2 > max_seq) {
                 break;
             }
 
             for (size_t i = 0; i < data[i1].common_prefix; ++i) {
-                llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1}, false);
+                llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
             }
             batch.logits[batch.n_tokens - 1] = true;
+            n_logits += 1;
 
             for (int s = 0; s < 2; ++s) {
+                // TODO: end before the last token, no need to predict past the end of the sequences
                 for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) {
                     llama_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true);
+                    n_logits += 1;
                 }
             }
 
-            data[i1].i_batch = i_batch;
-            i_batch += data[i1].required_tokens;
+            data[i1].i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += data[i1].required_tokens;
             if (++i1 == data.size()) {
@@ -1190,15 +1217,16 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
 
             const auto& n_base1 = skip_choice ? task.n_base1 : task.common_prefix;
             const int last_1st = task.seq_tokens[0].size() - n_base1 > 1 ? 1 : 0;
-            size_t li = n_base1 - 1;
+            size_t li = n_base1 - task.common_prefix;
             for (size_t j = n_base1-1; j < task.seq_tokens[0].size()-1-last_1st; ++j) {
-                eval_pairs.emplace_back(task.i_batch + li++, task.seq_tokens[0][j+1]);
+                eval_pairs.emplace_back(task.i_logits + li++, task.seq_tokens[0][j+1]);
             }
             const auto& n_base2 = skip_choice ? task.n_base2 : task.common_prefix;
             const int last_2nd = task.seq_tokens[1].size() - n_base2 > 1 ? 1 : 0;
-            li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - 1;
+            // FIXME: this uses the wrong first logits when not skipping the choice word
+            li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - task.common_prefix;
             for (size_t j = n_base2-1; j < task.seq_tokens[1].size()-1-last_2nd; ++j) {
-                eval_pairs.emplace_back(task.i_batch + li++, task.seq_tokens[1][j+1]);
+                eval_pairs.emplace_back(task.i_logits + li++, task.seq_tokens[1][j+1]);
             }
         }
         compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results);
@@ -1287,7 +1315,7 @@ struct multiple_choice_task {
     }
 
     // For evaluation
-    size_t i_batch;         // starting index in the llama_batch
+    size_t i_logits;        // starting index of logits in the llama_batch
     size_t common_prefix;   // max number of initial tokens that are the same in all sentences
     size_t required_tokens; // needed number of tokens to evaluate all answers
     std::vector<std::vector<llama_token>> seq_tokens;
@@ -1366,7 +1394,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
     std::vector<uint32_t> task_pos(n_task);
     strstream.read((char *)task_pos.data(), task_pos.size()*sizeof(uint32_t));
     if (strstream.fail()) {
-        printf("%s: failed to raad task positions from prompt\n", __func__);
+        printf("%s: failed to read task positions from prompt\n", __func__);
         return;
     }
 
@@ -1447,7 +1475,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
             return;
         }
     } else {
-        int n_dot = n_task/100;
+        int n_dot = std::max((int) n_task/100, 1);
         int i_task = 0;
         for (auto& task : tasks) {
             ++i_task;
@@ -1491,17 +1519,18 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0; // this tells us where in `llama_batch` we are currently
+        size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
         llama_batch_clear(batch);
 
         // batch as much tasks as possible into the available context
-        // each task has 4 unique seuqnce ids - one for each ending
+        // each task has 4 unique sequence ids - one for each ending
         // the common prefix is shared among the 4 sequences to save tokens
         // we extract logits only from the last common token and from all ending tokens of each sequence
         int s0 = 0;
         while (n_cur + (int) tasks[i1].required_tokens <= n_ctx) {
             auto& cur_task = tasks[i1];
+            int n_logits = 0;
 
             int num_answers = cur_task.seq_tokens.size();
             if (s0 + num_answers > max_seq) {
@@ -1518,17 +1547,22 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
                 llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
             }
             batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+            n_logits += 1;
 
             for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
-                for (size_t i = cur_task.common_prefix; i < cur_task.seq_tokens[s].size(); ++i) {
-                    llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, true);
+                const size_t seq_tokens_size = cur_task.seq_tokens[s].size();
+                // TODO: don't evaluate the last token of each sequence
+                for (size_t i = cur_task.common_prefix; i < seq_tokens_size; ++i) {
+                    const bool needs_logits = i < seq_tokens_size - 1;
+                    llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    n_logits += needs_logits;
                 }
             }
 
             s0 += num_answers;
 
-            cur_task.i_batch = i_batch;
-            i_batch += cur_task.required_tokens;
+            cur_task.i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += cur_task.required_tokens;
             if (++i1 == tasks.size()) {
@@ -1554,12 +1588,11 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         eval_pairs.clear();
         for (size_t i = i0; i < i1; ++i) {
             auto& cur_task = tasks[i];
-            size_t li = cur_task.common_prefix;
+            size_t li = 1; // skip the last logit of the common prefix (computed separately below)
             for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
                 for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) {
-                    eval_pairs.emplace_back(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1]);
+                    eval_pairs.emplace_back(cur_task.i_logits + li++, cur_task.seq_tokens[s][j + 1]);
                 }
-                ++li;
             }
         }
         // Then we do the actual calculation
@@ -1578,7 +1611,8 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
             //}
             //printf("\n    common_prefix: %zu\n", cur_task.common_prefix);
 
-            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(cur_task.i_batch + cur_task.common_prefix - 1), n_vocab*sizeof(float));
+            // get the logits of the last token of the common prefix
+            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*cur_task.i_logits, n_vocab*sizeof(float));
 
             const auto first_probs = softmax(tok_logits);
 
@@ -1730,6 +1764,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
+            // TODO: use llama_batch.logits instead of relying on logits_all == true
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return;

examples/quantize/quantize.cpp

@@ -26,6 +26,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "IQ2_S",  LLAMA_FTYPE_MOSTLY_IQ2_S,  " 2.5  bpw quantization",            },
     { "IQ2_M",  LLAMA_FTYPE_MOSTLY_IQ2_M,  " 2.7  bpw quantization",            },
     { "IQ1_S",  LLAMA_FTYPE_MOSTLY_IQ1_S,  " 1.56 bpw quantization",            },
+    { "IQ1_M",  LLAMA_FTYPE_MOSTLY_IQ1_M,  " 1.75 bpw quantization",            },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization",            },
@@ -87,13 +88,17 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
     printf("  --imatrix file_name: use data in file_name as importance matrix for quant optimizations\n");
     printf("  --include-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
+    printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
+    printf("  --override-kv KEY=TYPE:VALUE\n");
+    printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
     printf("Note: --include-weights and --exclude-weights cannot be used together\n");
     printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
@@ -107,56 +112,60 @@ static void usage(const char * executable) {
     exit(1);
 }
 
-static void load_imatrix(const std::string& imatrix_file, std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+static void load_imatrix(const std::string & imatrix_file, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
     std::ifstream in(imatrix_file.c_str(), std::ios::binary);
     if (!in) {
-        printf("%s: failed to open %s\n",__func__,imatrix_file.c_str());
-        return;
+        printf("%s: failed to open %s\n",__func__, imatrix_file.c_str());
+        exit(1);
     }
     int n_entries;
-    in.read((char*)&n_entries, sizeof(n_entries));
+    in.read((char *)&n_entries, sizeof(n_entries));
     if (in.fail() || n_entries < 1) {
         printf("%s: no data in file %s\n", __func__, imatrix_file.c_str());
-        return;
+        exit(1);
     }
     for (int i = 0; i < n_entries; ++i) {
         int len; in.read((char *)&len, sizeof(len));
         std::vector<char> name_as_vec(len+1);
         in.read((char *)name_as_vec.data(), len);
         if (in.fail()) {
-            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1,imatrix_file.c_str());
-            return;
+            printf("%s: failed reading name for entry %d from %s\n", __func__, i+1, imatrix_file.c_str());
+            exit(1);
         }
         name_as_vec[len] = 0;
         std::string name{name_as_vec.data()};
-        auto& e = imatrix_data[std::move(name)];
+        auto & e = imatrix_data[name];
         int ncall;
-        in.read((char*)&ncall, sizeof(ncall));
+        in.read((char *)&ncall, sizeof(ncall));
         int nval;
         in.read((char *)&nval, sizeof(nval));
         if (in.fail() || nval < 1) {
-            printf("%s: failed reading number of values for entry %d\n",__func__,i);
+            printf("%s: failed reading number of values for entry %d\n", __func__, i);
             imatrix_data = {};
-            return;
+            exit(1);
         }
         e.resize(nval);
-        in.read((char*)e.data(), nval*sizeof(float));
+        in.read((char *)e.data(), nval*sizeof(float));
         if (in.fail()) {
-            printf("%s: failed reading data for entry %d\n",__func__,i);
+            printf("%s: failed reading data for entry %d\n", __func__, i);
             imatrix_data = {};
-            return;
+            exit(1);
         }
         if (ncall > 0) {
             for (auto& v : e) v /= ncall;
         }
+
+        if (getenv("LLAMA_TRACE")) {
+            printf("%s: loaded data (size = %6d, ncall = %6d) for '%s'\n", __func__, int(e.size()), ncall, name.c_str());
+        }
     }
-    printf("%s: loaded %d importance matrix entries from %s\n",__func__,int(imatrix_data.size()),imatrix_file.c_str());
+    printf("%s: loaded %d importance matrix entries from %s\n", __func__, int(imatrix_data.size()), imatrix_file.c_str());
 }
 
-static void prepare_imatrix(const std::string& imatrix_file,
-        const std::vector<std::string>& included_weights,
-        const std::vector<std::string>& excluded_weights,
-        std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+static void prepare_imatrix(const std::string & imatrix_file,
+        const std::vector<std::string> & included_weights,
+        const std::vector<std::string> & excluded_weights,
+        std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
     if (!imatrix_file.empty()) {
         load_imatrix(imatrix_file, imatrix_data);
     }
@@ -201,6 +210,43 @@ static ggml_type parse_ggml_type(const char * arg) {
     return result;
 }
 
+static bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides) {
+    const char* sep = strchr(data, '=');
+    if (sep == nullptr || sep - data >= 128) {
+        fprintf(stderr, "%s: malformed KV override '%s'\n", __func__, data);
+        return false;
+    }
+    llama_model_kv_override kvo;
+    std::strncpy(kvo.key, data, sep - data);
+    kvo.key[sep - data] = 0;
+    sep++;
+    if (strncmp(sep, "int:", 4) == 0) {
+        sep += 4;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+        kvo.int_value = std::atol(sep);
+    } else if (strncmp(sep, "float:", 6) == 0) {
+        sep += 6;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
+        kvo.float_value = std::atof(sep);
+    } else if (strncmp(sep, "bool:", 5) == 0) {
+        sep += 5;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
+        if (std::strcmp(sep, "true") == 0) {
+            kvo.bool_value = true;
+        } else if (std::strcmp(sep, "false") == 0) {
+            kvo.bool_value = false;
+        } else {
+            fprintf(stderr, "%s: invalid boolean value for KV override '%s'\n", __func__, data);
+            return false;
+        }
+    } else {
+        fprintf(stderr, "%s: invalid type for KV override '%s'\n", __func__, data);
+        return false;
+    }
+    overrides.emplace_back(std::move(kvo));
+    return true;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -211,6 +257,7 @@ int main(int argc, char ** argv) {
     int arg_idx = 1;
     std::string imatrix_file;
     std::vector<std::string> included_weights, excluded_weights;
+    std::vector<llama_model_kv_override> kv_overrides;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -227,6 +274,10 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
+            if (arg_idx == argc-1 || !parse_kv_override(argv[++arg_idx], kv_overrides)) {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {
@@ -267,6 +318,11 @@ int main(int argc, char ** argv) {
     if (!imatrix_data.empty()) {
         params.imatrix = &imatrix_data;
     }
+    if (!kv_overrides.empty()) {
+        kv_overrides.emplace_back();
+        kv_overrides.back().key[0] = 0;
+        params.kv_overrides = &kv_overrides;
+    }
 
     llama_backend_init();
 
@@ -288,8 +344,7 @@ int main(int argc, char ** argv) {
         if (ftype_str == "COPY") {
             params.only_copy = true;
         }
-    }
-    else {
+    } else {
         fname_out = argv[arg_idx];
         arg_idx++;
 
@@ -320,10 +375,12 @@ int main(int argc, char ** argv) {
 
     if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
          params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
-        fprintf(stderr, "\n===============================================================================================\n");
-        fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
-        fprintf(stderr, "===============================================================================================\n\n\n");
+         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S  ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) && imatrix_data.empty()) {
+        fprintf(stderr, "\n==========================================================================================================\n");
+        fprintf(stderr, "Please do not use IQ1_S, IQ1_M, IQ2_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "==========================================================================================================\n\n\n");
         return 1;
     }
 

examples/server/README.md

@@ -16,52 +16,50 @@ The project is under active development, and we are [looking for feedback and co
 
 **Command line options:**
 
-- `--threads N`, `-t N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching, this parameter is used only if one token is to be processed on CPU backend.
+- `--threads N`, `-t N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching. This parameter is used only if one token is to be processed on CPU backend.
 - `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. Not used if model layers are offloaded to GPU.
-- `--threads-http N`: number of threads in the http server pool to process requests (default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`)
+- `--threads-http N`: Number of threads in the http server pool to process requests. Default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`
 - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
-- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (default: unused).
-- `-hfr REPO, --hf-repo REPO`: Hugging Face model repository (default: unused).
-- `-hff FILE, --hf-file FILE`: Hugging Face model file (default: unused).
+- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file. Default: unused
+- `-hfr REPO, --hf-repo REPO`: Hugging Face model repository. Default: unused
+- `-hff FILE, --hf-file FILE`: Hugging Face model file. Default: unused
 - `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
-- `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
+- `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is `512`, but LLaMA models were built with a context of `2048`, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of `4096`.
 - `-ngl N`, `--n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
-- `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
-- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
-- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `2048`.
-- `-ub N`, `--ubatch-size N`: physical maximum batch size. Default: `512`.
-- `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
+- `-mg i, --main-gpu i`: When using multiple GPUs, this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default, GPU `0` is used.
+- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs, this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default, the data is split in proportion to VRAM, but this may not be optimal for performance.
+- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `2048`
+- `-ub N`, `--ubatch-size N`: Physical maximum batch size. Default: `512`
 - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
-- `--numa STRATEGY`: Attempt one of the below optimization strategies  that help on some NUMA systems
+- `--numa STRATEGY`: Attempt one of the below optimization strategies that may help on some NUMA systems
 - `--numa distribute`: Spread execution evenly over all nodes
 - `--numa isolate`: Only spawn threads on CPUs on the node that execution started on
-- `--numa numactl`: Use the CPU map provided by numactl
-if run without this previously, it is recommended to drop the system page cache before using this
-see https://github.com/ggerganov/llama.cpp/issues/1437
+- `--numa numactl`: Use the CPU map provided by numactl. If run without this previously, it is recommended to drop the system
+page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/1437
 
-- `--numa`: Attempt optimizations that help on some NUMA systems.
+- `--numa`: Attempt optimizations that may help on some NUMA systems.
 - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
-- `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
-- `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
-- `--port`: Set the port to listen. Default: `8080`.
-- `--path`: path from which to serve static files (default: disabled)
-- `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
-- `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s.
-- `--embedding`: Enable embedding extraction, Default: disabled.
-- `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
-- `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
-- `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
+- `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`
+- `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`
+- `--port`: Set the port to listen. Default: `8080`
+- `--path`: Path from which to serve static files. Default: disabled
+- `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
+- `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s.
+- `--embedding`: Enable embedding extraction. Default: disabled
+- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`
+- `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching).  Default: disabled
+- `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
-- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
-- `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
-- `-n N, --n-predict N`: Set the maximum tokens to predict (default: -1)
+- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend. Used together with group attention width `--grp-attn-w`. Default: `1`, which is disabled.
+- `--grp-attn-w`: Set the group attention width to extend context size through self-extend.  Used together with group attention factor `--grp-attn-n`. Default: `512`
+- `-n N, --n-predict N`: Set the maximum tokens to predict. Default: `-1`
 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
-- `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled)
-- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-- `--log-disable`: Output logs to stdout only, not to `llama.log`. default: enabled.
-- `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json)
+- `--metrics`: enable prometheus `/metrics` compatible endpoint. Default: disabled
+- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name.  Default: template taken from model's metadata. We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
+- `--log-disable`: Output logs to stdout only, not to `llama.log`. Default: enabled
+- `--log-format FORMAT`: Define the log output to FORMAT: json or text Default: `json`
 
 **If compiled with `LLAMA_SERVER_SSL=ON`**
 - `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key
@@ -69,7 +67,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 
 ## Build
 
-server is build alongside everything else from the root of the project
+`server` is built alongside everything else from the root of the project
 
 - Using `make`:
 
@@ -85,7 +83,7 @@ server is build alongside everything else from the root of the project
 
 ## Build with SSL
 
-server can also be built with SSL support using OpenSSL 3
+`server` can also be built with SSL support using OpenSSL 3
 
 - Using `make`:
 
@@ -135,7 +133,7 @@ docker run -p 8080:8080 -v /path/to/models:/models --gpus all ghcr.io/ggerganov/
 
 ## Testing with CURL
 
-Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS.
+Using [curl](https://curl.se/). On Windows, `curl.exe` should be available in the base OS.
 
 ```sh
 curl --request POST \
@@ -159,7 +157,7 @@ mkdir llama-client
 cd llama-client
 ```
 
-Create a index.js file and put inside this:
+Create a index.js file and put this inside:
 
 ```javascript
 const prompt = `Building a website can be done in 10 simple steps:`;
@@ -190,8 +188,8 @@ node index.js
   - 503 -> `{"status": "loading model"}` if the model is still being loaded.
   - 500 -> `{"status": "error"}` if the model failed to load.
   - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
-  - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available.
-  - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slot are currently available.
+  - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slots are currently available.
+  - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slots are currently available.
 
   If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
 
@@ -205,75 +203,75 @@ node index.js
       - The model's `tokenizer.ggml.add_bos_token` metadata is `true`
       - The system prompt is empty
 
-    `temperature`: Adjust the randomness of the generated text (default: 0.8).
+    `temperature`: Adjust the randomness of the generated text. Default: `0.8`
 
-    `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` (default: 0.0, 0.0 = disabled).
+    `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` Default: `0.0`, which is disabled.
 
-    `dynatemp_exponent`: Dynamic temperature exponent (default: 1.0).
+    `dynatemp_exponent`: Dynamic temperature exponent. Default: `1.0`
 
-    `top_k`: Limit the next token selection to the K most probable tokens (default: 40).
+    `top_k`: Limit the next token selection to the K most probable tokens.  Default: `40`
 
-    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95).
+    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P. Default: `0.95`
 
-    `min_p`: The minimum probability for a token to be considered, relative to the probability of the most likely token (default: 0.05).
+    `min_p`: The minimum probability for a token to be considered, relative to the probability of the most likely token. Default: `0.05`
 
-    `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity).
+    `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity.
 
     `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded.
-    By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the prompt.
+    By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt.
 
     `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
 
     `stop`: Specify a JSON array of stopping strings.
-    These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).
+    These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`
 
-    `tfs_z`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
+    `tfs_z`: Enable tail free sampling with parameter z. Default: `1.0`, which is disabled.
 
-    `typical_p`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
+    `typical_p`: Enable locally typical sampling with parameter p. Default: `1.0`, which is disabled.
 
-    `repeat_penalty`: Control the repetition of token sequences in the generated text (default: 1.1).
+    `repeat_penalty`: Control the repetition of token sequences in the generated text. Default: `1.1`
 
-    `repeat_last_n`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
+    `repeat_last_n`: Last n tokens to consider for penalizing repetition. Default: `64`, where `0` is disabled and `-1` is ctx-size.
 
-    `penalize_nl`: Penalize newline tokens when applying the repeat penalty (default: true).
+    `penalize_nl`: Penalize newline tokens when applying the repeat penalty. Default: `true`
 
-    `presence_penalty`: Repeat alpha presence penalty (default: 0.0, 0.0 = disabled).
+    `presence_penalty`: Repeat alpha presence penalty. Default: `0.0`, which is disabled.
 
-    `frequency_penalty`: Repeat alpha frequency penalty (default: 0.0, 0.0 = disabled);
+    `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
 
-    `penalty_prompt`: This will replace the `prompt` for the purpose of the penalty evaluation. Can be either `null`, a string or an array of numbers representing tokens (default: `null` = use the original `prompt`).
+    `penalty_prompt`: This will replace the `prompt` for the purpose of the penalty evaluation. Can be either `null`, a string or an array of numbers representing tokens. Default: `null`, which is to use the original `prompt`.
 
-    `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
+    `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
 
-    `mirostat_tau`: Set the Mirostat target entropy, parameter tau (default: 5.0).
+    `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
 
-    `mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1).
+    `mirostat_eta`: Set the Mirostat learning rate, parameter eta.  Default: `0.1`
 
-    `grammar`: Set grammar for grammar-based sampling (default: no grammar)
+    `grammar`: Set grammar for grammar-based sampling.  Default: no grammar
 
-    `seed`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
+    `seed`: Set the random number generator (RNG) seed.  Default: `-1`, which is a random seed.
 
-    `ignore_eos`: Ignore end of stream token and continue generating (default: false).
+    `ignore_eos`: Ignore end of stream token and continue generating.  Default: `false`
 
-    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. (default: []).
+    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. Default: `[]`
 
-    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
+    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token. Default: `0`
 
-    `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum (default: 0)
+    `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum. Default: `0`
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
-    `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
+    `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot.  Default: `-1`
 
-    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false)
+    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch.  Default: `false`
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
-    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. (default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values)
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
 ### Result JSON
 
-- Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
+- Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
 
 - `completion_probabilities`: An array of token probabilities for each completion. The array's length is `n_predict`. Each item in the array has the following structure:
 
@@ -287,7 +285,7 @@ node index.js
     },
     {
       "prob": float,
-      "tok_str": "<second most likely tonen>"
+      "tok_str": "<second most likely token>"
     },
     ...
   ]
@@ -357,14 +355,14 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - `assistant_name` - the required assistant name to generate the prompt in case you have specified a system prompt for all slots.
 - `user_name` - the required anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
-- `default_generation_settings` - the default generation settings for the `/completion` endpoint, has the same fields as the `generation_settings` response object from the `/completion` endpoint.
+- `default_generation_settings` - the default generation settings for the `/completion` endpoint, which has the same fields as the `generation_settings` response object from the `/completion` endpoint.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 
 - **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only model with [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, ChatML template will be used.
 
     *Options:*
 
-    See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such are `mirostat` are supported.
+    See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
 
     *Examples:*
 
@@ -514,16 +512,16 @@ Available metrics:
 - `llamacpp:tokens_predicted_total`: Number of generation tokens processed.
 - `llamacpp:prompt_tokens_seconds`: Average prompt throughput in tokens/s.
 - `llamacpp:predicted_tokens_seconds`: Average generation throughput in tokens/s.
-- `llamacpp:kv_cache_usage_ratio`: KV-cache usage. 1 means 100 percent usage.
+- `llamacpp:kv_cache_usage_ratio`: KV-cache usage. `1` means 100 percent usage.
 - `llamacpp:kv_cache_tokens`: KV-cache tokens.
-- `llamacpp:requests_processing`: Number of request processing.
-- `llamacpp:requests_deferred`: Number of request deferred.
+- `llamacpp:requests_processing`: Number of requests processing.
+- `llamacpp:requests_deferred`: Number of requests deferred.
 
 ## More examples
 
 ### Change system prompt on runtime
 
-To use the server example to serve multiple chat-type clients while keeping the same system prompt, you can utilize the option `system_prompt` to achieve that. This only needs to be done once to establish it.
+To use the server example to serve multiple chat-type clients while keeping the same system prompt, you can utilize the option `system_prompt`. This only needs to be used once.
 
 `prompt`: Specify a context that you want all connecting clients to respect.
 
@@ -562,11 +560,11 @@ bash chat.sh
 
 ### OAI-like API
 
-The HTTP server supports OAI-like API: https://github.com/openai/openai-openapi
+The HTTP `server` supports an OAI-like API: https://github.com/openai/openai-openapi
 
 ### API errors
 
-Server returns error in the same format as OAI: https://github.com/openai/openai-openapi
+`server` returns errors in the same format as OAI: https://github.com/openai/openai-openapi
 
 Example of an error:
 

examples/server/bench/bench.py

@@ -0,0 +1,309 @@
+import argparse
+import json
+import os
+import re
+import signal
+import socket
+import subprocess
+import sys
+import threading
+import time
+import traceback
+from contextlib import closing
+from datetime import datetime
+
+import matplotlib
+import matplotlib.dates
+import matplotlib.pyplot as plt
+import requests
+from statistics import mean
+
+
+def main(args_in: list[str] | None = None) -> None:
+    parser = argparse.ArgumentParser(description="Start server benchmark scenario")
+    parser.add_argument("--name", type=str, help="Bench name", required=True)
+    parser.add_argument("--runner-label", type=str, help="Runner label", required=True)
+    parser.add_argument("--branch", type=str, help="Branch name", default="detached")
+    parser.add_argument("--commit", type=str, help="Commit name", default="dirty")
+    parser.add_argument("--host", type=str, help="Server listen host", default="0.0.0.0")
+    parser.add_argument("--port", type=int, help="Server listen host", default="8080")
+    parser.add_argument("--model-path-prefix", type=str, help="Prefix where to store the model files", default="models")
+    parser.add_argument("--n-prompts", type=int,
+                        help="SERVER_BENCH_N_PROMPTS: total prompts to randomly select in the benchmark", required=True)
+    parser.add_argument("--max-prompt-tokens", type=int,
+                        help="SERVER_BENCH_MAX_PROMPT_TOKENS: maximum prompt tokens to filter out in the dataset",
+                        required=True)
+    parser.add_argument("--max-tokens", type=int,
+                        help="SERVER_BENCH_MAX_CONTEXT: maximum context size of the completions request to filter out in the dataset: prompt + predicted tokens",
+                        required=True)
+    parser.add_argument("--hf-repo", type=str, help="Hugging Face model repository", required=True)
+    parser.add_argument("--hf-file", type=str, help="Hugging Face model file", required=True)
+    parser.add_argument("-ngl", "--n-gpu-layers", type=int, help="layers to the GPU for computation", required=True)
+    parser.add_argument("--ctx-size", type=int, help="Set the size of the prompt context", required=True)
+    parser.add_argument("--parallel", type=int, help="Set the number of slots for process requests", required=True)
+    parser.add_argument("--batch-size", type=int, help="Set the batch size for prompt processing", required=True)
+    parser.add_argument("--ubatch-size", type=int, help="physical maximum batch size", required=True)
+    parser.add_argument("--scenario", type=str, help="Scenario to run", required=True)
+    parser.add_argument("--duration", type=str, help="Bench scenario", required=True)
+
+    args = parser.parse_args(args_in)
+
+    start_time = time.time()
+
+    # Start the server and performance scenario
+    try:
+        server_process = start_server(args)
+    except Exception:
+        print("bench: server start error :")
+        traceback.print_exc(file=sys.stdout)
+        sys.exit(1)
+
+    # start the benchmark
+    try:
+        start_benchmark(args)
+
+        iterations = 0
+        with open("results.github.env", 'w') as github_env:
+            # parse output
+            with open('k6-results.json', 'r') as bench_results:
+                # Load JSON data from file
+                data = json.load(bench_results)
+                for metric_name in data['metrics']:
+                    for metric_metric in data['metrics'][metric_name]:
+                        value = data['metrics'][metric_name][metric_metric]
+                        if isinstance(value, float) or isinstance(value, int):
+                            value = round(value, 2)
+                            data['metrics'][metric_name][metric_metric]=value
+                            github_env.write(
+                                f"{escape_metric_name(metric_name)}_{escape_metric_name(metric_metric)}={value}\n")
+                token_seconds = data['metrics']['llamacpp_tokens_second']['avg']
+                iterations = data['root_group']['checks']['success completion']['passes']
+
+    except Exception:
+        print("bench: error :")
+        traceback.print_exc(file=sys.stdout)
+
+    # Stop the server
+    if server_process:
+        try:
+            print(f"bench: shutting down server pid={server_process.pid} ...")
+            if os.name == 'nt':
+                interrupt = signal.CTRL_C_EVENT
+            else:
+                interrupt = signal.SIGINT
+            server_process.send_signal(interrupt)
+            server_process.wait(0.5)
+
+        except subprocess.TimeoutExpired:
+            print(f"server still alive after 500ms, force-killing pid={server_process.pid} ...")
+            server_process.kill()  # SIGKILL
+            server_process.wait()
+
+        while is_server_listening(args.host, args.port):
+            time.sleep(0.1)
+
+    title = (f"llama.cpp {args.name} on {args.runner_label}\n "
+             f"duration={args.duration} {iterations} iterations")
+    xlabel = (f"{args.hf_repo}/{args.hf_file}\n"
+              f"parallel={args.parallel} ctx-size={args.ctx_size} ngl={args.n_gpu_layers} batch-size={args.batch_size} ubatch-size={args.ubatch_size} pp={args.max_prompt_tokens} pp+tg={args.max_tokens}\n"
+              f"branch={args.branch} commit={args.commit}")
+
+    # Prometheus
+    end_time = time.time()
+    prometheus_metrics = {}
+    if is_server_listening("0.0.0.0", 9090):
+        metrics = ['prompt_tokens_seconds', 'predicted_tokens_seconds',
+                   'kv_cache_usage_ratio', 'requests_processing', 'requests_deferred']
+
+        for metric in metrics:
+            resp = requests.get(f"http://localhost:9090/api/v1/query_range",
+                                params={'query': 'llamacpp:' + metric, 'start': start_time, 'end': end_time, 'step': 2})
+
+            with open(f"{metric}.json", 'w') as metric_json:
+                metric_json.write(resp.text)
+
+            if resp.status_code != 200:
+                print(f"bench: unable to extract prometheus metric {metric}: {resp.text}")
+            else:
+                metric_data = resp.json()
+                values = metric_data['data']['result'][0]['values']
+                timestamps, metric_values = zip(*values)
+                metric_values = [float(value) for value in metric_values]
+                prometheus_metrics[metric] = metric_values
+                timestamps_dt = [datetime.fromtimestamp(int(ts)) for ts in timestamps]
+                plt.figure(figsize=(16, 10), dpi=80)
+                plt.plot(timestamps_dt, metric_values, label=metric)
+                plt.xticks(rotation=0, fontsize=14, horizontalalignment='center', alpha=.7)
+                plt.yticks(fontsize=12, alpha=.7)
+
+                ylabel = f"llamacpp:{metric}"
+                plt.title(title,
+                          fontsize=14, wrap=True)
+                plt.grid(axis='both', alpha=.3)
+                plt.ylabel(ylabel, fontsize=22)
+                plt.xlabel(xlabel, fontsize=14, wrap=True)
+                plt.gca().xaxis.set_major_locator(matplotlib.dates.MinuteLocator())
+                plt.gca().xaxis.set_major_formatter(matplotlib.dates.DateFormatter("%Y-%m-%d %H:%M:%S"))
+                plt.gcf().autofmt_xdate()
+
+                # Remove borders
+                plt.gca().spines["top"].set_alpha(0.0)
+                plt.gca().spines["bottom"].set_alpha(0.3)
+                plt.gca().spines["right"].set_alpha(0.0)
+                plt.gca().spines["left"].set_alpha(0.3)
+
+                # Save the plot as a jpg image
+                plt.savefig(f'{metric}.jpg', dpi=60)
+                plt.close()
+
+                # Mermaid format in case images upload failed
+                with (open(f"{metric}.mermaid", 'w') as mermaid_f):
+                    mermaid = (
+                    f"""---
+config:
+    xyChart:
+        titleFontSize: 12
+        width: 900
+        height: 600
+    themeVariables:
+        xyChart:
+            titleColor: "#000000"
+---
+xychart-beta
+    title "{title}"
+    y-axis "llamacpp:{metric}"
+    x-axis "llamacpp:{metric}" {int(min(timestamps))} --> {int(max(timestamps))}
+    line [{', '.join([str(round(float(value), 2)) for value in metric_values])}]
+                    """)
+                    mermaid_f.write(mermaid)
+
+    # 140 chars max for commit status description
+    bench_results = {
+        "i": iterations,
+        "req": {
+            "p90": round(data['metrics']["http_req_duration"]["p(90)"], 2),
+            "avg": round(data['metrics']["http_req_duration"]["avg"], 2),
+        },
+        "pp": {
+            "p90": round(data['metrics']["llamacpp_prompt_tokens"]["p(90)"], 2),
+            "avg": round(data['metrics']["llamacpp_prompt_tokens"]["avg"], 2),
+            "0": round(mean(prometheus_metrics['prompt_tokens_seconds']), 2),
+        },
+        "tg": {
+            "p90": round(data['metrics']["llamacpp_tokens_second"]["p(90)"], 2),
+            "avg": round(data['metrics']["llamacpp_tokens_second"]["avg"], 2),
+            "0": round(mean(prometheus_metrics['predicted_tokens_seconds']), 2),
+        },
+    }
+    with open("results.github.env", 'a') as github_env:
+        github_env.write(f"BENCH_RESULTS={json.dumps(bench_results, indent=None, separators=(',', ':') )}\n")
+        github_env.write(f"BENCH_ITERATIONS={iterations}\n")
+
+        title = title.replace('\n', ' ')
+        xlabel = xlabel.replace('\n', ' ')
+        github_env.write(f"BENCH_GRAPH_TITLE={title}\n")
+        github_env.write(f"BENCH_GRAPH_XLABEL={xlabel}\n")
+
+
+def start_benchmark(args):
+    k6_path = 'k6'
+    if 'BENCH_K6_BIN_PATH' in os.environ:
+        k6_path = os.environ['BENCH_K6_BIN_PATH']
+    k6_args = [
+        'run', args.scenario,
+        '--no-color',
+    ]
+    k6_args.extend(['--duration', args.duration])
+    k6_args.extend(['--iterations', args.n_prompts])
+    k6_args.extend(['--vus', args.parallel])
+    k6_args.extend(['--summary-export', 'k6-results.json'])
+    args = f"SERVER_BENCH_N_PROMPTS={args.n_prompts} SERVER_BENCH_MAX_PROMPT_TOKENS={args.max_prompt_tokens} SERVER_BENCH_MAX_CONTEXT={args.max_tokens} "
+    args = args + ' '.join([str(arg) for arg in [k6_path, *k6_args]])
+    print(f"bench: starting k6 with: {args}")
+    k6_completed = subprocess.run(args, shell=True, stdout=sys.stdout, stderr=sys.stderr)
+    if k6_completed.returncode != 0:
+        raise Exception("bench: unable to run k6")
+
+
+def start_server(args):
+    server_process = start_server_background(args)
+
+    attempts = 0
+    max_attempts = 20
+    if 'GITHUB_ACTIONS' in os.environ:
+        max_attempts *= 2
+
+    while not is_server_listening(args.host, args.port):
+        attempts += 1
+        if attempts > max_attempts:
+            assert False, "server not started"
+        print(f"bench:     waiting for server to start ...")
+        time.sleep(0.5)
+
+    print("bench: server started.")
+    return server_process
+
+
+def start_server_background(args):
+    # Start the server
+    server_path = '../../../build/bin/server'
+    if 'LLAMA_SERVER_BIN_PATH' in os.environ:
+        server_path = os.environ['LLAMA_SERVER_BIN_PATH']
+    server_args = [
+        '--host', args.host,
+        '--port', args.port,
+    ]
+    model_file = args.model_path_prefix + os.path.sep + args.hf_file
+    model_dir  = os.path.dirname(model_file)
+    if not os.path.exists(model_dir):
+        os.makedirs(model_dir)
+    server_args.extend(['--model', model_file])
+    server_args.extend(['--hf-repo', args.hf_repo])
+    server_args.extend(['--hf-file', args.hf_file])
+    server_args.extend(['--n-gpu-layers', args.n_gpu_layers])
+    server_args.extend(['--ctx-size', args.ctx_size])
+    server_args.extend(['--parallel', args.parallel])
+    server_args.extend(['--batch-size', args.batch_size])
+    server_args.extend(['--ubatch-size', args.ubatch_size])
+    server_args.extend(['--n-predict', args.max_tokens * 2])
+    server_args.extend(['--defrag-thold', "0.1"])
+    server_args.append('--cont-batching')
+    server_args.append('--metrics')
+    server_args.extend(['--log-format', "text"])
+    args = [str(arg) for arg in [server_path, *server_args]]
+    print(f"bench: starting server with: {' '.join(args)}")
+    pkwargs = {
+        'stdout': subprocess.PIPE,
+        'stderr': subprocess.PIPE
+    }
+    server_process = subprocess.Popen(
+        args,
+        **pkwargs)
+
+    def server_log(in_stream, out_stream):
+        for line in iter(in_stream.readline, b''):
+            print(line.decode('utf-8'), end='', file=out_stream)
+
+    thread_stdout = threading.Thread(target=server_log, args=(server_process.stdout, sys.stdout))
+    thread_stdout.start()
+    thread_stderr = threading.Thread(target=server_log, args=(server_process.stderr, sys.stderr))
+    thread_stderr.start()
+
+    return server_process
+
+
+def is_server_listening(server_fqdn, server_port):
+    with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
+        result = sock.connect_ex((server_fqdn, server_port))
+        _is_server_listening = result == 0
+        if _is_server_listening:
+            print(f"server is listening on {server_fqdn}:{server_port}...")
+        return _is_server_listening
+
+
+def escape_metric_name(metric_name):
+    return re.sub('[^A-Z0-9]', '_', metric_name.upper())
+
+
+if __name__ == '__main__':
+    main()

examples/server/bench/prometheus.yml

@@ -0,0 +1,9 @@
+global:
+  scrape_interval:     10s
+  external_labels:
+    llamacpp: 'server'
+
+scrape_configs:
+  - job_name: 'llama.cpp server'
+    static_configs:
+      - targets: ['localhost:8080']

examples/server/bench/requirements.txt

@@ -0,0 +1,2 @@
+matplotlib
+requests

examples/server/bench/script.js

@@ -87,6 +87,7 @@ export default function () {
         ],
         "model": model,
         "stream": false,
+        "seed": 42,
         "max_tokens": max_tokens
     }
 

examples/server/completion.js.hpp

@@ -43,444 +43,454 @@ unsigned char completion_js[] = {
   0x7d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63,
   0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20,
   0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x72,
-  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x69, 0x66,
-  0x20, 0x28, 0x21, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65,
-  0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65,
-  0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72,
-  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d,
-  0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f,
-  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x50, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61,
-  0x72, 0x61, 0x6d, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x73, 0x2c,
-  0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20,
-  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x70, 0x6f,
-  0x6e, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20,
-  0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x22, 0x2f, 0x63, 0x6f, 0x6d, 0x70,
-  0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x2c, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x6d, 0x65, 0x74, 0x68, 0x6f, 0x64, 0x3a, 0x20, 0x27,
-  0x50, 0x4f, 0x53, 0x54, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x62,
-  0x6f, 0x64, 0x79, 0x3a, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74,
-  0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x63, 0x6f, 0x6d, 0x70,
-  0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73,
-  0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, 0x61, 0x64, 0x65,
-  0x72, 0x73, 0x3a, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x27, 0x43, 0x6f, 0x6e, 0x6e, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x27,
-  0x3a, 0x20, 0x27, 0x6b, 0x65, 0x65, 0x70, 0x2d, 0x61, 0x6c, 0x69, 0x76,
-  0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x27, 0x43,
-  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2d, 0x54, 0x79, 0x70, 0x65, 0x27,
-  0x3a, 0x20, 0x27, 0x61, 0x70, 0x70, 0x6c, 0x69, 0x63, 0x61, 0x74, 0x69,
-  0x6f, 0x6e, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x27, 0x2c, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x27, 0x41, 0x63, 0x63, 0x65, 0x70, 0x74, 0x27,
-  0x3a, 0x20, 0x27, 0x74, 0x65, 0x78, 0x74, 0x2f, 0x65, 0x76, 0x65, 0x6e,
-  0x74, 0x2d, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x27, 0x2c, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x28, 0x70, 0x61, 0x72,
-  0x61, 0x6d, 0x73, 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x20,
-  0x3f, 0x20, 0x7b, 0x27, 0x41, 0x75, 0x74, 0x68, 0x6f, 0x72, 0x69, 0x7a,
-  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x60, 0x42, 0x65, 0x61,
-  0x72, 0x65, 0x72, 0x20, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
-  0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x7d, 0x60, 0x7d, 0x20,
-  0x3a, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x3a,
-  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e,
-  0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x0a, 0x20, 0x20, 0x7d, 0x29,
+  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3b, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x61, 0x70, 0x69, 0x5f, 0x75, 0x72, 0x6c, 0x20, 0x3d,
+  0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x2e, 0x61, 0x70, 0x69, 0x5f,
+  0x75, 0x72, 0x6c, 0x20, 0x7c, 0x7c, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x0a,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x63, 0x6f, 0x6e, 0x74, 0x72,
+  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20,
+  0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43,
+  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e,
+  0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e,
+  0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x44, 0x65, 0x66, 0x61, 0x75,
+  0x6c, 0x74, 0x73, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x2c, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x7d,
   0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72,
-  0x65, 0x61, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x70,
-  0x6f, 0x6e, 0x73, 0x65, 0x2e, 0x62, 0x6f, 0x64, 0x79, 0x2e, 0x67, 0x65,
-  0x74, 0x52, 0x65, 0x61, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64,
-  0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x54, 0x65, 0x78,
-  0x74, 0x44, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a,
-  0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65,
-  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x6c,
-  0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20,
-  0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20, 0x42, 0x75, 0x66,
-  0x66, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x61, 0x72, 0x74,
-  0x69, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x72, 0x65, 0x61, 0x64, 0x20, 0x6c,
-  0x69, 0x6e, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x63,
-  0x6f, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c,
-  0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x65,
-  0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61, 0x64, 0x28, 0x29, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x6f, 0x6e, 0x65, 0x29, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72,
-  0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x41,
-  0x64, 0x64, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f,
-  0x76, 0x65, 0x72, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x74, 0x6f, 0x20,
-  0x74, 0x68, 0x65, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x20,
-  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61, 0x74,
-  0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x6c, 0x65, 0x66,
-  0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x2b, 0x20, 0x64, 0x65, 0x63, 0x6f,
-  0x64, 0x65, 0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x28, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29,
-  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
-  0x43, 0x68, 0x65, 0x63, 0x6b, 0x20, 0x69, 0x66, 0x20, 0x74, 0x68, 0x65,
-  0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x72, 0x61, 0x63,
-  0x74, 0x65, 0x72, 0x20, 0x69, 0x73, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e,
-  0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x6e, 0x64, 0x73,
-  0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65, 0x42, 0x72, 0x65, 0x61,
-  0x6b, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e, 0x65, 0x6e, 0x64,
-  0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x5c, 0x6e, 0x27, 0x29, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x53,
-  0x70, 0x6c, 0x69, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65, 0x78,
-  0x74, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c,
-  0x69, 0x6e, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e,
-  0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x5c, 0x6e, 0x27, 0x29, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x49,
-  0x66, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x64,
-  0x6f, 0x65, 0x73, 0x6e, 0x27, 0x74, 0x20, 0x65, 0x6e, 0x64, 0x20, 0x77,
-  0x69, 0x74, 0x68, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62,
-  0x72, 0x65, 0x61, 0x6b, 0x2c, 0x20, 0x74, 0x68, 0x65, 0x6e, 0x20, 0x74,
-  0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x6c, 0x69, 0x6e, 0x65,
-  0x20, 0x69, 0x73, 0x20, 0x69, 0x6e, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
-  0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
-  0x53, 0x74, 0x6f, 0x72, 0x65, 0x20, 0x69, 0x74, 0x20, 0x69, 0x6e, 0x20,
-  0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x20,
-  0x62, 0x65, 0x20, 0x61, 0x64, 0x64, 0x65, 0x64, 0x20, 0x74, 0x6f, 0x20,
-  0x74, 0x68, 0x65, 0x20, 0x6e, 0x65, 0x78, 0x74, 0x20, 0x63, 0x68, 0x75,
-  0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x65, 0x6e,
+  0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x77,
+  0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x60, 0x24,
+  0x7b, 0x61, 0x70, 0x69, 0x5f, 0x75, 0x72, 0x6c, 0x7d, 0x2f, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x60, 0x2c, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x74, 0x68, 0x6f, 0x64, 0x3a,
+  0x20, 0x27, 0x50, 0x4f, 0x53, 0x54, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x62, 0x6f, 0x64, 0x79, 0x3a, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e,
+  0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x50, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, 0x61,
+  0x64, 0x65, 0x72, 0x73, 0x3a, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x27, 0x43, 0x6f, 0x6e, 0x6e, 0x65, 0x63, 0x74, 0x69, 0x6f,
+  0x6e, 0x27, 0x3a, 0x20, 0x27, 0x6b, 0x65, 0x65, 0x70, 0x2d, 0x61, 0x6c,
+  0x69, 0x76, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x27, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2d, 0x54, 0x79, 0x70,
+  0x65, 0x27, 0x3a, 0x20, 0x27, 0x61, 0x70, 0x70, 0x6c, 0x69, 0x63, 0x61,
+  0x74, 0x69, 0x6f, 0x6e, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x27, 0x2c, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x27, 0x41, 0x63, 0x63, 0x65, 0x70,
+  0x74, 0x27, 0x3a, 0x20, 0x27, 0x74, 0x65, 0x78, 0x74, 0x2f, 0x65, 0x76,
+  0x65, 0x6e, 0x74, 0x2d, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x27, 0x2c,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x28, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65,
+  0x79, 0x20, 0x3f, 0x20, 0x7b, 0x27, 0x41, 0x75, 0x74, 0x68, 0x6f, 0x72,
+  0x69, 0x7a, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x60, 0x42,
+  0x65, 0x61, 0x72, 0x65, 0x72, 0x20, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x7d, 0x60,
+  0x7d, 0x20, 0x3a, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61,
+  0x6c, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65,
+  0x72, 0x2e, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x0a, 0x20, 0x20,
+  0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x72, 0x65,
+  0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x2e, 0x62, 0x6f, 0x64, 0x79, 0x2e,
+  0x67, 0x65, 0x74, 0x52, 0x65, 0x61, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63,
+  0x6f, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x54,
+  0x65, 0x78, 0x74, 0x44, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x28, 0x29,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20,
+  0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65,
+  0x72, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20, 0x42,
+  0x75, 0x66, 0x66, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x61,
+  0x72, 0x74, 0x69, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x72, 0x65, 0x61, 0x64,
+  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x74, 0x72,
+  0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20,
+  0x28, 0x63, 0x6f, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73,
+  0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20,
+  0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61, 0x64, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x6f, 0x6e, 0x65,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x41, 0x64, 0x64, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x6c, 0x65, 0x66,
+  0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x74,
+  0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e,
+  0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x6c,
+  0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x2b, 0x20, 0x64, 0x65,
+  0x63, 0x6f, 0x64, 0x65, 0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65,
+  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x43, 0x68, 0x65, 0x63, 0x6b, 0x20, 0x69, 0x66, 0x20, 0x74,
+  0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x72,
+  0x61, 0x63, 0x74, 0x65, 0x72, 0x20, 0x69, 0x73, 0x20, 0x61, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x6e,
   0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65, 0x42, 0x72,
-  0x65, 0x61, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20,
-  0x3d, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x28,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65,
-  0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d,
-  0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20, 0x52, 0x65, 0x73, 0x65,
-  0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x69,
-  0x66, 0x20, 0x77, 0x65, 0x20, 0x68, 0x61, 0x76, 0x65, 0x20, 0x61, 0x20,
-  0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x20, 0x61,
-  0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x65, 0x6e, 0x64, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x2f, 0x2f, 0x20, 0x50, 0x61, 0x72, 0x73, 0x65, 0x20, 0x61, 0x6c,
-  0x6c, 0x20, 0x73, 0x73, 0x65, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x73,
-  0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x64, 0x64, 0x20, 0x74, 0x68, 0x65,
-  0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x72, 0x65, 0x67, 0x65, 0x78, 0x20, 0x3d, 0x20, 0x2f, 0x5e, 0x28, 0x5c,
-  0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73, 0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f,
-  0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f,
-  0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x69, 0x6e,
-  0x65, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x29, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x20, 0x3d, 0x20,
-  0x72, 0x65, 0x67, 0x65, 0x78, 0x2e, 0x65, 0x78, 0x65, 0x63, 0x28, 0x6c,
-  0x69, 0x6e, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x5b, 0x6d, 0x61, 0x74, 0x63,
-  0x68, 0x5b, 0x31, 0x5d, 0x5d, 0x20, 0x3d, 0x20, 0x6d, 0x61, 0x74, 0x63,
-  0x68, 0x5b, 0x32, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6e, 0x63, 0x65, 0x20,
-  0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20, 0x74, 0x68, 0x69, 0x73,
-  0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70,
-  0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x6a, 0x75, 0x73,
-  0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x74, 0x68, 0x65,
-  0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x64, 0x61, 0x74,
-  0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c,
-  0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f,
-  0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75,
-  0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e,
-  0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x79, 0x69,
-  0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x79, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x72, 0x65,
-  0x73, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x69, 0x66,
-  0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20, 0x61, 0x20, 0x73, 0x74,
-  0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x20, 0x66, 0x72, 0x6f,
-  0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2c, 0x20, 0x77, 0x65,
-  0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x20,
-  0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f,
-  0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x65, 0x61, 0x6b, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e, 0x65,
+  0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x5c, 0x6e, 0x27,
+  0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x53, 0x70, 0x6c, 0x69, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74,
+  0x65, 0x78, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x6c, 0x69, 0x6e,
+  0x65, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74,
+  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78,
+  0x74, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x5c, 0x6e, 0x27,
+  0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x49, 0x66, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65, 0x78, 0x74,
+  0x20, 0x64, 0x6f, 0x65, 0x73, 0x6e, 0x27, 0x74, 0x20, 0x65, 0x6e, 0x64,
+  0x20, 0x77, 0x69, 0x74, 0x68, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65,
+  0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x2c, 0x20, 0x74, 0x68, 0x65, 0x6e,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x6c, 0x69,
+  0x6e, 0x65, 0x20, 0x69, 0x73, 0x20, 0x69, 0x6e, 0x63, 0x6f, 0x6d, 0x70,
+  0x6c, 0x65, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x53, 0x74, 0x6f, 0x72, 0x65, 0x20, 0x69, 0x74, 0x20, 0x69,
+  0x6e, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x74,
+  0x6f, 0x20, 0x62, 0x65, 0x20, 0x61, 0x64, 0x64, 0x65, 0x64, 0x20, 0x74,
+  0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6e, 0x65, 0x78, 0x74, 0x20, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61, 0x74, 0x61,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21,
+  0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65,
+  0x42, 0x72, 0x65, 0x61, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65,
+  0x72, 0x20, 0x3d, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x2e, 0x70, 0x6f,
+  0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72,
+  0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20, 0x52, 0x65,
+  0x73, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72,
+  0x20, 0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x68, 0x61, 0x76, 0x65, 0x20,
+  0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b,
+  0x20, 0x61, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x65, 0x6e, 0x64, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x50, 0x61, 0x72, 0x73, 0x65, 0x20,
+  0x61, 0x6c, 0x6c, 0x20, 0x73, 0x73, 0x65, 0x20, 0x65, 0x76, 0x65, 0x6e,
+  0x74, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x64, 0x64, 0x20, 0x74,
+  0x68, 0x65, 0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c,
+  0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x20, 0x3d, 0x20, 0x2f, 0x5e,
+  0x28, 0x5c, 0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73, 0x28, 0x2e, 0x2a, 0x29,
+  0x24, 0x2f, 0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x20,
+  0x3d, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x2e, 0x65, 0x78, 0x65, 0x63,
+  0x28, 0x6c, 0x69, 0x6e, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x61, 0x74, 0x63,
+  0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x5b, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x5b, 0x31, 0x5d, 0x5d, 0x20, 0x3d, 0x20, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x5b, 0x32, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6e, 0x63,
+  0x65, 0x20, 0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20, 0x74, 0x68,
+  0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e,
+  0x63, 0x70, 0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x6a,
+  0x75, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x74,
+  0x68, 0x65, 0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74,
+  0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73,
+  0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x4a,
+  0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x79, 0x69, 0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x79, 0x69, 0x65, 0x6c, 0x64, 0x20,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20, 0x61, 0x20,
+  0x73, 0x74, 0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x20, 0x66,
+  0x72, 0x6f, 0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2c, 0x20,
+  0x77, 0x65, 0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72, 0x65, 0x61,
+  0x6b, 0x20, 0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73,
+  0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69,
+  0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20,
+  0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74,
+  0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e,
+  0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20,
+  0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72, 0x65,
+  0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75,
+  0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75,
+  0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x4a,
+  0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c,
+  0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x6d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x2e, 0x69, 0x6e, 0x63, 0x6c, 0x75, 0x64, 0x65, 0x73,
+  0x28, 0x27, 0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61, 0x76, 0x61,
+  0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x54, 0x68, 0x72, 0x6f, 0x77,
+  0x20, 0x61, 0x6e, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x74, 0x6f,
+  0x20, 0x62, 0x65, 0x20, 0x63, 0x61, 0x75, 0x67, 0x68, 0x74, 0x20, 0x62,
+  0x79, 0x20, 0x75, 0x70, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x20, 0x63,
+  0x61, 0x6c, 0x6c, 0x65, 0x72, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74,
+  0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72,
+  0x6f, 0x72, 0x28, 0x27, 0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61,
+  0x76, 0x61, 0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e,
+  0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x5b, 0x24,
+  0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f,
+  0x72, 0x2e, 0x63, 0x6f, 0x64, 0x65, 0x7d, 0x20, 0x2d, 0x20, 0x24, 0x7b,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72,
+  0x2e, 0x74, 0x79, 0x70, 0x65, 0x7d, 0x5d, 0x3a, 0x20, 0x24, 0x7b, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e,
+  0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28,
+  0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
+  0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f,
+  0x72, 0x20, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e,
+  0x6e, 0x61, 0x6d, 0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62,
+  0x6f, 0x72, 0x74, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
+  0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22,
+  0x2c, 0x20, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b,
+  0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c,
+  0x6c, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72,
+  0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61,
+  0x6c, 0x6c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e,
+  0x74, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61,
+  0x74, 0x20, 0x79, 0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75,
+  0x62, 0x73, 0x63, 0x72, 0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f,
+  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65,
+  0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27,
+  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
+  0x67, 0x65, 0x74, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a,
+  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61,
+  0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65,
+  0x6e, 0x65, 0x72, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x22, 0x2c, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69,
+  0x74, 0x65, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74,
+  0x61, 0x69, 0x6c, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f,
+  0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74,
+  0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72,
+  0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69,
+  0x67, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65,
+  0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e,
+  0x65, 0x77, 0x20, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79,
+  0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20,
+  0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f,
+  0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c,
+  0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68,
+  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b,
+  0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65,
+  0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64,
+  0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74,
+  0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67,
+  0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c,
+  0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67,
   0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
   0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e,
-  0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20,
-  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e,
-  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
-  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x66, 0x61,
-  0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74,
-  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72,
-  0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74,
-  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f,
-  0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75,
-  0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e,
-  0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67,
-  0x65, 0x2e, 0x69, 0x6e, 0x63, 0x6c, 0x75, 0x64, 0x65, 0x73, 0x28, 0x27,
-  0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61, 0x76, 0x61, 0x69, 0x6c,
-  0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x2f, 0x2f, 0x20, 0x54, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x61,
-  0x6e, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x74, 0x6f, 0x20, 0x62,
-  0x65, 0x20, 0x63, 0x61, 0x75, 0x67, 0x68, 0x74, 0x20, 0x62, 0x79, 0x20,
-  0x75, 0x70, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x20, 0x63, 0x61, 0x6c,
-  0x6c, 0x65, 0x72, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72,
-  0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72,
-  0x28, 0x27, 0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61, 0x76, 0x61,
-  0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72,
-  0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70,
-  0x70, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x5b, 0x24, 0x7b, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e,
-  0x63, 0x6f, 0x64, 0x65, 0x7d, 0x20, 0x2d, 0x20, 0x24, 0x7b, 0x72, 0x65,
-  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x74,
-  0x79, 0x70, 0x65, 0x7d, 0x5d, 0x3a, 0x20, 0x24, 0x7b, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x6d, 0x65,
-  0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x65, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
-  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x2e, 0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20,
-  0x24, 0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72,
-  0x6f, 0x72, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20,
-  0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61,
-  0x6d, 0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62, 0x6f, 0x72,
-  0x74, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
-  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20,
-  0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20,
-  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
-  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65,
-  0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c,
-  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75,
-  0x72, 0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x20,
-  0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20,
-  0x79, 0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75, 0x62, 0x73,
-  0x63, 0x72, 0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a,
-  0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a,
-  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70,
-  0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45,
-  0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x7d,
-  0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70,
-  0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65,
-  0x72, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c,
-  0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f,
-  0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65,
-  0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74, 0x61, 0x69,
-  0x6c, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f,
-  0x2f, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65,
-  0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
-  0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d,
-  0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
-  0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20,
-  0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74,
-  0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77,
-  0x20, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
-  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63,
-  0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66,
-  0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66,
-  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
-  0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63,
-  0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
-  0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74,
+  0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20,
+  0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
+  0x22, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b,
+  0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74,
+  0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b,
   0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65,
   0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73,
   0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e,
   0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65,
-  0x6e, 0x74, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22,
+  0x6e, 0x74, 0x28, 0x22, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22,
   0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20,
-  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d,
-  0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68,
-  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e,
-  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
-  0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75,
-  0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x67,
-  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
-  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64,
-  0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b,
-  0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61,
-  0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
-  0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
-  0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e,
-  0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74,
-  0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61,
-  0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77,
-  0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74,
-  0x28, 0x22, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20,
-  0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68,
-  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d,
-  0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
-  0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75,
-  0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x64,
-  0x6f, 0x6e, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61,
-  0x69, 0x6c, 0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
-  0x74, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d,
-  0x29, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
-  0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c,
-  0x6c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74,
-  0x75, 0x72, 0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73,
-  0x65, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c,
-  0x76, 0x65, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63,
-  0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74, 0x65, 0x78,
-  0x74, 0x2e, 0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x65, 0x73,
-  0x20, 0x6e, 0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74,
-  0x20, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65,
-  0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65,
-  0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74, 0x68, 0x65,
-  0x6e, 0x28, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20,
-  0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77,
-  0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a,
-  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x72,
-  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72,
-  0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72,
-  0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
-  0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72,
-  0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f,
-  0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20,
-  0x3d, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67,
-  0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77,
-  0x20, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61, 0x73, 0x79,
-  0x6e, 0x63, 0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c,
-  0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74,
+  0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74,
+  0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20,
+  0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
+  0x22, 0x64, 0x6f, 0x6e, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65,
+  0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x7d, 0x29, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74,
+  0x75, 0x72, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
+  0x67, 0x65, 0x74, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43,
+  0x61, 0x6c, 0x6c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d,
+  0x69, 0x73, 0x65, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73,
+  0x6f, 0x6c, 0x76, 0x65, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65,
+  0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74,
+  0x65, 0x78, 0x74, 0x2e, 0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f,
+  0x65, 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f,
+  0x72, 0x74, 0x20, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67,
+  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70,
+  0x6c, 0x65, 0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69,
+  0x73, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74,
+  0x68, 0x65, 0x6e, 0x28, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74,
+  0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x6f, 0x72, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28,
+  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e,
+  0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70,
+  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66,
+  0x69, 0x67, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e,
+  0x65, 0x77, 0x20, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61,
+  0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76,
+  0x65, 0x2c, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77,
+  0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61,
+  0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67,
+  0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x73, 0x6f, 0x6c, 0x76, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61,
+  0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65,
+  0x63, 0x74, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d,
+  0x3b, 0x0a, 0x0a, 0x2f, 0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64,
+  0x65, 0x70, 0x72, 0x65, 0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20,
+  0x2a, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d,
+  0x70, 0x6c, 0x65, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e,
+  0x63, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63,
+  0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69,
   0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75,
   0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28,
-  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68,
-  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e,
-  0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x6f,
-  0x6c, 0x76, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63,
-  0x68, 0x20, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74,
-  0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a,
-  0x0a, 0x2f, 0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64, 0x65, 0x70,
-  0x72, 0x65, 0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f,
-  0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c,
-  0x65, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20,
-  0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e,
-  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c,
-  0x6c, 0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20,
-  0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b,
-  0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x61,
-  0x72, 0x61, 0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c,
-  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63,
-  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29,
-  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c,
-  0x62, 0x61, 0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b,
-  0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47,
-  0x65, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c,
-  0x20, 0x69, 0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74,
-  0x68, 0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x20, 0x54,
-  0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65, 0x66, 0x75,
-  0x6c, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74, 0x69, 0x6e,
-  0x67, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78,
-  0x74, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64,
-  0x20, 0x73, 0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f,
-  0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20,
-  0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d,
-  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x67,
-  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
-  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70,
-  0x73, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65,
-  0x74, 0x63, 0x68, 0x28, 0x22, 0x2f, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x22,
-  0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20,
-  0x72, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f,
-  0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d,
-  0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x64, 0x65, 0x66, 0x61, 0x75,
-  0x6c, 0x74, 0x5f, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f,
-  0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a,
-  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e,
-  0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
-  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70,
+  0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20,
+  0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61,
+  0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f,
+  0x20, 0x47, 0x65, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64,
+  0x65, 0x6c, 0x20, 0x69, 0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e,
+  0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65,
+  0x66, 0x75, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74,
+  0x69, 0x6e, 0x67, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x78, 0x74, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61,
+  0x6e, 0x64, 0x20, 0x73, 0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78,
+  0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66,
+  0x6f, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x63,
+  0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21,
+  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
+  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x61, 0x70, 0x69,
+  0x5f, 0x75, 0x72, 0x6c, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69,
+  0x67, 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x75, 0x72, 0x6c, 0x20, 0x7c, 0x7c,
+  0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x20, 0x3d, 0x20, 0x61,
+  0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x60,
+  0x24, 0x7b, 0x61, 0x70, 0x69, 0x5f, 0x75, 0x72, 0x6c, 0x7d, 0x2f, 0x70,
+  0x72, 0x6f, 0x70, 0x73, 0x60, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28,
+  0x72, 0x20, 0x3d, 0x3e, 0x20, 0x72, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28,
+  0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e,
+  0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x5f, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61,
+  0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
+  0x73, 0x3b, 0x0a, 0x7d, 0x0a
 };
-size_t completion_js_len = 5796;
+unsigned int completion_js_len = 5909;

examples/server/index.js.hpp

@@ -1928,4 +1928,4 @@ unsigned char index_js[] = {
   0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x50, 0x74, 0x20, 0x61, 0x73,
   0x20, 0x75, 0x73, 0x65, 0x53, 0x74, 0x61, 0x74, 0x65, 0x7d, 0x3b, 0x0a
 };
-size_t index_js_len = 23136;
+unsigned int index_js_len = 23136;

examples/server/public/completion.js

@@ -21,6 +21,7 @@ let generation_settings = null;
 //
 export async function* llama(prompt, params = {}, config = {}) {
   let controller = config.controller;
+  const api_url = config.api_url || "";
 
   if (!controller) {
     controller = new AbortController();
@@ -28,7 +29,7 @@ export async function* llama(prompt, params = {}, config = {}) {
 
   const completionParams = { ...paramDefaults, ...params, prompt };
 
-  const response = await fetch("/completion", {
+  const response = await fetch(`${api_url}/completion`, {
     method: 'POST',
     body: JSON.stringify(completionParams),
     headers: {
@@ -193,9 +194,10 @@ export const llamaComplete = async (params, controller, callback) => {
 }
 
 // Get the model info from the server. This is useful for getting the context window and so on.
-export const llamaModelInfo = async () => {
+export const llamaModelInfo = async (config = {}) => {
   if (!generation_settings) {
-    const props = await fetch("/props").then(r => r.json());
+    const api_url = config.api_url || "";
+    const props = await fetch(`${api_url}/props`).then(r => r.json());
     generation_settings = props.default_generation_settings;
   }
   return generation_settings;

examples/server/public/index.html

@@ -199,10 +199,10 @@
   <script type="module">
     import {
       html, h, signal, effect, computed, render, useSignal, useEffect, useRef, Component
-    } from '/index.js';
+    } from './index.js';
 
-    import { llama } from '/completion.js';
-    import { SchemaConverter } from '/json-schema-to-grammar.mjs';
+    import { llama } from './completion.js';
+    import { SchemaConverter } from './json-schema-to-grammar.mjs';
     let selected_image = false;
     var slot_id = -1;
 
@@ -222,6 +222,7 @@
       temperature: 0.7,
       repeat_last_n: 256, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.18, // 1.0 = disabled
+      penalize_nl: false,
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
@@ -405,7 +406,7 @@
         throw new Error("already running");
       }
       controller.value = new AbortController();
-      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
+      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value, api_url: document.baseURI.replace(/\/+$/, '') })) {
         const data = chunk.data;
 
         if (data.stop) {
@@ -627,6 +628,7 @@
       const updateParams = (el) => params.value = { ...params.value, [el.target.name]: el.target.value }
       const updateParamsFloat = (el) => params.value = { ...params.value, [el.target.name]: parseFloat(el.target.value) }
       const updateParamsInt = (el) => params.value = { ...params.value, [el.target.name]: Math.floor(parseFloat(el.target.value)) }
+      const updateParamsBool = (el) => params.value = { ...params.value, [el.target.name]: el.target.checked }
 
       const grammarJsonSchemaPropOrder = signal('')
       const updateGrammarJsonSchemaPropOrder = (el) => grammarJsonSchemaPropOrder.value = el.target.value
@@ -670,6 +672,15 @@
         `
       };
 
+      const BoolField = ({ label, name, value }) => {
+        return html`
+          <div>
+            <label for="${name}">${label}</label>
+            <input type="checkbox" id="${name}" name="${name}" checked="${value}" onclick=${updateParamsBool} />
+          </div>
+        `
+      };
+
       const userTemplateReset = (e) => {
         e.preventDefault();
         userTemplateResetToDefaultAndApply()
@@ -769,6 +780,7 @@
             ${FloatField({ label: "Temperature", max: 2.0, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
             ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })}
             ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })}
+            ${BoolField({ label: "Penalize repetition of newlines", name: "penalize_nl", value: params.value.penalize_nl })}
             ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })}
             ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
             ${FloatField({ label: "Min-P sampling", max: 1.0, min: 0.0, name: "min_p", step: 0.01, value: params.value.min_p })}

examples/server/server.cpp

@@ -747,7 +747,8 @@ struct server_context {
         {
             const int32_t n_batch = llama_n_batch(ctx);
 
-            batch = llama_batch_init(n_batch, 0, params.n_parallel);
+            // only a single seq_id per token is needed
+            batch = llama_batch_init(n_batch, 0, 1);
         }
 
         metrics.init();
@@ -2188,8 +2189,6 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
     printf("                            KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
     printf("  -b N, --batch-size N      logical maximum batch size (default: %d)\n", params.n_batch);
     printf("  -ub N, --ubatch-size N    physical maximum batch size (default: %d)\n", params.n_ubatch);
-    printf("  --memory-f32              use f32 instead of f16 for memory key+value (default: disabled)\n");
-    printf("                            not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_supports_mlock()) {
         printf("  --mlock                   force system to keep model in RAM rather than swapping or compressing\n");
     }
@@ -2212,6 +2211,8 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
         printf("                            fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
         printf("  -mg i, --main-gpu i       the GPU to use for the model (with split-mode = none),\n");
         printf("                            or for intermediate results and KV (with split-mode = row)\n");
+        printf("  -nkvo, --no-kv-offload\n");
+        printf("                            disable KV offload\n");
     }
     printf("  -m FNAME, --model FNAME\n");
     printf("                            model path (default: %s)\n", params.model.c_str());
@@ -2497,6 +2498,8 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                     "See main README.md for information on enabling GPU BLAS support",
                     {{"n_gpu_layers", params.n_gpu_layers}});
             }
+        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
+            params.no_kv_offload = true;
         } else if (arg == "--split-mode" || arg == "-sm") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -3565,6 +3568,7 @@ int main(int argc, char ** argv) {
     sigemptyset (&sigint_action.sa_mask);
     sigint_action.sa_flags = 0;
     sigaction(SIGINT, &sigint_action, NULL);
+    sigaction(SIGTERM, &sigint_action, NULL);
 #elif defined (_WIN32)
     auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
         return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;

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

@@ -1114,7 +1114,10 @@ def start_server_background(context):
         server_args.append('--verbose')
     if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
         server_args.extend(['--log-format', "text"])
-    print(f"starting server with: {context.server_path} {server_args}")
+
+    args = [str(arg) for arg in [context.server_path, *server_args]]
+    print(f"bench: starting server with: {' '.join(args)}")
+
     flags = 0
     if 'nt' == os.name:
         flags |= subprocess.DETACHED_PROCESS
@@ -1130,16 +1133,14 @@ def start_server_background(context):
         [str(arg) for arg in [context.server_path, *server_args]],
         **pkwargs)
 
-    def log_stdout(process):
-        for line in iter(process.stdout.readline, b''):
-            print(line.decode('utf-8'), end='')
-    thread_stdout = threading.Thread(target=log_stdout, args=(context.server_process,))
+    def server_log(in_stream, out_stream):
+        for line in iter(in_stream.readline, b''):
+            print(line.decode('utf-8'), end='', file=out_stream)
+
+    thread_stdout = threading.Thread(target=server_log, args=(context.server_process.stdout, sys.stdout))
     thread_stdout.start()
 
-    def log_stderr(process):
-        for line in iter(process.stderr.readline, b''):
-            print(line.decode('utf-8'), end='', file=sys.stderr)
-    thread_stderr = threading.Thread(target=log_stderr, args=(context.server_process,))
+    thread_stderr = threading.Thread(target=server_log, args=(context.server_process.stderr, sys.stderr))
     thread_stderr.start()
 
     print(f"server pid={context.server_process.pid}, behave pid={os.getpid()}")

examples/server/utils.hpp

@@ -49,12 +49,23 @@ extern bool server_log_json;
 #define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
 #define LOG_INFO(   MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
 
+static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra);
+
 template <typename T>
 static T json_value(const json &body, const std::string &key, const T &default_value) {
     // Fallback null to default value
-    return body.contains(key) && !body.at(key).is_null()
-        ? body.value(key, default_value)
-        : default_value;
+    if (body.contains(key) && !body.at(key).is_null()){
+        try {
+            return body.value(key, default_value);
+        }
+        catch (nlohmann::json_abi_v3_11_3::detail::type_error const&){
+            std::string message = "Wrong type supplied for parameter '" + key + "'. Expected '" + typeid(default_value).name() + "', using default value.";
+            server_log("WARN", __func__, __LINE__, message.c_str(), body);
+            return default_value;
+        }
+    } else {
+        return default_value;
+    }
 }
 
 static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra) {

examples/speculative/speculative.cpp

@@ -65,7 +65,6 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    params.logits_all = true;
     std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
 
     // load the draft model

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1711163522,
-        "narHash": "sha256-YN/Ciidm+A0fmJPWlHBGvVkcarYWSC+s3NTPk/P+q3c=",
+        "lastModified": 1711703276,
+        "narHash": "sha256-iMUFArF0WCatKK6RzfUJknjem0H9m4KgorO/p3Dopkk=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "44d0940ea560dee511026a53f0e2e2cde489b4d4",
+        "rev": "d8fe5e6c92d0d190646fb9f1056741a229980089",
         "type": "github"
       },
       "original": {

flake.nix

@@ -145,6 +145,7 @@
             # the same path you would with an overlay.
             legacyPackages = {
               llamaPackages = pkgs.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
+              llamaPackagesWindows = pkgs.pkgsCross.mingwW64.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
               llamaPackagesCuda = pkgsCuda.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
               llamaPackagesRocm = pkgsRocm.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
             };
@@ -155,6 +156,7 @@
               {
                 default = config.legacyPackages.llamaPackages.llama-cpp;
                 vulkan = config.packages.default.override { useVulkan = true; };
+                windows = config.legacyPackages.llamaPackagesWindows.llama-cpp;
               }
               // lib.optionalAttrs pkgs.stdenv.isLinux {
                 opencl = config.packages.default.override { useOpenCL = true; };
@@ -168,9 +170,14 @@
               };
 
             # Packages exposed in `.#checks` will be built by the CI and by
-            # `nix flake check`. Currently we expose all packages, but we could
-            # make more granular choices
-            checks = config.packages;
+            # `nix flake check`.
+            #
+            # We could test all outputs e.g. as `checks = confg.packages`.
+            #
+            # TODO: Build more once https://github.com/ggerganov/llama.cpp/issues/6346 has been addressed
+            checks = {
+              inherit (config.packages) default vulkan;
+            };
           };
       };
 }

ggml-alloc.c

@@ -705,8 +705,13 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         struct ggml_tensor * leaf = graph->leafs[i];
         struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
         galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
-        galloc->leaf_allocs[i].leaf.offset = hn->offset;
-        galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        if (leaf->view_src || leaf->data) {
+            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
+            galloc->leaf_allocs[i].leaf.size_max = 0;
+        } else {
+            galloc->leaf_allocs[i].leaf.offset = hn->offset;
+            galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        }
     }
 
     // reallocate buffers if needed

ggml-common.h

@@ -377,6 +377,27 @@ typedef struct {
 } block_iq1_s;
 static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
 
+// 1.75 bpw
+typedef struct {
+    uint8_t  qs[QK_K/8];      // grid index, low 8 bits
+    uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
+#if QK_K == 64
+    ggml_half d;
+#endif
+    uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
+} block_iq1_m;
+#if QK_K == 64
+static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
+#else
+static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
+#endif
+
+// Used by IQ1_M quants
+typedef union {
+    ggml_half f16;
+    uint16_t  u16;
+} iq1m_scale_t;
+
 // Non-linear quants
 #define QK4_NL 32
 typedef struct {
@@ -426,10 +447,11 @@ static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_
 
 #define GGML_COMMON_IMPL
 #elif defined(GGML_COMMON_IMPL_SYCL)
+
 #include <cstdint>
 
-#define GGML_TABLE_BEGIN(type, name, size) static dpct::global_memory<const type, 1> name(sycl::range<1>(size), {
-#define GGML_TABLE_END() });
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
 
 #define GGML_COMMON_IMPL
 #endif
@@ -1050,6 +1072,7 @@ GGML_TABLE_END()
 
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
+#define IQ1M_DELTA 0.125f
 #if defined(GGML_COMMON_IMPL_C)
 GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
     0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,

ggml-cuda.cu

@@ -401,10 +401,8 @@ GGML_CALL static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t
 GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
-    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+    if (tensor->view_src != NULL) {
         assert(tensor->view_src->buffer->buft == buffer->buft);
-        tensor->backend = tensor->view_src->backend;
-        tensor->extra = tensor->view_src->extra;
         return;
     }
 
@@ -615,6 +613,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -643,6 +642,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -1960,227 +1960,49 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     }
 }
 
-#if 0
-template<typename ... Srcs>
-static __global__ void k_compute_batched_ptrs_id(
-        const void ** ptrs_src, void ** ptrs_dst,
-        int ne12, int ne13,
-        int ne23,
-        int nb02, int nb03,
-        int nb12, int nb13,
-        int nb2, int nb3,
-        int r2, int r3,
-        ggml_type src0_type, half * src0_as_f16, int64_t src0_ne,
-        const half * src1_f16, half * dst_f16,
-        const int32_t * ids, const int id,
-        Srcs... src0s) {
-
-    int i = ids[id];
-
-    half * src0_f16;
-    const void * srcs_ar[] = { (const half *) src0s... };
-    if (src0_type == GGML_TYPE_F16) {
-        src0_f16 = (half *) srcs_ar[i];
-    } else {
-        src0_f16 = src0_as_f16;
-        if (threadIdx.x == 0 && threadIdx.y == 0) {
-            const to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(src0_type);
-            to_fp16(srcs_ar[i], src0_f16, src0_ne, cudaStreamFireAndForget);
-        }
-    }
-
-    int i13 = blockIdx.x * blockDim.x + threadIdx.x;
-    int i12 = blockIdx.y * blockDim.y + threadIdx.y;
-
-    if (i13 >= ne13 || i12 >= ne12) {
-        return;
-    }
-
-    int i03 = i13 / r3;
-    int i02 = i12 / r2;
-
-    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_f16 + i02*nb02   + i03*nb03;
-    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_f16 + i12*nb12/2 + i13*nb13/2;
-    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)  dst_f16 + i12* nb2/2 + i13* nb3/2;
-}
-
-static void ggml_cuda_mul_mat_id_cublas(ggml_tensor * dst) {
-    const struct ggml_tensor * ids = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-    const struct ggml_tensor * src00 = dst->src[2];
-
-    const int id = dst->op_params[0];
-
-    GGML_ASSERT(!ggml_is_transposed(src00));
-    GGML_ASSERT(!ggml_is_transposed(src1));
-
-    GGML_ASSERT(src00->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00);
-    const int64_t ne01 = src00->ne[1];
-    const int64_t ne02 = src00->ne[2];
-    const int64_t ne03 = src00->ne[3];
-
-    //const int64_t nb01 = src00->nb[1];
-    const int64_t nb02 = src00->nb[2]; GGML_UNUSED(nb02);
-    const int64_t nb03 = src00->nb[3]; GGML_UNUSED(nb03);
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
-
-    //const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
-    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
-
-    const int64_t ne1 = ggml_nelements(src1);
-    const int64_t ne  = ggml_nelements(dst);
-
-    ggml_cuda_set_device(g_main_device);
-    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
-
-    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream));
-
-    //ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    //void * src0_ddq = src0_extra->data_device[g_main_device];
-    //half * src0_as_f16 = (half *) src0_ddq;
-
-    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
-
-    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
-    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
-
-    // convert src1 to fp16
-    const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
-    GGML_ASSERT(to_fp16_cuda != nullptr);
-
-    size_t src1_as = 0;
-    half * src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne1 * sizeof(half), &src1_as);
-    to_fp16_cuda(src1_ddf, src1_as_f16, ne1, main_stream);
-
-    size_t dst_as = 0;
-    half * dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as);
-
-    GGML_ASSERT(ne12 % ne02 == 0);
-    GGML_ASSERT(ne13 % ne03 == 0);
-
-    // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
-
-    const half alpha_f16 = 1.0f;
-    const half beta_f16  = 0.0f;
-
-    // use cublasGemmBatchedEx
-    const int ne23 = ne12*ne13;
-
-    const void ** ptrs_src = nullptr;
-          void ** ptrs_dst = nullptr;
-
-    size_t ptrs_src_s = 0;
-    size_t ptrs_dst_s = 0;
-
-    ptrs_src = (const void **) ggml_cuda_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s);
-    ptrs_dst = (      void **) ggml_cuda_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s);
-
-    int64_t src0_ne = ggml_nelements(src00);
-    half * src0_as_f16 = nullptr;
-    size_t src0_as = 0;
-    if (src00->type != GGML_TYPE_F16) {
-        src0_as_f16 = (half *) ggml_cuda_pool_malloc(src0_ne * sizeof(half), &src0_as);
-    }
-
-    static_assert(GGML_MAX_SRC == 6, "GGML_MAX_SRC == 6");
-    dim3 block_dims(ne13, ne12);
-    k_compute_batched_ptrs_id<<<1, block_dims, 0, main_stream>>>(
-            ptrs_src, ptrs_dst,
-            ne12, ne13,
-            ne23,
-            ne00*ne01*sizeof(half), ne00*ne01*ne02*sizeof(half),
-            nb12, nb13,
-            dst->nb[2], dst->nb[3],
-            r2, r3,
-            src00->type, src0_as_f16, src0_ne,
-            src1_as_f16, dst_f16,
-            (const int *)((ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device], id,
-            dst->src[2] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[2]->extra)->data_device[g_main_device] : nullptr,
-            dst->src[3] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[3]->extra)->data_device[g_main_device] : nullptr,
-            dst->src[4] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[4]->extra)->data_device[g_main_device] : nullptr,
-            dst->src[5] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[5]->extra)->data_device[g_main_device] : nullptr
-    );
-    CUDA_CHECK(cudaGetLastError());
-
-    CUBLAS_CHECK(
-    cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-            ne01, ne11, ne10,
-            &alpha_f16, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F, ne00,
-                        (const void **) (ptrs_src + 1*ne23), CUDA_R_16F, ne10,
-            &beta_f16,  (      void **) (ptrs_dst + 0*ne23), CUDA_R_16F, ne01,
-            ne23,
-            CUBLAS_COMPUTE_16F,
-            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-
-    if (src0_as != 0) {
-        ggml_cuda_pool_free(src0_as_f16, src0_as);
-    }
-    if (ptrs_src_s != 0) {
-        ggml_cuda_pool_free(ptrs_src, ptrs_src_s);
-    }
-    if (ptrs_dst_s != 0) {
-        ggml_cuda_pool_free(ptrs_dst, ptrs_dst_s);
-    }
-
-    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
-    to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream);
-
-    ggml_cuda_pool_free(src1_as_f16, src1_as);
-    ggml_cuda_pool_free(dst_f16, dst_as);
-}
-#endif
-
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-#if 0
-    ggml_cuda_mul_mat_id_cublas(dst);
-    // TODO: mmq/mmv support
-#endif
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * ids  = dst->src[2];
+
+    GGML_ASSERT(!ggml_backend_buffer_is_cuda_split(src0->buffer) && "mul_mat_id does not support split buffers");
 
     cudaStream_t stream = ctx.stream();
 
     const size_t nb11 = src1->nb[1];
     const size_t nb1  =  dst->nb[1];
 
-    const struct ggml_tensor * ids = src0;
     const int32_t id = ((int32_t *) dst->op_params)[0];
-    const int32_t n_as = ((int32_t *) dst->op_params)[1];
+    const int32_t n_as = src0->ne[2];
 
     std::vector<char> ids_host(ggml_nbytes(ids));
     const char * ids_dev = (const char *) ids->data;
     CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
     CUDA_CHECK(cudaStreamSynchronize(stream));
 
+    ggml_tensor src0_row = *src0;
     ggml_tensor src1_row = *src1;
     ggml_tensor dst_row = *dst;
 
+    char * src0_original = (char *) src0->data;
     char * src1_original = (char *) src1->data;
     char * dst_original  = (char *)  dst->data;
 
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[3] = src0->nb[2];
+
     if (src1->ne[1] == 1) {
         for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
             const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
 
             GGML_ASSERT(row_id >= 0 && row_id < n_as);
 
-            const struct ggml_tensor * src0_row = dst->src[row_id + 2];
-
+            src0_row.data = src0_original + row_id*src0->nb[2];
             src1_row.data = src1_original + i01*src1->nb[1];
             dst_row.data  =  dst_original + i01*dst->nb[1];
 
-            ggml_cuda_mul_mat(ctx, src0_row, &src1_row, &dst_row);
+            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
         }
     } else {
         ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
@@ -2190,8 +2012,6 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         dst_row.data  =  dst_contiguous.get();
 
         for (int32_t row_id = 0; row_id < n_as; ++row_id) {
-            const struct ggml_tensor * src0_row = dst->src[row_id + 2];
-
             int64_t num_src1_rows = 0;
             for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
                 const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
@@ -2211,6 +2031,8 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
                 continue;
             }
 
+            src0_row.data = src0_original + row_id*src0->nb[2];
+
             src1_row.ne[1] = num_src1_rows;
             dst_row.ne[1] = num_src1_rows;
 
@@ -2222,7 +2044,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
             dst_row.nb[2] = num_src1_rows*nb1;
             dst_row.nb[3] = num_src1_rows*nb1;
 
-            ggml_cuda_mul_mat(ctx, src0_row, &src1_row, &dst_row);
+            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
 
             num_src1_rows = 0;
             for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
@@ -2387,7 +2209,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
     cudaError_t err = cudaGetLastError();
     if (err != cudaSuccess) {
         fprintf(stderr, "%s: %s failed\n", __func__, ggml_op_desc(dst));
-        GGML_ASSERT(false);
+        CUDA_CHECK(err);
     }
 
     return true;
@@ -2503,7 +2325,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
 
-        if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
             continue;
         }
 
@@ -2560,7 +2382,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 ggml_type a_type = a->type;
                 if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
                     a_type == GGML_TYPE_IQ1_S   || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S   ||
-                    a_type == GGML_TYPE_IQ2_S   || a_type == GGML_TYPE_IQ4_XS) {
+                    a_type == GGML_TYPE_IQ1_M   || a_type == GGML_TYPE_IQ2_S  || a_type == GGML_TYPE_IQ4_XS) {
                     if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
                         return false;
                     }

ggml-cuda/argsort.cu

@@ -8,32 +8,41 @@ static inline __device__ void ggml_cuda_swap(T & a, T & b) {
 }
 
 template<ggml_sort_order order>
-static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols) {
+static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
     // bitonic sort
     int col = threadIdx.x;
     int row = blockIdx.y;
 
-    if (col >= ncols) return;
+    if (col >= ncols_pad) {
+        return;
+    }
 
     const float * x_row = x + row * ncols;
-    int * dst_row = dst + row * ncols;
+    extern __shared__ int dst_row[];
 
     // initialize indices
-    if (col < ncols) {
-        dst_row[col] = col;
-    }
+    dst_row[col] = col;
+
     __syncthreads();
 
-    for (int k = 2; k <= ncols; k *= 2) {
+    for (int k = 2; k <= ncols_pad; k *= 2) {
         for (int j = k / 2; j > 0; j /= 2) {
             int ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
+                    if (dst_row[col] >= ncols ||
+                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    ) {
                         ggml_cuda_swap(dst_row[col], dst_row[ixj]);
                     }
                 } else {
-                    if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
+                    if (dst_row[ixj] >= ncols ||
+                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    ) {
                         ggml_cuda_swap(dst_row[col], dst_row[ixj]);
                     }
                 }
@@ -41,18 +50,35 @@ static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int n
             __syncthreads();
         }
     }
+
+    // copy the result to dst without the padding
+    if (col < ncols) {
+        dst[row * ncols + col] = dst_row[col];
+    }
+}
+
+static int next_power_of_2(int x) {
+    int n = 1;
+    while (n < x) {
+        n *= 2;
+    }
+    return n;
 }
 
 static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
     // bitonic sort requires ncols to be power of 2
-    GGML_ASSERT((ncols & (ncols - 1)) == 0);
+    const int ncols_pad = next_power_of_2(ncols);
 
-    const dim3 block_dims(ncols, 1, 1);
+    const dim3 block_dims(ncols_pad, 1, 1);
     const dim3 block_nums(1, nrows, 1);
+    const size_t shared_mem = ncols_pad * sizeof(int);
+
+    GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
+
     if (order == GGML_SORT_ORDER_ASC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else if (order == GGML_SORT_ORDER_DESC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else {
         GGML_ASSERT(false);
     }

ggml-cuda/common.cuh

@@ -1,7 +1,8 @@
 #pragma once
 
-#include "../ggml.h"
-#include "../ggml-cuda.h"
+#include "ggml.h"
+#include "ggml-cuda.h"
+
 #include <memory>
 
 #if defined(GGML_USE_HIPBLAS)
@@ -11,7 +12,7 @@
 #define GGML_COMMON_DECL_CUDA
 #define GGML_COMMON_IMPL_CUDA
 #endif
-#include "../ggml-common.h"
+#include "ggml-common.h"
 
 #include <cstdio>
 #include <array>

ggml-cuda/convert.cu

@@ -373,7 +373,7 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 #else
-    assert(false);
+    NO_DEVICE_CODE;
 #endif
 
 }
@@ -395,7 +395,7 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 #else
-    assert(false);
+    NO_DEVICE_CODE;
 #endif
 
 }
@@ -416,7 +416,7 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 #else
-    assert(false);
+    NO_DEVICE_CODE;
 #endif
 
 }
@@ -444,7 +444,7 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
 #else
-    assert(false);
+    NO_DEVICE_CODE;
 #endif
 
 }
@@ -470,7 +470,7 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
 #else
-    assert(false);
+    NO_DEVICE_CODE;
 #endif
 
 }
@@ -496,11 +496,42 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
         y[j] = d * (q[j] + delta);
     }
 #else
-    assert(false);
+    NO_DEVICE_CODE;
 #endif
 
 }
 
+template<typename dst_t>
+static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int i   = blockIdx.x;
+    const block_iq1_m * x = (const block_iq1_m  *) vx;
+
+    const int tid = threadIdx.x;
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * sc = (const uint16_t *)x[i].scales;
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const int ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
+    const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
+    const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+#else
+    NO_DEVICE_CODE;
+#endif
+
+}
+
+
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
@@ -658,6 +689,12 @@ static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k,
     dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template<typename dst_t>
+static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
+}
+
 template<typename dst_t>
 static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int nb = (k + QK_K - 1) / QK_K;
@@ -724,6 +761,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq3_xxs_cuda;
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_cuda;
         case GGML_TYPE_IQ4_NL:
             return dequantize_row_iq4_nl_cuda;
         case GGML_TYPE_IQ4_XS:
@@ -769,6 +808,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq3_xxs_cuda;
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_cuda;
         case GGML_TYPE_IQ4_NL:
             return dequantize_row_iq4_nl_cuda;
         case GGML_TYPE_IQ4_XS:

ggml-cuda/dmmv.cu

@@ -2,14 +2,6 @@
 #include "dequantize.cuh"
 #include "convert.cuh"
 
-// dmmv = dequantize_mul_mat_vec
-#ifndef GGML_CUDA_DMMV_X
-#define GGML_CUDA_DMMV_X 32
-#endif
-#ifndef GGML_CUDA_MMV_Y
-#define GGML_CUDA_MMV_Y 1
-#endif
-
 #ifndef K_QUANTS_PER_ITERATION
 #define K_QUANTS_PER_ITERATION 2
 #else

ggml-cuda/dmmv.cuh

@@ -1,5 +1,16 @@
 #include "common.cuh"
 
+// dmmv = dequantize_mul_mat_vec
+
+// TODO: remove this?
+#ifndef GGML_CUDA_DMMV_X
+#define GGML_CUDA_DMMV_X 32
+#endif
+
+#ifndef GGML_CUDA_MMV_Y
+#define GGML_CUDA_MMV_Y 1
+#endif
+
 void ggml_cuda_op_dequantize_mul_mat_vec(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,

ggml-cuda/mmvq.cu

@@ -282,6 +282,14 @@ static void mul_mat_vec_iq1_s_q8_1_cuda(
         (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
 }
 
+static void mul_mat_vec_iq1_m_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI1_S, block_iq1_m, 1, vec_dot_iq1_m_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
 static void mul_mat_vec_iq4_nl_q8_1_cuda(
     const void * vx, const void * vy, float * dst,
     const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
@@ -373,6 +381,9 @@ void ggml_cuda_op_mul_mat_vec_q(
         case GGML_TYPE_IQ1_S:
             mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
             break;
+        case GGML_TYPE_IQ1_M:
+            mul_mat_vec_iq1_m_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
         case GGML_TYPE_IQ4_NL:
             mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
             break;

ggml-cuda/vecdotq.cuh

@@ -961,8 +961,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     return d * (sumi1 + sumi2);
 #endif
 #else
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 }
 
@@ -1001,13 +1000,11 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
 #else
     GGML_UNUSED(ksigns64);
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 #else
     GGML_UNUSED(ksigns64);
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 }
 
@@ -1049,13 +1046,11 @@ static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
     return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
 #else
     GGML_UNUSED(ksigns64);
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 #else
     GGML_UNUSED(ksigns64);
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 }
 
@@ -1085,12 +1080,10 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
     const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f;
     return d * sumi;
 #else
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 #else
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 }
 
@@ -1119,12 +1112,10 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
     const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds);
     return d * sumi;
 #else
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 #else
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 }
 
@@ -1159,8 +1150,50 @@ static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
     const float m = d1q * __high2float(bq8_1[ib32].ds);
     return d * sumi + m * delta;
 #else
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if QK_K == 256
+    const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
+
+    const int ib32 = iqs;
+    int   sumi[2] = {0, 0};
+    float sumf[2] = {0.f, 0.f};
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    const int * q8 = (const int *)bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8)));
+        int grid0 = grid[0] & 0x0f0f0f0f;
+        int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
+        sumi[l/2] = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi[l/2]));
+        const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
+        const int sumy = __dp4a(q8[2*l+1], 0x01010101, __dp4a(q8[2*l+0], 0x01010101, 0));
+        sumf[l/2] += delta*sumy;
+    }
+#else
+    const int8_t * q8 = bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
+        int sumy = 0;
+        for (int j = 0; j < 4; ++j) {
+            sumi[l/2] += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4);
+            sumy += q8[j] + q8[j+4];
+        }
+        const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
+        sumf[l/2] += delta*sumy;
+        q8 += 8;
+    }
+#endif
+    iq1m_scale_t scale;
+    const uint16_t * sc = (const uint16_t *)bq1->scales;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds);
+    return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
+#else
+    NO_DEVICE_CODE;
 #endif
 }
 
@@ -1223,27 +1256,6 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
     const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
     const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
 
-    //// iqs is 0...7
-    //const int ib64 = iqs/2;
-    //const int il = iqs%2;
-    //const int32_t  * q8_1 = (const int *)bq8_1[2*ib64+0].qs + 2*il;
-    //const int32_t  * q8_2 = (const int *)bq8_1[2*ib64+1].qs + 2*il;
-    //const uint32_t * q4_1 = (const uint32_t *)bq4->qs + 8*ib64 + 2*il;
-    //const uint32_t * q4_2 = q4_1 + 4;
-    //const int8_t ls1 = (bq4->scales_l[ib64] & 0xf) | (((bq4->scales_h >> (4*ib64+0)) & 3) << 4);
-    //const int8_t ls2 = (bq4->scales_l[ib64] >>  4) | (((bq4->scales_h >> (4*ib64+2)) & 3) << 4);
-    //const float d1 = (float)bq4->d * (ls1 - 32) * __low2float(bq8_1[2*ib64+0].ds);
-    //const float d2 = (float)bq4->d * (ls2 - 32) * __low2float(bq8_1[2*ib64+1].ds);
-    //int v1, v2;
-    //int sumi1 = 0, sumi2 = 0;
-    //for (int j = 0; j < 2; ++j) {
-    //    get_int_from_table_16(q4_1[j], values, v1, v2);
-    //    sumi1 = __dp4a(v2, q8_1[j+4], __dp4a(v1, q8_1[j+0], sumi1));
-    //    get_int_from_table_16(q4_2[j], values, v1, v2);
-    //    sumi2 = __dp4a(v2, q8_2[j+4], __dp4a(v1, q8_2[j+0], sumi2));
-    //}
-    //return d1 * sumi1 + d2 * sumi2;
-
     // iqs is 0...7
     const int ib32 = iqs;
     const int32_t  * q8 = (const int *)bq8_1[ib32].qs;
@@ -1259,24 +1271,8 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
     }
     return d * (sumi1 + sumi2);
 
-    //// iqs is 0...15
-    //const int ib32 = iqs/2;
-    //const int il = iqs%2;
-    //const int32_t  * q8 = (const int *)bq8_1[ib32].qs + 2*il;
-    //const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32 + 2*il;
-    //const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
-    //const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
-    //int v1, v2;
-    //int sumi1 = 0, sumi2 = 0;
-    //for (int j = 0; j < 2; ++j) {
-    //    get_int_from_table_16(q4[j], values, v1, v2);
-    //    sumi1 = __dp4a(v1, q8[j+0], sumi1);
-    //    sumi2 = __dp4a(v2, q8[j+4], sumi2);
-    //}
-    //return d * (sumi1 + sumi2);
 #else
-    assert(false);
-    return 0.f;
+    NO_DEVICE_CODE;
 #endif
 #else
     return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);

ggml-kompute.cpp

@@ -1430,6 +1430,10 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
             struct ggml_tensor * dst = gf->nodes[i];
             GGML_ASSERT(dst->data != nullptr);
 
+            if (ggml_is_empty(dst)) {
+                continue;
+            }
+
             switch (dst->op) {
                 case GGML_OP_NONE:
                 case GGML_OP_RESHAPE:

ggml-metal.m

@@ -64,6 +64,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
@@ -91,6 +92,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
@@ -114,6 +116,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
@@ -134,6 +137,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
@@ -154,6 +158,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F32,
@@ -490,6 +495,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,            get_rows_iq3_s,         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,            get_rows_iq2_s,         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,            get_rows_iq1_s,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,            get_rows_iq1_m,         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,           get_rows_iq4_nl,        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,           get_rows_iq4_xs,        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,              get_rows_i32,           true);
@@ -517,6 +523,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,          mul_mv_iq3_s_f32,       ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,          mul_mv_iq2_s_f32,       ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,          mul_mv_iq1_s_f32,       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,          mul_mv_iq1_m_f32,       ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,         mul_mv_iq4_nl_f32,      ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,         mul_mv_iq4_xs_f32,      ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,         mul_mv_id_f32_f32,      ctx->support_simdgroup_reduction);
@@ -540,6 +547,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,       mul_mv_id_iq3_s_f32,    ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,       mul_mv_id_iq2_s_f32,    ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,       mul_mv_id_iq1_s_f32,    ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,       mul_mv_id_iq1_m_f32,    ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,      mul_mv_id_iq4_nl_f32,   ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,      mul_mv_id_iq4_xs_f32,   ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,            mul_mm_f32_f32,         ctx->support_simdgroup_mm);
@@ -560,6 +568,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,          mul_mm_iq3_s_f32,       ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,          mul_mm_iq2_s_f32,       ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,          mul_mm_iq1_s_f32,       ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,          mul_mm_iq1_m_f32,       ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,         mul_mm_iq4_nl_f32,      ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,         mul_mm_iq4_xs_f32,      ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,         mul_mm_id_f32_f32,      ctx->support_simdgroup_mm);
@@ -580,6 +589,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,       mul_mm_id_iq3_s_f32,    ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,       mul_mm_id_iq2_s_f32,    ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,       mul_mm_id_iq1_s_f32,    ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,       mul_mm_id_iq1_m_f32,    ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,      mul_mm_id_iq4_nl_f32,   ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,      mul_mm_id_iq4_xs_f32,   ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                  rope_f32,               true);
@@ -837,6 +847,10 @@ static enum ggml_status ggml_metal_graph_compute(
             struct ggml_tensor * src2 = gf->nodes[i]->src[2];
             struct ggml_tensor * dst  = gf->nodes[i];
 
+            if (ggml_is_empty(dst)) {
+                continue;
+            }
+
             switch (dst->op) {
                 case GGML_OP_NONE:
                 case GGML_OP_RESHAPE:
@@ -1421,6 +1435,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
+                                case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
                                 case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
@@ -1575,6 +1590,12 @@ static enum ggml_status ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
                                     } break;
+                                case GGML_TYPE_IQ1_M:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
+                                    } break;
                                 case GGML_TYPE_IQ4_NL:
                                     {
                                         nth0 = 4;
@@ -1619,9 +1640,9 @@ static enum ggml_status ggml_metal_graph_compute(
                             [encoder setBytes:&r2   length:sizeof(r2)   atIndex:17];
                             [encoder setBytes:&r3   length:sizeof(r3)   atIndex:18];
 
-                            if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1  ||
-                                src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0 ||
-                                src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ2_S) {
+                            if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
+                                src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
+                                src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@@ -1664,37 +1685,31 @@ static enum ggml_status ggml_metal_graph_compute(
                     {
                         //GGML_ASSERT(ne00 == ne10);
                         //GGML_ASSERT(ne03 == ne13);
-
-                        GGML_ASSERT(src0t == GGML_TYPE_I32);
-
-                        const int n_as = ((int32_t *) dst->op_params)[1];
-
-                        // TODO: make this more general
-                        GGML_ASSERT(n_as <= 8);
+                        const int n_as = src0->ne[2];
 
                         // max size of the src1ids array in the kernel shared buffer
                         GGML_ASSERT(ne11 <= 4096);
 
-                        const int64_t  ne20 = src2 ? src2->ne[0] : 0;
-                        const int64_t  ne21 = src2 ? src2->ne[1] : 0;
-                        const int64_t  ne22 = src2 ? src2->ne[2] : 0;
-                        const int64_t  ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
+                        // src2 = ids
+                        const int64_t  ne20 = src2->ne[0]; GGML_UNUSED(ne20);
+                        const int64_t  ne21 = src2->ne[1];
+                        const int64_t  ne22 = src2->ne[2]; GGML_UNUSED(ne22);
+                        const int64_t  ne23 = src2->ne[3]; GGML_UNUSED(ne23);
 
-                        const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
-                        const uint64_t nb21 = src2 ? src2->nb[1] : 0;
-                        const uint64_t nb22 = src2 ? src2->nb[2] : 0;
-                        const uint64_t nb23 = src2 ? src2->nb[3] : 0; GGML_UNUSED(nb23);
+                        const uint64_t nb20 = src2->nb[0]; GGML_UNUSED(nb20);
+                        const uint64_t nb21 = src2->nb[1];
+                        const uint64_t nb22 = src2->nb[2]; GGML_UNUSED(nb22);
+                        const uint64_t nb23 = src2->nb[3]; GGML_UNUSED(nb23);
 
-                        const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
+                        const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
 
-                        GGML_ASSERT(!ggml_is_transposed(src2));
+                        GGML_ASSERT(src2t == GGML_TYPE_I32);
+
+                        GGML_ASSERT(!ggml_is_transposed(src0));
                         GGML_ASSERT(!ggml_is_transposed(src1));
 
                         GGML_ASSERT(src1t == GGML_TYPE_F32);
 
-                        const uint r2 = ne12/ne22;
-                        const uint r3 = ne13/ne23;
-
                         // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                         // to the matrix-vector kernel
                         int ne11_mm_min = n_as;
@@ -1702,7 +1717,10 @@ static enum ggml_status ggml_metal_graph_compute(
                         const int idx = ((int32_t *) dst->op_params)[0];
 
                         // batch size
-                        GGML_ASSERT(ne01 == ne11);
+                        GGML_ASSERT(ne21 == ne11); // ?
+                        GGML_ASSERT(ne12 == 1 && ne13 == 1); // no broadcasting
+                        const uint r2 = 1;
+                        const uint r3 = 1;
 
                         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
@@ -1711,7 +1729,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         //       indirect matrix multiplication
                         // !!!
                         if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
-                            ne20 % 32 == 0 && ne20 >= 64 &&
+                            ne00 % 32 == 0 && ne00 >= 64 &&
                             ne11 > ne11_mm_min) {
 
                             // some Metal matrix data types require aligned pointers
@@ -1724,7 +1742,7 @@ static enum ggml_status ggml_metal_graph_compute(
 
                             id<MTLComputePipelineState> pipeline = nil;
 
-                            switch (src2->type) {
+                            switch (src0->type) {
                                 case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32    ].pipeline; break;
                                 case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32    ].pipeline; break;
                                 case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32   ].pipeline; break;
@@ -1743,6 +1761,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32  ].pipeline; break;
+                                case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
                                 case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
@@ -1752,36 +1771,27 @@ static enum ggml_status ggml_metal_graph_compute(
                             [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                            [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:3];
-                            [encoder setBytes:&ne20    length:sizeof(ne20) atIndex:4];
-                            [encoder setBytes:&ne22    length:sizeof(ne22) atIndex:5];
-                            [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:6];
-                            [encoder setBytes:&nb22    length:sizeof(nb22) atIndex:7];
-                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:8];
-                            [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:9];
-                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:10];
-                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:11];
-                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:12];
-                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
-                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
-                            [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:15];
-                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:16];
-                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:17];
-                            [encoder setBytes:&idx     length:sizeof(idx)  atIndex:18];
-                            // TODO: how to make this an array? read Metal docs
-                            for (int j = 0; j < 8; ++j) {
-                                // NOTE: this is done like this to avoid uninitialized kernel arguments when n_as < 8
-                                struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
-
-                                size_t offs_src_cur = 0;
-                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur);
-
-                                [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j];
-                            }
+                            [encoder setBuffer:id_src2 offset:offs_src2    atIndex:3];
+                            [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:4];
+                            [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:5];
+                            [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:6];
+                            [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:7];
+                            [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:8];
+                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:9];
+                            [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:10];
+                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:11];
+                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:12];
+                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:13];
+                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:14];
+                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:15];
+                            [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:16];
+                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:17];
+                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:18];
+                            [encoder setBytes:&idx     length:sizeof(idx)  atIndex:19];
 
                             [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + 2*ne11, 16) atIndex:0];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne21 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne01 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                         } else {
                             int nth0 = 32;
                             int nth1 = 1;
@@ -1791,7 +1801,7 @@ static enum ggml_status ggml_metal_graph_compute(
                             id<MTLComputePipelineState> pipeline = nil;
 
                             // use custom matrix x vector kernel
-                            switch (src2t) {
+                            switch (src0t) {
                                 case GGML_TYPE_F32:
                                     {
                                         GGML_ASSERT(src1t == GGML_TYPE_F32);
@@ -1900,6 +1910,12 @@ static enum ggml_status ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
                                     } break;
+                                case GGML_TYPE_IQ1_M:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline;
+                                    } break;
                                 case GGML_TYPE_IQ4_NL:
                                     {
                                         nth0 = 4;
@@ -1919,8 +1935,8 @@ static enum ggml_status ggml_metal_graph_compute(
                                     }
                             };
 
-                            if (ggml_is_quantized(src2t)) {
-                                GGML_ASSERT(ne20 >= nth0*nth1);
+                            if (ggml_is_quantized(src0t)) {
+                                GGML_ASSERT(ne00 >= nth0*nth1);
                             }
 
                             const int64_t _ne1 = 1; // kernels needs a reference in constant memory
@@ -1929,75 +1945,66 @@ static enum ggml_status ggml_metal_graph_compute(
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:3];
-                            [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
-                            [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
-                            [encoder setBytes:&ne22 length:sizeof(ne22) atIndex:6];
-                            [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:7];
-                            [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:8];
-                            [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:9];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
-                            [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:11];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
-                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:17];
-                            [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:18];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:19];
-                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:20];
-                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:21];
-                            [encoder setBytes:&idx  length:sizeof(idx)  atIndex:22];
-                            // TODO: how to make this an array? read Metal docs
-                            for (int j = 0; j < 8; ++j) {
-                                // NOTE: this is done like this to avoid uninitialized kernel arguments when n_as < 8
-                                struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
+                            [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
+                            [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:4];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:5];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:6];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:7];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:8];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:9];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:10];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                            [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:12];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:18];
+                            [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:19];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:20];
+                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:21];
+                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:22];
+                            [encoder setBytes:&idx  length:sizeof(idx)  atIndex:23];
 
-                                size_t offs_src_cur = 0;
-                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur);
-
-                                [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
+                            if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
+                                src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
+                                src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-
-                            if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1  ||
-                                src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1  || src2t == GGML_TYPE_Q8_0 ||
-                                src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S || src2t == GGML_TYPE_IQ2_S) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
-                                const int mem_size = src2t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                            else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
+                                const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_IQ3_XXS || src2t == GGML_TYPE_IQ3_S) {
-                                const int mem_size = src2t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                            else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
+                                const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_IQ4_NL || src2t == GGML_TYPE_IQ4_XS) {
+                            else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
                                 const int mem_size = 32*sizeof(float);
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_Q4_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            else if (src0t == GGML_TYPE_Q4_K) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_Q3_K) {
+                            else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 1)/2, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #else
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #endif
                             }
-                            else if (src2t == GGML_TYPE_Q5_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            else if (src0t == GGML_TYPE_Q5_K) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_Q6_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 1)/2, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            else if (src0t == GGML_TYPE_Q6_K) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             } else {
                                 const int64_t ny = (_ne1 + nrows - 1)/nrows;
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne21, ny, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                         }
                     } break;
@@ -2024,6 +2031,7 @@ static enum ggml_status ggml_metal_graph_compute(
                             case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S  ].pipeline; break;
                             case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S  ].pipeline; break;
                             case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S  ].pipeline; break;
+                            case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M  ].pipeline; break;
                             case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
                             case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
                             case GGML_TYPE_I32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32    ].pipeline; break;
@@ -2403,6 +2411,16 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
 
+                        // bitonic sort requires the number of elements to be power of 2
+                        int64_t ne00_padded = 1;
+                        while (ne00_padded < ne00) {
+                            ne00_padded *= 2;
+                        }
+
+                        // Metal kernels require the buffer size to be multiple of 16 bytes
+                        // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
+                        const int mem_size = GGML_PAD(ne00_padded*sizeof(int32_t), 16);
+
                         id<MTLComputePipelineState> pipeline = nil;
 
                         switch (order) {
@@ -2412,11 +2430,13 @@ static enum ggml_status ggml_metal_graph_compute(
                         };
 
                         [encoder setComputePipelineState:pipeline];
-                        [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                        [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                        [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                        [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
+                        [encoder setBuffer:id_dst      offset:offs_dst         atIndex:1];
+                        [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:2];
+                        [encoder setBytes:&ne00_padded length:sizeof( int64_t) atIndex:3];
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
 
-                        [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00, 1, 1)];
+                        [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00_padded, 1, 1)];
                     } break;
                 case GGML_OP_LEAKY_RELU:
                     {

ggml-opencl.cpp

@@ -2234,6 +2234,11 @@ static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(gg
 static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
     for (int i = 0; i < graph->n_nodes; ++i) {
         ggml_tensor * node = graph->nodes[i];
+
+        if (ggml_is_empty(node)) {
+            continue;
+        }
+
         switch (node->op) {
             case GGML_OP_MUL_MAT:
                 ggml_cl_mul_mat(node->src[0], node->src[1], node, nullptr, 0);

ggml-quants.c

@@ -3474,6 +3474,65 @@ void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, in
     }
 }
 
+void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, int k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    float delta[4];
+    uint16_t idx[4];
+
+#if QK_K != 64
+    iq1m_scale_t scale;
+#endif
+
+    for (int i = 0; i < nb; i++) {
+
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+#if QK_K == 64
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+#else
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+        const float d = GGML_FP16_TO_FP32(scale.f16);
+#endif
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+#if QK_K == 64
+            const float dl1 = d * (2*((sc[ib/2] >> (8*(ib%2)+0)) & 0xf) + 1);
+            const float dl2 = d * (2*((sc[ib/2] >> (8*(ib%2)+4)) & 0xf) + 1);
+#else
+            const float dl1 = d * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1);
+            const float dl2 = d * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1);
+#endif
+            idx[0] = qs[0] | ((qh[0] << 8) & 0x700);
+            idx[1] = qs[1] | ((qh[0] << 4) & 0x700);
+            idx[2] = qs[2] | ((qh[1] << 8) & 0x700);
+            idx[3] = qs[3] | ((qh[1] << 4) & 0x700);
+            delta[0] = qh[0] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[1] = qh[0] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[2] = qh[1] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[3] = qh[1] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
+            for (int l = 0; l < 2; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = dl1 * (grid[j] + delta[l]);
+                }
+                y += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = dl2 * (grid[j] + delta[l]);
+                }
+                y += 8;
+            }
+            qs += 4;
+            qh += 2;
+        }
+    }
+}
+
 static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
 
 void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, int k) {
@@ -9695,6 +9754,248 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
 #endif
 }
 
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if QK_K != 64
+    iq1m_scale_t scale;
+#endif
+
+#if defined __ARM_NEON
+
+#if QK_K == 64
+    const int32x4_t mask  = vdupq_n_s32(0xf);
+#else
+    const int32x4_t mask  = vdupq_n_s32(0x7);
+#endif
+    const int32x4_t mone  = vdupq_n_s32(1);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x4_t deltas;
+    deltas.val[0] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(+1));
+    deltas.val[1] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(+1));
+    deltas.val[2] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(-1));
+    deltas.val[3] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(-1));
+
+    ggml_int8x16x4_t q1b;
+    ggml_int8x16x4_t q8b;
+
+    uint32_t aux32;
+    const uint8_t * aux8 = (const uint8_t *)&aux32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+#if QK_K != 64
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+#endif
+
+        int32x4_t sumi1 = mzero;
+        int32x4_t sumi2 = mzero;
+
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+
+            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[0] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[0] << 4) & 0x700)))));
+            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[1] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[1] << 4) & 0x700)))));
+            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[2] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[2] << 4) & 0x700)))));
+            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[3] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[3] << 4) & 0x700)))));
+
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const int32x4_t p1 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[0], q8b.val[0]), ggml_vdotq_s32(mzero, q1b.val[1], q8b.val[1]));
+            const int32x4_t p2 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[2], q8b.val[2]), ggml_vdotq_s32(mzero, q1b.val[3], q8b.val[3]));
+            const int32x4_t p12 = vpaddq_s32(p1, p2);
+
+            const uint32_t * qh32 = (const uint32_t *)qh; // we are 4-byte aligned, so we can do that
+            aux32 = ((qh32[0] >> 3) & 0x01010101) | ((qh32[0] >> 6) & 0x02020202);
+
+            const int32x4_t p3 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[0]], q8b.val[0]), ggml_vdotq_s32(mzero, deltas.val[aux8[1]], q8b.val[1]));
+            const int32x4_t p4 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[2]], q8b.val[2]), ggml_vdotq_s32(mzero, deltas.val[aux8[3]], q8b.val[3]));
+            const int32x4_t p34 = vpaddq_s32(p3, p4);
+
+#if QK_K == 64
+            int32x4_t scales_4 = ggml_vld1q_u32(sc[0] >> 0, sc[0] >> 4, sc[0] >> 8, sc[0] >> 12);
+#else
+            int32x4_t scales_4 = ggml_vld1q_u32(sc[ib/2] >> 0, sc[ib/2] >> 3, sc[ib/2] >> 6, sc[ib/2] >> 9);
+#endif
+            scales_4 = vaddq_s32(vshlq_n_s32(vandq_s32(scales_4, mask), 1), mone);
+
+            sumi1 = vmlaq_s32(sumi1, scales_4, p12);
+            sumi2 = vmlaq_s32(sumi2, scales_4, p34);
+
+            qs += 8; qh += 4;
+
+        }
+
+#if QK_K == 64
+        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
+#else
+        sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
+#endif
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+#if QK_K == 64
+    const __m256i mask = _mm256_set1_epi16(0xf);
+#else
+    const __m256i mask = _mm256_set1_epi16(0x7);
+#endif
+    const __m256i mone = _mm256_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+#if QK_K != 64
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+#endif
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m256i q1b_1 = _mm256_set_epi64x(
+                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)],
+                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]
+            );
+            const __m256i q1b_2 = _mm256_set_epi64x(
+                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)],
+                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]
+            );
+            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+
+            const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m256i delta2 = _mm256_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+            const __m256i dot3 = mul_add_epi8(delta1, q8b_1);
+            const __m256i dot4 = mul_add_epi8(delta2, q8b_2);
+#if QK_K == 64
+            __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[0] >>  4), _mm_set1_epi16(sc[0] >> 0));
+            __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[0] >> 12), _mm_set1_epi16(sc[0] >> 8));
+#else
+            __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0));
+            __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6));
+#endif
+            scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone);
+            scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone);
+            const __m256i p1 = _mm256_madd_epi16(dot1, scale1);
+            const __m256i p2 = _mm256_madd_epi16(dot2, scale2);
+            const __m256i p3 = _mm256_madd_epi16(dot3, scale1);
+            const __m256i p4 = _mm256_madd_epi16(dot4, scale2);
+
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_add_epi32(p1, p2));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_add_epi32(p3, p4));
+
+            qs += 8; qh += 4;
+        }
+
+#if QK_K == 64
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+#else
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+#endif
+        accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
+        accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
+
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#else
+
+    int sum1[2], sum2[2], delta[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+#if QK_K != 64
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+#endif
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            delta[0] = qh[0] & 0x08 ? -1 : 1;
+            delta[1] = qh[0] & 0x80 ? -1 : 1;
+            delta[2] = qh[1] & 0x08 ? -1 : 1;
+            delta[3] = qh[1] & 0x80 ? -1 : 1;
+            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
+                int lsum1 = 0, lsum2 = 0;
+                for (int j = 0; j < 8; ++j) {
+                    lsum1 += q8[j] * grid[j];
+                    lsum2 += q8[j];
+                }
+                q8 += 8;
+                sum1[l/2] += lsum1;
+                sum2[l/2] += lsum2*delta[l];
+            }
+#if QK_K == 64
+            const int ls1 = 2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1;
+            const int ls2 = 2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1;
+#else
+            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
+            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
+#endif
+            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
+            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
+            qs += 4;
+            qh += 2;
+        }
+
+#if QK_K == 64
+        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+#else
+        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+#endif
+    }
+
+    *s = sumf;
+
+#endif
+}
+
 void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
@@ -9938,17 +10239,17 @@ static iq2_entry_t iq2_data[4] = {
 };
 
 static inline int iq2_data_index(enum ggml_type type) {
-    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ2_S);
+    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
     return type == GGML_TYPE_IQ2_XXS ? 0 :
            type == GGML_TYPE_IQ2_XS  ? 1 :
-           type == GGML_TYPE_IQ1_S   ? 2 : 3;
+           type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? 2 : 3;
 }
 
 static inline int iq2_grid_size(enum ggml_type type) {
-    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ2_S);
+    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
     return type == GGML_TYPE_IQ2_XXS ? 256 :
            type == GGML_TYPE_IQ2_XS  ? 512 :
-           type == GGML_TYPE_IQ1_S   ? NGRID_IQ1S : 1024;
+           type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? NGRID_IQ1S : 1024;
 }
 
 static int iq2_compare_func(const void * left, const void * right) {
@@ -10214,10 +10515,10 @@ void iq2xs_init_impl(enum ggml_type type) {
 
     const int kmap_size = 43692;
     //const int nwant = type == GGML_TYPE_IQ1_S ? 3 : 2;
-    const int nwant = type == GGML_TYPE_IQ1_S ? 3 : type == GGML_TYPE_IQ2_S ? 1 : 2;
+    const int nwant = type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? 3 : type == GGML_TYPE_IQ2_S ? 1 : 2;
     const uint16_t * kgrid = type == GGML_TYPE_IQ2_XXS ? kgrid_2bit_256 :
                              type == GGML_TYPE_IQ2_XS  ? kgrid_2bit_512 :
-                             type == GGML_TYPE_IQ1_S   ? kgrid_1bit_2048 : kgrid_2bit_1024;
+                             type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? kgrid_1bit_2048 : kgrid_2bit_1024;
     uint64_t * kgrid_q2xs;
     int      * kmap_q2xs;
     uint16_t * kneighbors_q2xs;
@@ -10314,7 +10615,7 @@ void iq2xs_init_impl(enum ggml_type type) {
 }
 
 void iq2xs_free_impl(enum ggml_type type) {
-    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ2_S);
+    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
     const int gindex = iq2_data_index(type);
     if (iq2_data[gindex].grid) {
         free(iq2_data[gindex].grid);       iq2_data[gindex].grid = NULL;
@@ -11520,7 +11821,16 @@ static int iq1_sort_helper(const void * left, const void * right) {
 }
 
 #define IQ1S_BLOCK_SIZE 32
-static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights) {
+#define IQ1M_BLOCK_SIZE 16
+static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights,
+        float    * scales,
+        float    * weight,
+        float    * sumx,
+        float    * sumw,
+        float    * pairs,
+        int8_t   * L,
+        uint16_t * index,
+        int8_t   * shifts) {
 
     const int gindex = iq2_data_index(GGML_TYPE_IQ1_S);
 
@@ -11534,22 +11844,17 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
     GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
     GGML_ASSERT(n%QK_K == 0);
 
+    block_iq1_s * y = vy;
+
     const int nbl = n/QK_K;
 
-    block_iq1_s * y = vy;
+    const int block_size = IQ1S_BLOCK_SIZE;
 
     const float x_p[3] = {-1 + IQ1S_DELTA,  IQ1S_DELTA, 1 + IQ1S_DELTA};
     const float x_m[3] = {-1 - IQ1S_DELTA, -IQ1S_DELTA, 1 - IQ1S_DELTA};
 
-    float  scales[QK_K/IQ1S_BLOCK_SIZE];
-    float  weight[IQ1S_BLOCK_SIZE];
-    int8_t L[IQ1S_BLOCK_SIZE];
-    float  sumx[IQ1S_BLOCK_SIZE+1];
-    float  sumw[IQ1S_BLOCK_SIZE+1];
-    float  pairs[2*IQ1S_BLOCK_SIZE];
+
     int * idx = (int *)(pairs + 1);
-    uint16_t index[IQ1S_BLOCK_SIZE/8];
-    int8_t shifts[QK_K/IQ1S_BLOCK_SIZE];
 
     for (int ibl = 0; ibl < nbl; ++ibl) {
 
@@ -11564,15 +11869,15 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
         for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
         float sigma2 = 2*sumx2/QK_K;
 
-        for (int ib = 0; ib < QK_K/IQ1S_BLOCK_SIZE; ++ib) {
-            const float * xb = xbl + IQ1S_BLOCK_SIZE*ib;
-            const float * qw = quant_weights + QK_K*ibl + IQ1S_BLOCK_SIZE*ib;
-            for (int i = 0; i < IQ1S_BLOCK_SIZE; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
+            const float * xb = xbl + block_size*ib;
+            const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+            for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
             float max = fabsf(xb[0]);
-            for (int i = 1; i < IQ1S_BLOCK_SIZE; ++i) max = MAX(max, fabsf(xb[i]));
+            for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
             if (!max) {
                 scales[ib] = 0;
-                memset(L, 1, IQ1S_BLOCK_SIZE);
+                memset(L, 1, block_size);
                 continue;
             }
             // Here we solve exactly the sum of squared difference (SSD) weighted minimization problem.
@@ -11581,14 +11886,14 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
             // in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and
             // Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale
             // for each possible and score for each split.
-            for (int j = 0; j < IQ1S_BLOCK_SIZE; ++j) {
+            for (int j = 0; j < block_size; ++j) {
                 pairs[2*j] = xb[j];
                 idx[2*j] = j;
             }
-            qsort(pairs, IQ1S_BLOCK_SIZE, 2*sizeof(float), iq1_sort_helper);
+            qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
             {
                 sumx[0] = sumw[0] = 0;
-                for (int j = 0; j < IQ1S_BLOCK_SIZE; ++j) {
+                for (int j = 0; j < block_size; ++j) {
                     int i = idx[2*j];
                     sumx[j+1] = sumx[j] + weight[i]*xb[i];
                     sumw[j+1] = sumw[j] + weight[i];
@@ -11596,16 +11901,16 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
             }
             float best_score = 0, scale = max;
             int besti1 = -1, besti2 = -1, best_shift = 0;
-            for (int i1 = 0; i1 <= IQ1S_BLOCK_SIZE; ++i1) {
-                for (int i2 = i1; i2 <= IQ1S_BLOCK_SIZE; ++i2) {
-                    float sumqx = (sumx[i1] - sumx[0])*x_p[0] + (sumx[i2] - sumx[i1])*x_p[1] + (sumx[IQ1S_BLOCK_SIZE] - sumx[i2])*x_p[2];
-                    float sumq2 = (sumw[i1] - sumw[0])*x_p[0]*x_p[0] + (sumw[i2] - sumw[i1])*x_p[1]*x_p[1] + (sumw[IQ1S_BLOCK_SIZE] - sumw[i2])*x_p[2]*x_p[2];
+            for (int i1 = 0; i1 <= block_size; ++i1) {
+                for (int i2 = i1; i2 <= block_size; ++i2) {
+                    float sumqx = (sumx[i1] - sumx[0])*x_p[0] + (sumx[i2] - sumx[i1])*x_p[1] + (sumx[block_size] - sumx[i2])*x_p[2];
+                    float sumq2 = (sumw[i1] - sumw[0])*x_p[0]*x_p[0] + (sumw[i2] - sumw[i1])*x_p[1]*x_p[1] + (sumw[block_size] - sumw[i2])*x_p[2]*x_p[2];
                     if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
                         scale = sumqx/sumq2; best_score = scale*sumqx;
                         besti1 = i1; besti2 = i2; best_shift = 1;
                     }
-                    sumqx = (sumx[i1] - sumx[0])*x_m[0] + (sumx[i2] - sumx[i1])*x_m[1] + (sumx[IQ1S_BLOCK_SIZE] - sumx[i2])*x_m[2];
-                    sumq2 = (sumw[i1] - sumw[0])*x_m[0]*x_m[0] + (sumw[i2] - sumw[i1])*x_m[1]*x_m[1] + (sumw[IQ1S_BLOCK_SIZE] - sumw[i2])*x_m[2]*x_m[2];
+                    sumqx = (sumx[i1] - sumx[0])*x_m[0] + (sumx[i2] - sumx[i1])*x_m[1] + (sumx[block_size] - sumx[i2])*x_m[2];
+                    sumq2 = (sumw[i1] - sumw[0])*x_m[0]*x_m[0] + (sumw[i2] - sumw[i1])*x_m[1]*x_m[1] + (sumw[block_size] - sumw[i2])*x_m[2]*x_m[2];
                     if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
                         scale = sumqx/sumq2; best_score = scale*sumqx;
                         besti1 = i1; besti2 = i2; best_shift = -1;
@@ -11615,14 +11920,14 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
             GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_shift != 0);
             for (int j =      0; j < besti1; ++j) L[idx[2*j]] = 0;
             for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
-            for (int j = besti2; j < IQ1S_BLOCK_SIZE; ++j) L[idx[2*j]] = 2;
+            for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2;
             if (scale < 0) {
-                for (int j = 0; j < IQ1S_BLOCK_SIZE; ++j) L[j] = 2 - L[j];
+                for (int j = 0; j < block_size; ++j) L[j] = 2 - L[j];
                 scale = -scale; best_shift = -best_shift;
             }
             bool all_on_grid = true;
             const float * xx = best_shift == 1 ? x_p : x_m;
-            for (int k = 0; k < IQ1S_BLOCK_SIZE/8; ++k) {
+            for (int k = 0; k < block_size/8; ++k) {
                 uint16_t u = 0;
                 for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j);
                 int grid_index = kmap_q2xs[u];
@@ -11636,7 +11941,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
             }
             if (!all_on_grid) {
                 float sumqx = 0, sumq2 = 0;
-                for (int k = 0; k < IQ1S_BLOCK_SIZE/8; ++k) {
+                for (int k = 0; k < block_size/8; ++k) {
                     const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]);
                     for (int j = 0; j < 8; ++j) {
                         float w = weight[8*k + j];
@@ -11648,8 +11953,8 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
                 if (sumqx > 0 && sumq2 > 0) scale = sumqx/sumq2;
             }
             uint16_t h = 0;
-            for (int k = 0; k < IQ1S_BLOCK_SIZE/8; ++k) {
-                y[ibl].qs[(IQ1S_BLOCK_SIZE/8)*ib + k] = index[k] & 255;
+            for (int k = 0; k < block_size/8; ++k) {
+                y[ibl].qs[(block_size/8)*ib + k] = index[k] & 255;
                 h |= (index[k] >> 8) << 3*k;
             }
             y[ibl].qh[ib] = h;
@@ -11660,14 +11965,13 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
         }
 
         if (!max_scale) {
-            memset(y[ibl].qs, 0, QK_K/8);
             continue;
         }
 
         float d = max_scale/15;
-        y[ibl].d = GGML_FP32_TO_FP16(d*1.125f); // 1.085f is another fudge factor. Don't ask me why it is needed.
+        y[ibl].d = GGML_FP32_TO_FP16(d*1.125f); // 1.125f is another fudge factor. Don't ask me why it is needed.
         float id = 1/d;
-        for (int ib = 0; ib < QK_K/IQ1S_BLOCK_SIZE; ++ib) {
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
             int l = nearest_int(0.5f*(id*scales[ib]-1));
             l = MAX(0, MIN(7, l));
             if (shifts[ib] == -1) l |= 8;
@@ -11678,16 +11982,307 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
 
 size_t quantize_iq1_s(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
+    float  scales[QK_K/IQ1S_BLOCK_SIZE];
+    float  weight[IQ1S_BLOCK_SIZE];
+    int8_t L[IQ1S_BLOCK_SIZE];
+    float  sumx[IQ1S_BLOCK_SIZE+1];
+    float  sumw[IQ1S_BLOCK_SIZE+1];
+    float  pairs[2*IQ1S_BLOCK_SIZE];
+    uint16_t index[IQ1S_BLOCK_SIZE/8];
+    int8_t shifts[QK_K/IQ1S_BLOCK_SIZE];
     int nblock = n_per_row/QK_K;
     char * qrow = (char *)dst;
     for (int row = 0; row < nrow; ++row) {
-        quantize_row_iq1_s_impl(src, qrow, n_per_row, quant_weights);
+        quantize_row_iq1_s_impl(src, qrow, n_per_row, quant_weights, scales, weight, sumx, sumw, pairs, L, index, shifts);
         src += n_per_row;
         qrow += nblock*sizeof(block_iq1_s);
     }
     return nrow * nblock * sizeof(block_iq1_s);
 }
 
+static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights,
+        float    * scales,
+        float    * weight,
+        float    * pairs,
+        int8_t   * L,
+        uint16_t * index,
+        int8_t   * shifts) {
+
+    const int gindex = iq2_data_index(GGML_TYPE_IQ1_M);
+
+    const uint64_t * kgrid_q2xs      = iq2_data[gindex].grid;
+    const int      * kmap_q2xs       = iq2_data[gindex].map;
+    const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+    //GGML_ASSERT(quant_weights   && "missing quantization weights");
+    GGML_ASSERT(kgrid_q2xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kmap_q2xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    block_iq1_m * y = vy;
+
+    const int nbl = n/QK_K;
+
+    const int block_size = IQ1M_BLOCK_SIZE;
+
+    const float x_p[3] = {-1 + IQ1M_DELTA,  IQ1M_DELTA, 1 + IQ1M_DELTA};
+    const float x_m[3] = {-1 - IQ1M_DELTA, -IQ1M_DELTA, 1 - IQ1M_DELTA};
+    const uint8_t masks[4] = {0x00, 0x80, 0x08, 0x88};
+
+    int * idx = (int *)(pairs + 1);
+
+    float sumqx[4], sumq2[4];
+
+    iq1m_scale_t s;
+    const float * xx;
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+#if QK_K == 64
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+#endif
+        memset(y[ibl].qs, 0, QK_K/8);
+        memset(y[ibl].qh, 0, QK_K/16);
+        memset(y[ibl].scales, 0, QK_K/32);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = 2*sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
+            const float * xb = xbl + block_size*ib;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+                for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            } else {
+                for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
+            }
+            float max = fabsf(xb[0]);
+            for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
+            if (!max) {
+                scales[ib] = 0;
+                memset(L, 1, block_size);
+                continue;
+            }
+            // Here we solve exactly the sum of squared difference (SSD) weighted minimization problem.
+            // With just 3 allowed quant values (-1, 0, 1), we can search exhaustively for the two
+            // boundaries that split the weights xb[i] into 3 groups. To do so, we sort the weights
+            // in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and
+            // Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale
+            // for each possible and score for each split.
+            for (int j = 0; j < block_size; ++j) {
+                pairs[2*j] = xb[j];
+                idx[2*j] = j;
+            }
+            qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
+            float best_score = 0, scale = max;
+            int besti1 = -1, besti2 = -1, best_k = -1;
+            // 0: +, +
+            // 1: +, -
+            // 2: -, +
+            // 3: -, -
+            for (int i1 = 0; i1 <= block_size; ++i1) {
+                for (int i2 = i1; i2 <= block_size; ++i2) {
+                    memset(sumqx, 0, 4*sizeof(float));
+                    memset(sumq2, 0, 4*sizeof(float));
+                    for (int j = 0; j < i1; ++j) {
+                        int i = idx[2*j];
+                        if (i < block_size/2) {
+                            sumqx[0] += weight[i]*x_p[0]*xb[i];
+                            sumqx[1] += weight[i]*x_p[0]*xb[i];
+                            sumqx[2] += weight[i]*x_m[0]*xb[i];
+                            sumqx[3] += weight[i]*x_m[0]*xb[i];
+                            sumq2[0] += weight[i]*x_p[0]*x_p[0];
+                            sumq2[1] += weight[i]*x_p[0]*x_p[0];
+                            sumq2[2] += weight[i]*x_m[0]*x_m[0];
+                            sumq2[3] += weight[i]*x_m[0]*x_m[0];
+                        } else {
+                            sumqx[0] += weight[i]*x_p[0]*xb[i];
+                            sumqx[2] += weight[i]*x_p[0]*xb[i];
+                            sumqx[1] += weight[i]*x_m[0]*xb[i];
+                            sumqx[3] += weight[i]*x_m[0]*xb[i];
+                            sumq2[0] += weight[i]*x_p[0]*x_p[0];
+                            sumq2[2] += weight[i]*x_p[0]*x_p[0];
+                            sumq2[1] += weight[i]*x_m[0]*x_m[0];
+                            sumq2[3] += weight[i]*x_m[0]*x_m[0];
+                        }
+                    }
+                    for (int j = i1; j < i2; ++j) {
+                        int i = idx[2*j];
+                        if (i < block_size/2) {
+                            sumqx[0] += weight[i]*x_p[1]*xb[i];
+                            sumqx[1] += weight[i]*x_p[1]*xb[i];
+                            sumqx[2] += weight[i]*x_m[1]*xb[i];
+                            sumqx[3] += weight[i]*x_m[1]*xb[i];
+                            sumq2[0] += weight[i]*x_p[1]*x_p[1];
+                            sumq2[1] += weight[i]*x_p[1]*x_p[1];
+                            sumq2[2] += weight[i]*x_m[1]*x_m[1];
+                            sumq2[3] += weight[i]*x_m[1]*x_m[1];
+                        } else {
+                            sumqx[0] += weight[i]*x_p[1]*xb[i];
+                            sumqx[2] += weight[i]*x_p[1]*xb[i];
+                            sumqx[1] += weight[i]*x_m[1]*xb[i];
+                            sumqx[3] += weight[i]*x_m[1]*xb[i];
+                            sumq2[0] += weight[i]*x_p[1]*x_p[1];
+                            sumq2[2] += weight[i]*x_p[1]*x_p[1];
+                            sumq2[1] += weight[i]*x_m[1]*x_m[1];
+                            sumq2[3] += weight[i]*x_m[1]*x_m[1];
+                        }
+                    }
+                    for (int j = i2; j < block_size; ++j) {
+                        int i = idx[2*j];
+                        if (i < block_size/2) {
+                            sumqx[0] += weight[i]*x_p[2]*xb[i];
+                            sumqx[1] += weight[i]*x_p[2]*xb[i];
+                            sumqx[2] += weight[i]*x_m[2]*xb[i];
+                            sumqx[3] += weight[i]*x_m[2]*xb[i];
+                            sumq2[0] += weight[i]*x_p[2]*x_p[2];
+                            sumq2[1] += weight[i]*x_p[2]*x_p[2];
+                            sumq2[2] += weight[i]*x_m[2]*x_m[2];
+                            sumq2[3] += weight[i]*x_m[2]*x_m[2];
+                        } else {
+                            sumqx[0] += weight[i]*x_p[2]*xb[i];
+                            sumqx[2] += weight[i]*x_p[2]*xb[i];
+                            sumqx[1] += weight[i]*x_m[2]*xb[i];
+                            sumqx[3] += weight[i]*x_m[2]*xb[i];
+                            sumq2[0] += weight[i]*x_p[2]*x_p[2];
+                            sumq2[2] += weight[i]*x_p[2]*x_p[2];
+                            sumq2[1] += weight[i]*x_m[2]*x_m[2];
+                            sumq2[3] += weight[i]*x_m[2]*x_m[2];
+                        }
+                    }
+                    for (int k = 0; k < 4; ++k) {
+                        if (sumq2[k] > 0 && sumqx[k]*sumqx[k] > best_score*sumq2[k]) {
+                            scale = sumqx[k]/sumq2[k]; best_score = scale*sumqx[k];
+                            besti1 = i1; besti2 = i2; best_k = k;
+                        }
+                    }
+                }
+            }
+            GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_k >= 0);
+            for (int j =      0; j < besti1; ++j) L[idx[2*j]] = 0;
+            for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
+            for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2;
+            if (scale < 0) {
+                for (int j = 0; j < block_size; ++j) L[j] = 2 - L[j];
+                scale = -scale;
+                best_k = best_k == 0 ? 3 : best_k == 1 ? 2 : best_k == 2 ? 1 : 0;
+            }
+            bool all_on_grid = true;
+            for (int k = 0; k < block_size/8; ++k) {
+                if (k == 0) xx = best_k < 2 ? x_p : x_m;
+                else xx = best_k%2 == 0 ? x_p : x_m;
+                uint16_t u = 0;
+                for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j);
+                int grid_index = kmap_q2xs[u];
+                if (grid_index < 0) {
+                    all_on_grid = false;
+                    const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                    grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
+                    GGML_ASSERT(grid_index >= 0);
+                }
+                index[k] = grid_index;
+            }
+            if (!all_on_grid) {
+                float sumqx_f = 0, sumq2_f = 0;
+                for (int k = 0; k < block_size/8; ++k) {
+                    if (k == 0) xx = best_k < 2 ? x_p : x_m;
+                    else xx = best_k%2 == 0 ? x_p : x_m;
+                    const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]);
+                    for (int j = 0; j < 8; ++j) {
+                        float w = weight[8*k + j];
+                        float q = xx[(pg[j] - 1)/2];
+                        sumqx_f += w*q*xb[8*k+j];
+                        sumq2_f += w*q*q;
+                    }
+                }
+                if (sumqx_f > 0 && sumq2_f > 0) scale = sumqx_f/sumq2_f;
+            }
+            y[ibl].qs[2*ib + 0] = index[0] & 255;
+            y[ibl].qs[2*ib + 1] = index[1] & 255;
+            y[ibl].qh[ib] = (index[0] >> 8) | ((index[1] >> 8) << 4);
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            shifts[ib] = best_k;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            continue;
+        }
+
+        uint16_t * sc = (uint16_t *)y[ibl].scales;
+#if QK_K == 64
+        float d = max_scale/31;
+#else
+        float d = max_scale/15;
+#endif
+        float id = 1/d;
+        float sumqx_f = 0, sumq2_f = 0;
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib+0]-1));
+#if QK_K == 64
+            l = MAX(0, MIN(15, l));
+            sc[ib/4] |= (l << 4*(ib%4));
+#else
+            l = MAX(0, MIN(7, l));
+            sc[ib/4] |= (l << 3*(ib%4));
+#endif
+            y[ibl].qh[ib] |= masks[shifts[ib]];
+            const float * xb = xbl + block_size*ib;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+                for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            } else {
+                for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
+            }
+            for (int k = 0; k < block_size/8; ++k) {
+                if (k == 0) xx = shifts[ib] < 2 ? x_p : x_m;
+                else xx = shifts[ib]%2 == 0 ? x_p : x_m;
+                const int8_t * pg = (const int8_t *)(kgrid_q2xs + y[ibl].qs[2*ib+k] + ((y[ibl].qh[ib] << (8 - 4*k)) & 0x700));
+                for (int j = 0; j < 8; ++j) {
+                    float w = weight[8*k + j];
+                    float q = xx[(pg[j] - 1)/2]*(2*l+1);
+                    sumqx_f += w*q*xb[8*k+j];
+                    sumq2_f += w*q*q;
+                }
+            }
+        }
+        if (sumq2_f > 0) d = sumqx_f/sumq2_f;
+        s.f16 = GGML_FP32_TO_FP16(d*1.1125f); // 1.1125f is another fudge factor. Don't ask me why it is needed.
+#if QK_K == 64
+        y[ibl].d = s.f16;
+#else
+        sc[0] |= ((s.u16 & 0x000f) << 12);
+        sc[1] |= ((s.u16 & 0x00f0) <<  8);
+        sc[2] |= ((s.u16 & 0x0f00) <<  4);
+        sc[3] |= ((s.u16 & 0xf000) <<  0);
+#endif
+    }
+}
+
+size_t quantize_iq1_m(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    float  scales[QK_K/IQ1M_BLOCK_SIZE];
+    float  weight[IQ1M_BLOCK_SIZE];
+    int8_t L[IQ1M_BLOCK_SIZE];
+    float  pairs[2*IQ1M_BLOCK_SIZE];
+    uint16_t index[IQ1M_BLOCK_SIZE/8];
+    int8_t shifts[QK_K/IQ1M_BLOCK_SIZE];
+    int nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int row = 0; row < nrow; ++row) {
+        quantize_row_iq1_m_impl(src, qrow, n_per_row, quant_weights, scales, weight, pairs, L, index, shifts);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq1_m);
+    }
+    return nrow * nblock * sizeof(block_iq1_m);
+}
+
 // ============================ 4-bit non-linear quants
 
 static inline int best_index_int8(int n, const int8_t * val, float x) {

ggml-quants.h

@@ -72,6 +72,7 @@ void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_
 void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
@@ -94,6 +95,7 @@ void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const
 void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -104,6 +106,7 @@ size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT ds
 size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
 size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
 size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
 size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
 size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
 size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);

ggml-sycl.cpp

@@ -1664,24 +1664,6 @@ namespace dpct
                               const void *alpha, const void *a, int lda, const void *b,
                               int ldb, const void *beta, void *c, int ldc)
         {
-#ifndef __INTEL_MKL__
-            GGML_UNUSED(q);
-            GGML_UNUSED(a_trans);
-            GGML_UNUSED(b_trans);
-            GGML_UNUSED(m);
-            GGML_UNUSED(n);
-            GGML_UNUSED(k);
-            GGML_UNUSED(alpha);
-            GGML_UNUSED(a);
-            GGML_UNUSED(lda);
-            GGML_UNUSED(b);
-            GGML_UNUSED(ldb);
-            GGML_UNUSED(beta);
-            GGML_UNUSED(c);
-            GGML_UNUSED(ldc);
-            throw std::runtime_error("The oneAPI Math Kernel Library (oneMKL) Interfaces "
-                                     "Project does not support this API.");
-#else
             Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
             Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
             auto data_a = get_memory<const Ta>(a);
@@ -1690,7 +1672,6 @@ namespace dpct
             oneapi::mkl::blas::column_major::gemm(
                 q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
                 data_b, ldb, beta_value, data_c, ldc);
-#endif
         }
 
         template <typename VecT, class BinaryOperation, class = void>
@@ -2330,6 +2311,7 @@ namespace dpct
                                           lda, b, ldb, beta, c, ldc);
             break;
         }
+#ifdef __INTEL_MKL__
         case detail::get_type_combination_id(
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_float, library_data_t::real_float):
@@ -2391,6 +2373,7 @@ namespace dpct
                 q, a_trans, b_trans, m, n, k, &alpha_float, a, lda, b, ldb, &beta_float, c, ldc);
             break;
         }
+#endif // __INTEL_MKL__
         default:
             throw std::runtime_error("the combination of data type is unsupported");
         }
@@ -2968,7 +2951,7 @@ namespace dpct
 #include "ggml-common.h"
 
 static int g_ggml_sycl_debug=0;
-#define GGML_SYCL_DEBUG(...) do{if(g_ggml_sycl_debug) printf(__VA_ARGS__);}while(0)
+#define GGML_SYCL_DEBUG(...) do{if(g_ggml_sycl_debug) fprintf(stderr, __VA_ARGS__);}while(0)
 
 #define CHECK_TRY_ERROR(expr)                                                  \
   [&]() {                                                                      \
@@ -8079,7 +8062,7 @@ template <bool need_check> static void
 template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_sycl_t vec_dot_q_sycl>
 static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows,
                           const sycl::nd_item<3> &item_ct1,
-                          const uint32_t *iq3xxs_grid_ptr, const uint64_t *ksigns64_ptr) {
+                          const uint32_t *iq3xxs_grid_ptr=nullptr, const uint64_t *ksigns64_ptr=nullptr) {
     const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
                     item_ct1.get_local_id(1);
 
@@ -9956,17 +9939,14 @@ static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int k,
                                         dpct::queue_ptr stream) {
     const int nb = k / QK_K;
     {
-        iq2xxs_grid.init(*stream);
-        ksigns_iq2xs.init(*stream);
-        kmask_iq2xs.init(*stream);
 
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq2xxs_grid_ptr_ct1 = iq2xxs_grid.get_ptr();
-            auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr();
-            auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr();
+            auto iq2xxs_grid_ptr_ct1 = &iq2xxs_grid[0];
+            auto ksigns_iq2xs_ptr_ct1 = &ksigns_iq2xs[0];
+            auto kmask_iq2xs_ptr_ct1 = &kmask_iq2xs[0];
 
             cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
                                                    sycl::range<3>(1, 1, 32),
@@ -9985,17 +9965,14 @@ static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int k,
                                        dpct::queue_ptr stream) {
     const int nb = k / QK_K;
     {
-        iq2xs_grid.init(*stream);
-        ksigns_iq2xs.init(*stream);
-        kmask_iq2xs.init(*stream);
 
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq2xs_grid_ptr_ct1 = iq2xs_grid.get_ptr();
-            auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr();
-            auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr();
+            auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0];
+            auto ksigns_iq2xs_ptr_ct1 = &ksigns_iq2xs[0];
+            auto kmask_iq2xs_ptr_ct1 = &kmask_iq2xs[0];
 
             cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
                                                    sycl::range<3>(1, 1, 32),
@@ -10014,17 +9991,14 @@ static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int k,
                                         dpct::queue_ptr stream) {
     const int nb = k / QK_K;
     {
-        iq3xxs_grid.init(*stream);
-        ksigns_iq2xs.init(*stream);
-        kmask_iq2xs.init(*stream);
 
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr();
-            auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr();
+            auto iq3xxs_grid_ptr_ct1 = &iq3xxs_grid[0];
+            auto ksigns_iq2xs_ptr_ct1 = &ksigns_iq2xs[0];
+            auto kmask_iq2xs_ptr_ct1 = &kmask_iq2xs[0];
 
             cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
                                                    sycl::range<3>(1, 1, 32),
@@ -10043,17 +10017,14 @@ static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int k,
                                         dpct::queue_ptr stream) {
     const int nb = k / QK_K;
     {
-        iq3s_grid.init(*stream);
-        ksigns_iq2xs.init(*stream);
-        kmask_iq2xs.init(*stream);
 
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3s_grid_ptr_ct1 = iq3s_grid.get_ptr();
-            auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr();
-            auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr();
+            auto iq3s_grid_ptr_ct1 = &iq3s_grid[0];
+            auto ksigns_iq2xs_ptr_ct1 = &ksigns_iq2xs[0];
+            auto kmask_iq2xs_ptr_ct1 = &kmask_iq2xs[0];
 
             cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
                                                    sycl::range<3>(1, 1, 32),
@@ -10072,17 +10043,14 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int k,
                                         dpct::queue_ptr stream) {
     const int nb = k / QK_K;
     {
-        iq1s_grid_gpu.init(*stream);
-        ksigns_iq2xs.init(*stream);
-        kmask_iq2xs.init(*stream);
 
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq1s_grid_ptr_ct1 = iq1s_grid_gpu.get_ptr();
-            auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr();
-            auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr();
+            auto iq1s_grid_ptr_ct1 = &iq1s_grid_gpu[0];
+            auto ksigns_iq2xs_ptr_ct1 = &ksigns_iq2xs[0];
+            auto kmask_iq2xs_ptr_ct1 = &kmask_iq2xs[0];
 
             cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
                                                    sycl::range<3>(1, 1, 32),
@@ -10415,12 +10383,8 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10428,8 +10392,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK4_0, QI4_0, block_q4_0,
                                       VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10444,12 +10407,8 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10457,8 +10416,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK4_0, QI4_1, block_q4_1,
                                       VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10473,12 +10431,8 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10486,8 +10440,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK5_0, QI5_0, block_q5_0,
                                       VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10502,12 +10455,8 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10515,8 +10464,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK5_1, QI5_1, block_q5_1,
                                       VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10531,12 +10479,8 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10544,8 +10488,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK8_0, QI8_0, block_q8_0,
                                       VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10560,12 +10503,8 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10573,8 +10512,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK_K, QI2_K, block_q2_K,
                                       VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10589,12 +10527,8 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10602,8 +10536,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK_K, QI3_K, block_q3_K,
                                       VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10618,12 +10551,8 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10631,8 +10560,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK_K, QI4_K, block_q4_K,
                                       VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10647,12 +10575,8 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10660,8 +10584,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK_K, QI5_K, block_q5_K,
                                       VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
@@ -10676,12 +10599,8 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10689,13 +10608,13 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
                     [[intel::reqd_sub_group_size(32)]] {
                         mul_mat_vec_q<QK_K, QI6_K, block_q6_K,
                                       VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1,
-                            iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1);
+                            vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
     }
 }
 
+
 static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy,
                                           float *dst, const int ncols,
                                           const int nrows,
@@ -10705,15 +10624,11 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq2xxs_grid.init(*stream);
-        ksigns_iq2xs.init(*stream);
-        kmask_iq2xs.init(*stream);
-
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq2xxs_grid_ptr_ct1 = iq2xxs_grid.get_ptr();
-            auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr();
-            auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr();
+            auto iq2xxs_grid_ptr_ct1 = &iq2xxs_grid[0];
+            auto ksigns_iq2xs_ptr_ct1 = &ksigns_iq2xs[0];
+            auto kmask_iq2xs_ptr_ct1 = &kmask_iq2xs[0];
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10736,12 +10651,10 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq2xs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq2xs_grid_ptr_ct1 = iq2xs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
+            auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0];
+            auto ksigns64_ptr_ct1 = &ksigns64[0];
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10764,12 +10677,10 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3xxs_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
+            auto iq3xxs_grid_ptr_ct1 = &iq3xxs_grid[0];
+            auto ksigns64_ptr_ct1 = &ksigns64[0];
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10792,12 +10703,10 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq3s_grid.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3s_grid_ptr_ct1 = iq3s_grid.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
+            auto iq3s_grid_ptr_ct1 = &iq3s_grid[0];
+            auto ksigns64_ptr_ct1 = &ksigns64[0];
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -10820,12 +10729,10 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq1s_grid_gpu.init(*stream);
-        ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq1s_grid_ptr_ct1 = iq1s_grid_gpu.get_ptr();
-            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
+            auto iq1s_grid_ptr_ct1 = &iq1s_grid_gpu[0];
+            auto ksigns64_ptr_ct1 = &ksigns64[0];
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
@@ -12868,6 +12775,7 @@ void print_device_detail(int id, sycl::device &device, std::string device_type)
 }
 
 void ggml_backend_sycl_print_sycl_devices() {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_print_sycl_devices\n");
     int device_count = dpct::dev_mgr::instance().device_count();
     std::map<std::string, size_t> DeviceNums;
     fprintf(stderr, "found %d SYCL devices:\n", device_count);
@@ -12925,7 +12833,9 @@ static void ggml_init_sycl() try {
     static bool initialized = false;
 
     if (!initialized) {
+        fprintf(stderr, "[SYCL] call ggml_init_sycl\n");
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
+
         fprintf(stderr, "%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug);
 
 #if defined(GGML_SYCL_F16)
@@ -14986,6 +14896,9 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
     SYCL_CHECK(ggml_sycl_set_device(g_main_device));
     dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0];
 
+    bool no_mixed_dtypes = main_stream->get_backend() == sycl::backend::ext_oneapi_cuda ||
+                           main_stream->get_backend() == sycl::backend::ext_oneapi_hip;
+
     SYCL_CHECK(
         CHECK_TRY_ERROR(g_sycl_handles[g_main_device] = main_stream));
 
@@ -15016,24 +14929,38 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
 
     dpct::library_data_t cu_compute_type = dpct::library_data_t::real_float;
     dpct::library_data_t cu_data_type = dpct::library_data_t::real_float;
+    if (no_mixed_dtypes) {
+        cu_compute_type = dpct::library_data_t::real_half;
+        cu_data_type = dpct::library_data_t::real_half;
+    }
 
     // dst strides
     size_t nbd2 = dst->nb[2];
     size_t nbd3 = dst->nb[3];
 
+    const float alpha_f32 = 1.0f;
+    const float beta_f32 = 0.0f;
+
     const sycl::half alpha_f16 = 1.0f;
     const sycl::half beta_f16 = 0.0f;
 
-    const float alpha_f32 = 1.0f;
-    const float beta_f32  = 0.0f;
-
     const void * alpha = &alpha_f32;
     const void * beta  = &beta_f32;
+    if (no_mixed_dtypes) {
+        alpha = &alpha_f16;
+        beta  = &beta_f16;
+    }
 
     // TODO: Renable (dst->op_params[0] =! GGML_PREC_DEFAULT) pathway
-    // oneMKL open source supports half, half, float, float: datatypes
+    // when oneMKL open source supports half, half, float, float: datatypes
 
     dst_t = (char *) dst_ddf;
+    if (no_mixed_dtypes) {
+        dst_t = (char *) dst_f16.alloc(ne_dst);
+
+        nbd2 /= sizeof(float) / sizeof(sycl::half);
+        nbd3 /= sizeof(float) / sizeof(sycl::half);
+    }
 
     GGML_ASSERT(ne12 % ne02 == 0);
     GGML_ASSERT(ne13 % ne03 == 0);
@@ -15119,6 +15046,10 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
     }
 #endif
 
+    if (no_mixed_dtypes) {
+        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
+        to_fp32_sycl(dst_f16.get(), dst_ddf, ne_dst, main_stream);
+    }
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -16018,6 +15949,7 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_
 }
 
 GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len) try {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_sycl_get_gpu_list\n");
     for(int i=0;i<max_len;i++) id_list[i] = -1;
 
     if (!g_sycl_gpu_mgr) {
@@ -16052,6 +15984,7 @@ catch (sycl::exception const &exc) {
 
 GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description,
                                       size_t description_size) try {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_sycl_get_device_description\n");
     dpct::device_info prop;
     int device_id = g_sycl_gpu_mgr->gpus[device];
     SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
@@ -16066,6 +15999,7 @@ catch (sycl::exception const &exc) {
 
 GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free,
                                                    size_t *total) try {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_memory\n");
     ggml_sycl_set_device(device);
 
     /*
@@ -16417,7 +16351,8 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = {
 };
 
 ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device_index) {
-    ggml_init_sycl();
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_buffer_type\n");
+
     if (device_index>=g_device_count or device_index<0) {
         printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
             device_index, g_device_count-1);
@@ -16787,6 +16722,7 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface
 };
 
 GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_split_buffer_type\n");
     ggml_init_sycl();
     // FIXME: this is not thread safe
     static std::map<std::array<float, GGML_SYCL_MAX_DEVICES>, struct ggml_backend_buffer_type> buft_map;
@@ -16859,6 +16795,7 @@ static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggm
 }
 
 ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_host_buffer_type\n");
     static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_type_host = {
         /* .iface    = */ {
             /* .get_name         = */ ggml_backend_sycl_host_buffer_type_name,
@@ -16973,7 +16910,7 @@ GGML_CALL static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t back
     params.ith = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
-        if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
             continue;
         }
 #ifndef NDEBUG
@@ -17155,6 +17092,7 @@ static ggml_guid_t ggml_backend_sycl_guid() {
 }
 
 GGML_CALL ggml_backend_t ggml_backend_sycl_init(int device) {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_init\n");
     ggml_init_sycl();
 
     check_allow_gpu_index(device);
@@ -17181,6 +17119,7 @@ bool ggml_backend_is_sycl(ggml_backend_t backend) {
 }
 
 GGML_CALL int ggml_backend_sycl_get_device_count() {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_count\n");
     if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr();
     return g_sycl_gpu_mgr->get_gpu_count();
 }
@@ -17193,16 +17132,21 @@ GGML_CALL static ggml_backend_t ggml_backend_reg_sycl_init(const char * params,
 }
 
 GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id) {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_index\n");
     return g_sycl_gpu_mgr->get_index(device_id);
 }
 
 GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index) {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_id\n");
     return g_sycl_gpu_mgr->gpus[device_index];
 }
 
 GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id) {
-    GGML_ASSERT(main_gpu_id<g_all_sycl_device_count);
+    ggml_init_sycl();
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_set_single_device_mode\n");
     fprintf(stderr, "ggml_backend_sycl_set_single_device: use single device: [%d]\n", main_gpu_id);
+    GGML_ASSERT(main_gpu_id<g_all_sycl_device_count);
+
     if (g_sycl_gpu_mgr) {
         delete g_sycl_gpu_mgr;
     }
@@ -17213,6 +17157,9 @@ GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id
 }
 
 GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode() {
+    ggml_init_sycl();
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_set_mul_device_mode\n");
+
     if (g_ggml_sycl_backend_gpu_mode == SYCL_MUL_GPU_MODE) {
         return;
     }

ggml-vulkan.h

@@ -11,17 +11,6 @@ extern "C" {
 #define GGML_VK_MAX_DEVICES 16
 
 GGML_API void ggml_vk_instance_init(void);
-GGML_API void ggml_vk_init_cpu_assist(void);
-
-GGML_API void ggml_vk_preallocate_buffers_graph_cpu_assist(struct ggml_tensor * node);
-GGML_API void ggml_vk_preallocate_buffers_cpu_assist(void);
-GGML_API void ggml_vk_build_graph_cpu_assist(struct ggml_tensor * node, bool last_node);
-GGML_API bool ggml_vk_compute_forward_cpu_assist(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-#ifdef GGML_VULKAN_CHECK_RESULTS
-void ggml_vk_check_results_1_cpu_assist(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-#endif
-GGML_API void ggml_vk_graph_cleanup_cpu_assist(void);
-GGML_API void ggml_vk_free_cpu_assist(void);
 
 // backend API
 GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);

ggml.c

@@ -278,8 +278,6 @@ inline static void * ggml_calloc(size_t num, size_t size) {
 #include <Accelerate/Accelerate.h>
 #if defined(GGML_USE_CLBLAST) // allow usage of CLBlast alongside Accelerate functions
 #include "ggml-opencl.h"
-#elif defined(GGML_USE_VULKAN)
-#include "ggml-vulkan.h"
 #endif
 #elif defined(GGML_USE_OPENBLAS)
 #if defined(GGML_BLAS_USE_MKL)
@@ -289,8 +287,6 @@ inline static void * ggml_calloc(size_t num, size_t size) {
 #endif
 #elif defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
-#elif defined(GGML_USE_VULKAN)
-#include "ggml-vulkan.h"
 #endif
 
 // floating point type used to accumulate sums
@@ -794,6 +790,18 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
+    [GGML_TYPE_IQ1_M] = {
+        .type_name                = "iq1_m",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq1_m),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq1_m,
+        .from_float               = NULL,
+        .from_float_reference     = NULL,
+        .vec_dot                  = ggml_vec_dot_iq1_m_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_IQ4_NL] = {
         .type_name                = "iq4_nl",
         .blck_size                = QK4_NL,
@@ -2539,6 +2547,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ3_XXS:       wtype = GGML_TYPE_IQ3_XXS;  break;
         case GGML_FTYPE_MOSTLY_IQ1_S:         wtype = GGML_TYPE_IQ1_S;    break;
+        case GGML_FTYPE_MOSTLY_IQ1_M:         wtype = GGML_TYPE_IQ1_M;    break;
         case GGML_FTYPE_MOSTLY_IQ4_NL:        wtype = GGML_TYPE_IQ4_NL;   break;
         case GGML_FTYPE_MOSTLY_IQ4_XS:        wtype = GGML_TYPE_IQ4_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ3_S:         wtype = GGML_TYPE_IQ3_S;    break;
@@ -2594,6 +2603,16 @@ static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
+GGML_CALL bool ggml_is_empty(const struct ggml_tensor * tensor) {
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        if (tensor->ne[i] == 0) {
+            // empty if any dimension has no elements
+            return true;
+        }
+    }
+    return false;
+}
+
 bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -2608,7 +2627,7 @@ bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor
 static inline bool ggml_can_repeat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
-    return
+    return ggml_is_empty(t0) ? ggml_is_empty(t1) :
         (t1->ne[0]%t0->ne[0] == 0) &&
         (t1->ne[1]%t0->ne[1] == 0) &&
         (t1->ne[2]%t0->ne[2] == 0) &&
@@ -2694,8 +2713,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 
 #if defined(GGML_USE_CLBLAST)
         ggml_cl_init();
-#elif defined(GGML_USE_VULKAN)
-        ggml_vk_init_cpu_assist();
 #endif
 
         ggml_setup_op_has_task_pass();
@@ -2915,7 +2932,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         data_size *= ne[i];
     }
 
-    GGML_ASSERT(view_src == NULL || data_size + view_offs <= ggml_nbytes(view_src));
+    GGML_ASSERT(view_src == NULL || data_size == 0 || data_size + view_offs <= ggml_nbytes(view_src));
 
     void * data = view_src != NULL ? view_src->data : NULL;
     if (data != NULL) {
@@ -4556,45 +4573,38 @@ void ggml_mul_mat_set_prec(
 
 // ggml_mul_mat_id
 
+// NOTE: id will be removed in the future and instead all the experts listed in ids will be computed
+//       this will allow computing all the used experts in a single matrix multiplication
 struct ggml_tensor * ggml_mul_mat_id(
         struct ggml_context * ctx,
-        struct ggml_tensor  * const as[],
-        int                   n_as,
+        struct ggml_tensor  * as,
         struct ggml_tensor  * ids,
         int                   id,
         struct ggml_tensor  * b) {
 
     GGML_ASSERT(ids->type == GGML_TYPE_I32);
-    GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1);
-    GGML_ASSERT(ids->ne[1] == b->ne[1]);
+    GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1); // ids is 2d
+    GGML_ASSERT(ids->ne[1] == b->ne[1]); // must have an expert per b row
     GGML_ASSERT(ids->ne[2] == b->ne[2] && ids->ne[3] == b->ne[3]);
-    GGML_ASSERT(n_as > 0 && n_as <= GGML_MAX_SRC - 2);
-    GGML_ASSERT(id >= 0 && id < ids->ne[0]);
+    GGML_ASSERT(id >= 0 && id < ids->ne[0]); // valid id
+    GGML_ASSERT(as->ne[0] == b->ne[0]); // can_mul_mat
 
     bool is_node = false;
 
-    if (as[0]->grad || b->grad) {
+    if (as->grad || b->grad) {
         is_node = true;
     }
 
-    const int64_t ne[4] = { as[0]->ne[1], b->ne[1], b->ne[2], b->ne[3] };
+    const int64_t ne[4] = { as->ne[1], b->ne[1], b->ne[2], b->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     ggml_set_op_params_i32(result, 0, id);
-    ggml_set_op_params_i32(result, 1, n_as);
 
     result->op   = GGML_OP_MUL_MAT_ID;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = ids;
+    result->src[0] = as;
     result->src[1] = b;
-
-    for (int i = 0; i < n_as; i++) {
-        struct ggml_tensor * a = as[i];
-        GGML_ASSERT(ggml_are_same_shape(as[0], a));
-        GGML_ASSERT(ggml_can_mul_mat(a, b));
-        GGML_ASSERT(!ggml_is_transposed(a));
-        result->src[i + 2] = a;
-    }
+    result->src[2] = ids;
 
     return result;
 }
@@ -8135,6 +8145,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -8417,6 +8428,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -8544,6 +8556,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -10928,10 +10941,9 @@ static void ggml_compute_forward_mul_mat_id(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * ids = dst->src[0];
+    const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
-
-    const struct ggml_tensor * src0 = dst->src[2]; // only for GGML_TENSOR_BINARY_OP_LOCALS
+    const struct ggml_tensor * ids = dst->src[2];
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -10961,13 +10973,13 @@ static void ggml_compute_forward_mul_mat_id(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
+    // broadcast is not supported with mmid
+    assert(ne12 == 1);
+    assert(ne13 == 1);
 
     // row groups
     const int id   = ggml_get_op_params_i32(dst, 0);
-    const int n_as = ggml_get_op_params_i32(dst, 1);
+    const int n_as = src0->ne[2];
 
     char * wdata_src1_end = (src1->type == vec_dot_type) ?
             (char *) params->wdata :
@@ -11027,7 +11039,7 @@ static void ggml_compute_forward_mul_mat_id(
             continue;
         }
 
-        const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
+        size_t src0_offset = cur_a*src0->nb[2];
 
         const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
         const size_t row_size = ggml_row_size(vec_dot_type, ne10);
@@ -11062,9 +11074,6 @@ static void ggml_compute_forward_mul_mat_id(
             continue;
         }
 
-        assert(ne12 % ne02 == 0);
-        assert(ne13 % ne03 == 0);
-
         // block-tiling attempt
         const int64_t blck_0 = 16;
         const int64_t blck_1 = 16;
@@ -11081,14 +11090,14 @@ static void ggml_compute_forward_mul_mat_id(
                     const int64_t  i11 = MMID_MATRIX_ROW(cur_a, _i11);
 
                     // broadcast src0 into src1
-                    const int64_t i03 = i13/r3;
-                    const int64_t i02 = i12/r2;
+                    //const int64_t i03 = i13/r3;
+                    //const int64_t i02 = i12/r2;
 
                     const int64_t i1 = i11;
                     const int64_t i2 = i12;
                     const int64_t i3 = i13;
 
-                    const char * src0_row = (const char *) src0_cur->data + (0 + i02*nb02 + i03*nb03);
+                    const char * src0_row = (const char *) src0->data + src0_offset;
 
                     // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
                     //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
@@ -11447,6 +11456,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -11638,6 +11648,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -11861,6 +11872,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -12564,6 +12576,7 @@ static void ggml_compute_forward_alibi(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -12652,6 +12665,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -16093,24 +16107,10 @@ static void ggml_compute_forward_cross_entropy_loss_back(
 static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
     GGML_ASSERT(params);
 
-    if (tensor->op == GGML_OP_NONE) {
+    if (tensor->op == GGML_OP_NONE || ggml_is_empty(tensor)) {
         return;
     }
 
-#if defined(GGML_USE_VULKAN)
-    const bool skip_cpu = ggml_vk_compute_forward_cpu_assist(params, tensor);
-#ifdef GGML_VULKAN_CHECK_RESULTS
-    if (skip_cpu) {
-        ggml_vk_check_results_1_cpu_assist(params, tensor);
-    }
-#endif
-    if (skip_cpu) {
-        return;
-    }
-    GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_TYPE_CPU);
-    GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU);
-#endif // GGML_USE_VULKAN
-
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
@@ -17962,6 +17962,12 @@ static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const
 static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_threads) {
     int n_tasks = 0;
 
+    if (ggml_is_empty(node)) {
+        // no need to multi-thread a no-op
+        n_tasks = 1;
+        return n_tasks;
+    }
+
     switch (node->op) {
         case GGML_OP_CPY:
         case GGML_OP_DUP:
@@ -18447,13 +18453,13 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
             case GGML_OP_MUL_MAT_ID:
                 {
                     cur = 0;
-                    const struct ggml_tensor * src0 = node->src[2];
+                    const struct ggml_tensor * src0 = node->src[0];
                     const struct ggml_tensor * src1 = node->src[1];
                     const enum ggml_type vec_dot_type = type_traits[src0->type].vec_dot_type;
                     if (src1->type != vec_dot_type) {
                         cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
                     }
-                    const int n_as = ggml_get_op_params_i32(node, 1);
+                    const int n_as = src0->ne[2];
                     cur += GGML_PAD(cur, sizeof(int64_t));       // align
                     cur += n_as * sizeof(int64_t);               // matrix_row_counts
                     cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows
@@ -18580,17 +18586,6 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         }
     }
 
-#ifdef GGML_USE_VULKAN
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_vk_preallocate_buffers_graph_cpu_assist(cgraph->nodes[i]);
-    }
-    ggml_vk_preallocate_buffers_cpu_assist();
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_vk_build_graph_cpu_assist(cgraph->nodes[i], i == cgraph->n_nodes - 1);
-    }
-#endif
-
     const int n_threads = cplan->n_threads;
 
     struct ggml_compute_state_shared state_shared = {
@@ -18647,10 +18642,6 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         }
     }
 
-#ifdef GGML_USE_VULKAN
-    ggml_vk_graph_cleanup_cpu_assist();
-#endif
-
     // performance stats (graph)
     {
         int64_t perf_cycles_cur  = ggml_perf_cycles()  - perf_start_cycles;
@@ -20306,7 +20297,8 @@ void ggml_quantize_init(enum ggml_type type) {
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_IQ1_S:   iq2xs_init_impl(type); break;
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:   iq2xs_init_impl(type); break;
         case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
         case GGML_TYPE_IQ3_S:   iq3xs_init_impl(512); break;
         default: // nothing
@@ -20331,7 +20323,8 @@ bool ggml_quantize_requires_imatrix(enum ggml_type type) {
     return
         type == GGML_TYPE_IQ2_XXS ||
         type == GGML_TYPE_IQ2_XS  ||
-        type == GGML_TYPE_IQ1_S;
+        type == GGML_TYPE_IQ1_S;//   ||
+        //type == GGML_TYPE_IQ1_M;
 }
 
 size_t ggml_quantize_chunk(
@@ -20375,6 +20368,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ3_S:   result = quantize_iq3_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_S:   result = quantize_iq2_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_S:   result = quantize_iq1_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
 #if QK_K == 64
         case GGML_TYPE_IQ4_XS:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;

ggml.h

@@ -369,6 +369,7 @@ extern "C" {
         GGML_TYPE_I32     = 26,
         GGML_TYPE_I64     = 27,
         GGML_TYPE_F64     = 28,
+        GGML_TYPE_IQ1_M   = 29,
         GGML_TYPE_COUNT,
     };
 
@@ -408,6 +409,7 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ3_S   = 20, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ2_S   = 21, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
     };
 
     // available tensor operations:
@@ -748,6 +750,7 @@ extern "C" {
     GGML_API GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor);
     GGML_API GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor);
     GGML_API GGML_CALL bool ggml_is_permuted  (const struct ggml_tensor * tensor);
+    GGML_API GGML_CALL bool ggml_is_empty     (const struct ggml_tensor * tensor);
     GGML_API           bool ggml_is_scalar    (const struct ggml_tensor * tensor);
     GGML_API           bool ggml_is_vector    (const struct ggml_tensor * tensor);
     GGML_API           bool ggml_is_matrix    (const struct ggml_tensor * tensor);
@@ -1161,8 +1164,7 @@ extern "C" {
     //  ggml_mul_mat_id(ctx, as, ids, id, b) ~= ggml_mul_mat(as[ids[id]], b)
     GGML_API struct ggml_tensor * ggml_mul_mat_id(
             struct ggml_context * ctx,
-            struct ggml_tensor  * const as[],
-            int                   n_as,
+            struct ggml_tensor  * as,
             struct ggml_tensor  * ids,
             int                   id,
             struct ggml_tensor  * b);

ggml_vk_generate_shaders.py

@@ -18,6 +18,12 @@ shader_int8_ext = """
 """
 
 # Type-specific defines
+shader_f32_defines = """
+#define QUANT_K 1
+#define QUANT_R 1
+
+#define A_TYPE float
+"""
 shader_f16_defines = """
 #define QUANT_K 1
 #define QUANT_R 1
@@ -157,8 +163,8 @@ struct block_q6_K
 """
 
 # Dequant functions
-shader_f16_dequant_func = """
-#define DEQUANT_FUNC vec2 v = vec2(data_a[ib + 0], data_a[ib + 1]);
+shader_float_dequant_func = """
+#define DEQUANT_FUNC vec2 v = vec2(ib, ib);  // data_a[ib], data_a[ib + 1]);
 """
 
 shader_q4_0_dequant_func = """
@@ -410,6 +416,133 @@ mulmat_load_q8_0 = """
             buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
             buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);"""
 
+
+mulmat_load_q2_K = """
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                         // 2 values per idx
+            const uint iqs = idx % 128;                        // 0..127
+
+            const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
+            const uint scalesi = iqs / 8;                      // 0..15
+            const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
+
+            const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
+            const uint scales = data_a[ib].scales[scalesi];
+            const vec2 d = vec2(data_a[ib].d);
+
+            const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);"""
+
+mulmat_load_q3_K = """
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                   // 2 values per idx
+            const uint iqs = idx % 128;                  // 0..127
+
+            const uint n = iqs / 64;                     // 0,1
+            const uint qsi = n * 32 + (iqs % 16) * 2;    // 0,2,4..62
+            const uint hmi =          (iqs % 16) * 2;    // 0,2,4..30
+            const uint j = (iqs % 64) / 4;               // 0..3
+            const uint is = iqs / 8;                     // 0..15
+            const uint halfsplit = ((iqs % 64) / 16);    // 0,1,2,3
+            const uint qsshift = halfsplit * 2;          // 0,2,4,6
+            const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128
+
+            const int8_t us = int8_t(is <  4 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+8] >> 0) & 3) << 4) :
+                                    is <  8 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+4] >> 2) & 3) << 4) :
+                                    is < 12 ? (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is+0] >> 4) & 3) << 4) :
+                                            (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is-4] >> 6) & 3) << 4));
+            const float dl = float(data_a[ib].d) * float(us - 32);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi    ] & m) != 0) ? 0 : 4)));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));"""
+
+mulmat_load_q4_K = """
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                 // 2 values per idx
+            const uint iqs = idx % 128;                // 0..127
+
+            const uint n = iqs / 32;                   // 0,1,2,3
+            const uint b = (iqs % 32) / 16;            // 0,1
+            const uint is = 2 * n + b;                 // 0..7
+            const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+
+            const vec2 loadd = vec2(data_a[ib].d);
+
+            uint8_t sc;
+            uint8_t mbyte;
+            if (is < 4) {
+                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
+                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
+            } else {
+                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
+                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
+            }
+            const float d = loadd.x * sc;
+            const float m = loadd.y * mbyte;
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) - m);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) - m);"""
+
+mulmat_load_q5_K = """
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                 // 2 values per idx
+            const uint iqs = idx % 128;                // 0..127
+
+            const uint n = iqs / 32;                   // 0,1,2,3
+            const uint b = (iqs % 32) / 16;            // 0,1
+            const uint is = 2 * n + b;                 // 0..7
+            const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+            const uint qhi = (iqs % 16) * 2;           // 0,2,4..30
+
+            const uint8_t hm = uint8_t(1 << (iqs / 16));
+
+            const vec2 loadd = vec2(data_a[ib].d);
+
+            uint8_t sc;
+            uint8_t mbyte;
+            if (is < 4) {
+                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
+                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
+            } else {
+                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
+                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
+            }
+            const float d = loadd.x * sc;
+            const float m = loadd.y * mbyte;
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0)) - m);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0)) - m);"""
+
+mulmat_load_q6_K = """
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint iqs = idx % 128;                 // 0..127
+
+            const uint n = iqs / 64;                    // 0,1
+            const uint b = (iqs % 64) / 32;             // 0,1
+            const uint is_b = (iqs % 16) / 8;           // 0,1
+            const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
+            const uint is = 8 * n + qhshift + is_b;     // 0..15
+            const uint qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
+            const uint qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+
+            const float dscale = float(data_a[ib].d) * float(data_a[ib].scales[is]);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));"""
+
 mulmat_body2 = """
         }
         [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
@@ -1611,8 +1744,9 @@ layout (push_constant) uniform parameter
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint d_offset;
     float param1; float param2;
-} p;
+} p;"""
 
+generic_unary_op_funcs = """
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -1636,14 +1770,17 @@ uint dst_idx(uint idx) {
     const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
     const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
     return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
-}
+}"""
 
+generic_unary_op_main = """
 void main() {
     if (gl_GlobalInvocationID.x >= p.ne) {
         return;
     }
 """
 
+generic_unary_op_combined = f"{generic_unary_op_head}\n{generic_unary_op_funcs}\n{generic_unary_op_main}"
+
 generic_binary_op_head = """#version 450
 
 #extension GL_EXT_shader_16bit_storage : require
@@ -1655,13 +1792,14 @@ layout (push_constant) uniform parameter
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
     uint d_offset;
-    uint param1; uint param2;
-} p;
+    float param1; float param2;
+} p;"""
 
+generic_binary_op_funcs = """
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer B {A_TYPE data_b[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 uint src0_idx(uint idx) {
@@ -1693,14 +1831,17 @@ uint dst_idx(uint idx) {
     const uint i21 = (idx - i23_offset - i22_offset) / p.ne20;
     const uint i20 = idx - i23_offset - i22_offset - i21*p.ne20;
     return i23*p.nb23 + i22*p.nb22 + i21*p.nb21 + i20*p.nb20;
-}
+}"""
 
+generic_binary_op_main = """
 void main() {
     if (gl_GlobalInvocationID.x >= p.ne) {
         return;
     }
 """
 
+generic_binary_op_combined = f"{generic_binary_op_head}\n{generic_binary_op_funcs}\n{generic_binary_op_main}"
+
 # MUL F32
 mul_body = """
     data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
@@ -1745,39 +1886,55 @@ cpy_f16_f16_end = """
 """
 
 # GET_ROWS
-get_rows_body = """
-#extension GL_EXT_control_flow_attributes : enable
-#extension GL_EXT_shader_8bit_storage : require
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer Y {int data_b[];};
-layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
-
+get_rows_float_body = """
 void main() {
-    const uint col = int(gl_GlobalInvocationID.x) * 2;
-    const uint row = int(gl_GlobalInvocationID.y);
+    const uint i00 = gl_GlobalInvocationID.x;
+    const uint i10 = gl_GlobalInvocationID.y;
+    const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
+    const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
 
-    if (col >= p.KY) {
+    if (i00 >= p.ne00) {
         return;
     }
 
-    const uint r = uint(data_b[row]);
+    const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
 
-    // copy data_a[r*p.KY + col] to dst[row*p.KX + col]
-    const uint xi = r*p.KY + col;
-    const uint di = row*p.KY + col;
+    const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+    const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
-    const uint ib = xi/QUANT_K; // block index
-    const uint iqs = (xi%QUANT_K)/QUANT_R; // quant index
-    const uint iybs = di - di%QUANT_K; // y block start index
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);
+#else
+    data_d[d_offset + i00] = data_a[a_offset + i00];
+#endif
+}
+"""
+
+get_rows_body = """
+void main() {
+    const uint i00 = (gl_GlobalInvocationID.x)*2;
+    const uint i10 = gl_GlobalInvocationID.y;
+    const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
+    const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
+
+    if (i00 >= p.ne00) {
+        return;
+    }
+
+    const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+
+    const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+    const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+
+    const uint ib = a_offset + i00/QUANT_K; // block index
+    const uint iqs = (i00%QUANT_K)/QUANT_R; // quant index
+    const uint iybs = i00 - i00%QUANT_K; // dst block start index
     const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
     DEQUANT_FUNC
 
-    dst[iybs + iqs + 0]        = D_TYPE(v.x);
-    dst[iybs + iqs + y_offset] = D_TYPE(v.y);
+    data_d[d_offset + iybs + iqs           ] = D_TYPE(v.x);
+    data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);
 }
 """
 
@@ -2418,6 +2575,31 @@ async def main():
         tasks.append(string_to_spv("matmul_q8_0_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q8_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
         tasks.append(string_to_spv("matmul_q8_0_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q8_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
 
+        stream.clear()
+        stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q2_K_defines, mulmat_body1, mulmat_load_q2_K, mulmat_body2))
+        tasks.append(string_to_spv("matmul_q2_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q2_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_q2_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q2_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
+
+        stream.clear()
+        stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q3_K_defines, mulmat_body1, mulmat_load_q3_K, mulmat_body2))
+        tasks.append(string_to_spv("matmul_q3_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q3_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_q3_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q3_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
+
+        stream.clear()
+        stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_K_defines, mulmat_body1, mulmat_load_q4_K, mulmat_body2))
+        tasks.append(string_to_spv("matmul_q4_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q4_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_q4_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
+
+        stream.clear()
+        stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_K_defines, mulmat_body1, mulmat_load_q5_K, mulmat_body2))
+        tasks.append(string_to_spv("matmul_q5_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q5_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_q5_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
+
+        stream.clear()
+        stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q6_K_defines, mulmat_body1, mulmat_load_q6_K, mulmat_body2))
+        tasks.append(string_to_spv("matmul_q6_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q6_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_q6_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q6_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
+
     # Shaders where precision is needed, so no fp16 version
 
     # mul mat vec
@@ -2426,7 +2608,7 @@ async def main():
         stream.extend((mul_mat_vec_head, shader_int8_ext, shader_f32))
 
         if i == GGML_TYPE_F16:
-            stream.extend((shader_f16_defines, shader_f16_dequant_func, mul_mat_vec_body))
+            stream.extend((shader_f16_defines, shader_float_dequant_func, mul_mat_vec_body))
         elif i == GGML_TYPE_Q4_0:
             stream.extend((shader_q4_0_defines, shader_q4_0_dequant_func, mul_mat_vec_body))
         elif i == GGML_TYPE_Q4_1:
@@ -2488,25 +2670,32 @@ async def main():
     # get_rows
     for i in range(0, VK_NUM_TYPES):
         stream.clear()
-        stream.extend((generic_head, shader_int8_ext, shader_f32))
+        stream.extend((generic_binary_op_head, shader_int8_ext, shader_f32))
+        optimization_workaround = False
 
-        if i == GGML_TYPE_F16:
-            stream.extend((shader_f16_defines,  shader_f16_dequant_func,  get_rows_body))
+        if i == GGML_TYPE_F32:
+            stream.extend((shader_f32_defines, generic_binary_op_funcs, get_rows_float_body))
+        elif i == GGML_TYPE_F16:
+            stream.extend((shader_f16_defines, generic_binary_op_funcs, get_rows_float_body))
+            optimization_workaround = True
         elif i == GGML_TYPE_Q4_0:
-            stream.extend((shader_q4_0_defines, shader_q4_0_dequant_func, get_rows_body))
+            stream.extend((shader_q4_0_defines, shader_q4_0_dequant_func, generic_binary_op_funcs, get_rows_body))
         elif i == GGML_TYPE_Q4_1:
-            stream.extend((shader_q4_1_defines, shader_q4_1_dequant_func, get_rows_body))
+            stream.extend((shader_q4_1_defines, shader_q4_1_dequant_func, generic_binary_op_funcs, get_rows_body))
         elif i == GGML_TYPE_Q5_0:
-            stream.extend((shader_q5_0_defines, shader_q5_0_dequant_func, get_rows_body))
+            stream.extend((shader_q5_0_defines, shader_q5_0_dequant_func, generic_binary_op_funcs, get_rows_body))
         elif i == GGML_TYPE_Q5_1:
-            stream.extend((shader_q5_1_defines, shader_q5_1_dequant_func, get_rows_body))
+            stream.extend((shader_q5_1_defines, shader_q5_1_dequant_func, generic_binary_op_funcs, get_rows_body))
         elif i == GGML_TYPE_Q8_0:
-            stream.extend((shader_q8_0_defines, shader_q8_0_dequant_func, get_rows_body))
+            stream.extend((shader_q8_0_defines, shader_q8_0_dequant_func, generic_binary_op_funcs, get_rows_body))
         else:
             continue
 
-        tasks.append(string_to_spv(f"get_rows_{type_names[i]}", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float16_t"}))
-        tasks.append(string_to_spv(f"get_rows_{type_names[i]}_f32", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float"}))
+        if optimization_workaround:
+            tasks.append(string_to_spv(f"get_rows_{type_names[i]}", "".join(stream), {"B_TYPE": "int", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
+        else:
+            tasks.append(string_to_spv(f"get_rows_{type_names[i]}", "".join(stream), {"B_TYPE": "int", "D_TYPE": "float16_t"}))
+        tasks.append(string_to_spv(f"get_rows_{type_names[i]}_f32", "".join(stream), {"B_TYPE": "int", "D_TYPE": "float"}))
 
     tasks.append(string_to_spv("mul_mat_vec_p021_f16_f32", mul_mat_p021_src, {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
     tasks.append(string_to_spv("mul_mat_vec_nc_f16_f32", mul_mat_nc_src, {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
@@ -2515,20 +2704,20 @@ async def main():
     tasks.append(string_to_spv("norm_f32", f"{generic_head}\n{shader_f32}\n{norm_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
     tasks.append(string_to_spv("rms_norm_f32", f"{generic_head}\n{shader_f32}\n{rms_norm_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
 
-    tasks.append(string_to_spv("cpy_f32_f32", f"{generic_unary_op_head}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("cpy_f32_f16", f"{generic_unary_op_head}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
-    tasks.append(string_to_spv("cpy_f16_f16", f"{generic_unary_op_head}\n{cpy_f16_f16_end}", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
+    tasks.append(string_to_spv("cpy_f32_f32", f"{generic_unary_op_combined}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("cpy_f32_f16", f"{generic_unary_op_combined}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
+    tasks.append(string_to_spv("cpy_f16_f16", f"{generic_unary_op_combined}\n{cpy_f16_f16_end}", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
 
-    tasks.append(string_to_spv("add_f32", f"{generic_binary_op_head}\n{add_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+    tasks.append(string_to_spv("add_f32", f"{generic_binary_op_combined}\n{add_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
 
     tasks.append(string_to_spv("split_k_reduce", mulmat_split_k_reduce_src, {}))
-    tasks.append(string_to_spv("mul_f32", f"{generic_binary_op_head}\n{mul_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+    tasks.append(string_to_spv("mul_f32", f"{generic_binary_op_combined}\n{mul_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
 
-    tasks.append(string_to_spv("scale_f32", f"{generic_unary_op_head}\n{scale_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+    tasks.append(string_to_spv("scale_f32", f"{generic_unary_op_combined}\n{scale_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
 
-    tasks.append(string_to_spv("sqr_f32", f"{generic_unary_op_head}\n{sqr_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+    tasks.append(string_to_spv("sqr_f32", f"{generic_unary_op_combined}\n{sqr_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
 
-    tasks.append(string_to_spv("clamp_f32", f"{generic_unary_op_head}\n{clamp_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+    tasks.append(string_to_spv("clamp_f32", f"{generic_unary_op_combined}\n{clamp_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
 
     tasks.append(string_to_spv("gelu_f32", f"{generic_head}\n{shader_f32}\n{gelu_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
     tasks.append(string_to_spv("silu_f32", f"{generic_head}\n{shader_f32}\n{silu_body}", {"A_TYPE": "float", "D_TYPE": "float"}))

gguf-py/gguf/constants.py

@@ -123,6 +123,7 @@ class MODEL_ARCH(IntEnum):
     GEMMA      = auto()
     STARCODER2 = auto()
     MAMBA      = auto()
+    XVERSE     = auto()
     COMMAND_R  = auto()
 
 
@@ -191,6 +192,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.GEMMA:          "gemma",
     MODEL_ARCH.STARCODER2:     "starcoder2",
     MODEL_ARCH.MAMBA:          "mamba",
+    MODEL_ARCH.XVERSE:         "xverse",
     MODEL_ARCH.COMMAND_R:      "command-r",
 }
 
@@ -219,9 +221,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_DOWN:        "blk.{bid}.ffn_down",
     MODEL_TENSOR.FFN_UP:          "blk.{bid}.ffn_up",
     MODEL_TENSOR.FFN_ACT:         "blk.{bid}.ffn",
-    MODEL_TENSOR.FFN_GATE_EXP:    "blk.{bid}.ffn_gate.{xid}",
-    MODEL_TENSOR.FFN_DOWN_EXP:    "blk.{bid}.ffn_down.{xid}",
-    MODEL_TENSOR.FFN_UP_EXP:      "blk.{bid}.ffn_up.{xid}",
+    MODEL_TENSOR.FFN_GATE_EXP:    "blk.{bid}.ffn_gate_exps",
+    MODEL_TENSOR.FFN_DOWN_EXP:    "blk.{bid}.ffn_down_exps",
+    MODEL_TENSOR.FFN_UP_EXP:      "blk.{bid}.ffn_up_exps",
     MODEL_TENSOR.LAYER_OUT_NORM:  "blk.{bid}.layer_output_norm",
     MODEL_TENSOR.SSM_IN:          "blk.{bid}.ssm_in",
     MODEL_TENSOR.SSM_CONV1D:      "blk.{bid}.ssm_conv1d",
@@ -365,6 +367,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.FFN_ACT,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.POS_EMBD,
     ],
     MODEL_ARCH.GPTJ: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -606,6 +611,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.SSM_D,
         MODEL_TENSOR.SSM_OUT,
     ],
+    MODEL_ARCH.XVERSE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.COMMAND_R: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -650,6 +671,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
     ],
+    MODEL_ARCH.XVERSE: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }
 
 #
@@ -706,6 +731,7 @@ class GGMLQuantizationType(IntEnum):
     I32     = 26
     I64     = 27
     F64     = 28
+    IQ1_M   = 29
 
 
 class GGUFEndian(IntEnum):

gguf-py/gguf/tensor_mapping.py

@@ -231,9 +231,8 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
-            "layers.{bid}.feed_forward.experts.{xid}.w3",           # mixtral
-            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w3", # mixtral
-            "transformer.decoder_layer.{bid}.moe.{xid}.linear_v",   # Grok
+            "layers.{bid}.feed_forward.experts.w3",                 # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_v",         # Grok (merged)
         ),
 
         # AWQ-activation gate
@@ -252,9 +251,8 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
-            "layers.{bid}.feed_forward.experts.{xid}.w1",           # mixtral
-            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w1", # mixtral
-            "transformer.decoder_layer.{bid}.moe.{xid}.linear"      # Grok
+            "layers.{bid}.feed_forward.experts.w1",                 # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear"            # Grok (merged)
         ),
 
         # Feed-forward down
@@ -280,20 +278,20 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
-            "layers.{bid}.feed_forward.experts.{xid}.w2",           # mixtral
-            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w2", # mixtral
-            "transformer.decoder_layer.{bid}.moe.{xid}.linear_1",   # Grok
-
+            "layers.{bid}.feed_forward.experts.w2",                 # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_1",         # Grok (merged)
         ),
 
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
+            "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
         ),
 
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
+            "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
         ),
 
         MODEL_TENSOR.ROPE_FREQS: (

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.8.0"
+version = "0.9.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

llama.h

@@ -39,7 +39,7 @@
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 4
+#define LLAMA_SESSION_VERSION 5
 
 #ifdef __cplusplus
 extern "C" {
@@ -60,9 +60,9 @@ extern "C" {
 
     enum llama_vocab_type {
         LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
-        LLAMA_VOCAB_TYPE_SPM  = 1, // SentencePiece
-        LLAMA_VOCAB_TYPE_BPE  = 2, // Byte Pair Encoding
-        LLAMA_VOCAB_TYPE_WPM  = 3, // WordPiece
+        LLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
+        LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
+        LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
     };
 
     // note: these values should be synchronized with ggml_rope
@@ -117,6 +117,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ2_S         = 28, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ2_M         = 29, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -284,6 +285,7 @@ extern "C" {
         bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
         bool pure;                           // quantize all tensors to the default type
         void * imatrix;                      // pointer to importance matrix data
+        void * kv_overrides;                 // pointer to vector containing overrides
     } llama_model_quantize_params;
 
     // grammar types
@@ -676,23 +678,29 @@ extern "C" {
     LLAMA_API void llama_synchronize(struct llama_context * ctx);
 
     // Token logits obtained from the last call to llama_decode()
-    // The logits for the last token are stored in the last row
-    // Logits for which llama_batch.logits[i] == 0 are undefined
-    // Rows: n_tokens provided with llama_batch
+    // The logits for which llama_batch.logits[i] != 0 are stored contiguously
+    // in the order they have appeared in the batch.
+    // Rows: number of tokens for which llama_batch.logits[i] != 0
     // Cols: n_vocab
     LLAMA_API float * llama_get_logits(struct llama_context * ctx);
 
     // Logits for the ith token. Equivalent to:
-    // llama_get_logits(ctx) + i*n_vocab
+    // llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab
+    // returns NULL for invalid ids.
     LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
 
-    // Get all output token embeddings
-    // shape: [n_tokens*n_embd] (1-dimensional)
+    // Get all output token embeddings.
+    // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model,
+    // the embeddings for which llama_batch.logits[i] != 0 are stored contiguously
+    // in the order they have appeared in the batch.
+    // shape: [n_outputs*n_embd]
+    // Otherwise, returns NULL.
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
-    // Get the embeddings for the ith token
-    // llama_get_embeddings(ctx) + i*n_embd
+    // Get the embeddings for the ith token. Equivalent to:
+    // llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd
     // shape: [n_embd] (1-dimensional)
+    // returns NULL for invalid ids.
     LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
 
     // Get the embeddings for a sequence id
@@ -999,10 +1007,38 @@ extern "C" {
 
 struct ggml_tensor;
 
+struct llama_partial_utf8 {
+    uint32_t value;    // bit value so far (unshifted)
+    int      n_remain; // num bytes remaining; -1 indicates invalid sequence
+};
+
+struct llama_grammar {
+    const std::vector<std::vector<llama_grammar_element>>   rules;
+    std::vector<std::vector<const llama_grammar_element *>> stacks;
+
+    // buffer for partially generated UTF-8 sequence from accepted tokens
+    llama_partial_utf8                                      partial_utf8;
+};
+
+struct llama_grammar_candidate {
+    size_t               index;
+    const uint32_t     * code_points;
+    llama_partial_utf8   partial_utf8;
+};
+
 const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
     struct llama_context * ctx
 );
 
+std::vector<std::vector<const llama_grammar_element *>> llama_grammar_accept(
+        const std::vector<std::vector<llama_grammar_element>>         & rules,
+        const std::vector<std::vector<const llama_grammar_element *>> & stacks,
+        const uint32_t                                                  chr);
+
+std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
+        const std::string & src,
+        llama_partial_utf8   partial_start);
+
 #endif // LLAMA_API_INTERNAL
 
 #endif // LLAMA_H

scripts/compare-llama-bench.py

@@ -178,6 +178,9 @@ def get_commit_hexsha8(name):
     for t in repo.tags:
         if t.name == name:
             return t.commit.hexsha[:8]
+    for c in repo.iter_commits("--all"):
+        if c.hexsha[:8] == name[:8]:
+            return c.hexsha[:8]
     return None
 
 
@@ -224,7 +227,7 @@ if known_args.compare is not None:
         hexsha8_compare = get_commit_hexsha8(known_args.compare)
         name_compare = known_args.compare
     if hexsha8_compare is None:
-        print(f"ERROR: cannot find data for baseline={known_args.compare}.")
+        print(f"ERROR: cannot find data for compare={known_args.compare}.")
         sys.exit(1)
 # Otherwise, search for the commit for llama-bench was most recently run
 # and that is not a parent of master:

scripts/sync-ggml-am.sh

@@ -95,6 +95,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-backend-impl.h     -> ggml-backend-impl.h
     # src/ggml-backend.c          -> ggml-backend.c
     # src/ggml-common.h           -> ggml-common.h
+    # src/ggml-cuda/*             -> ggml-cuda/
     # src/ggml-cuda.cu            -> ggml-cuda.cu
     # src/ggml-cuda.h             -> ggml-cuda.h
     # src/ggml-impl.h             -> ggml-impl.h
@@ -128,6 +129,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/src\/ggml-backend-impl\.h/ggml-backend-impl.h/g' \
         -e 's/src\/ggml-backend\.c/ggml-backend.c/g' \
         -e 's/src\/ggml-common\.h/ggml-common.h/g' \
+        -e 's/src\/ggml-cuda\//ggml-cuda\//g' \
         -e 's/src\/ggml-cuda\.cu/ggml-cuda.cu/g' \
         -e 's/src\/ggml-cuda\.h/ggml-cuda.h/g' \
         -e 's/src\/ggml-impl\.h/ggml-impl.h/g' \

scripts/sync-ggml.sh

@@ -5,6 +5,7 @@ cp -rpv ../ggml/src/ggml-alloc.c            ./ggml-alloc.c
 cp -rpv ../ggml/src/ggml-backend-impl.h     ./ggml-backend-impl.h
 cp -rpv ../ggml/src/ggml-backend.c          ./ggml-backend.c
 cp -rpv ../ggml/src/ggml-common.h           ./ggml-common.h
+cp -rpv ../ggml/src/ggml-cuda/*             ./ggml-cuda/
 cp -rpv ../ggml/src/ggml-cuda.cu            ./ggml-cuda.cu
 cp -rpv ../ggml/src/ggml-cuda.h             ./ggml-cuda.h
 cp -rpv ../ggml/src/ggml-impl.h             ./ggml-impl.h

tests/test-backend-ops.cpp

@@ -979,17 +979,13 @@ struct test_mul_mat_id : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
-        std::vector<ggml_tensor *> mats;
-        for (int i = 0; i < n_mats; i++) {
-            ggml_tensor * a = ggml_new_tensor_2d(ctx, type_a, k, m);
-            mats.push_back(a);
-        }
+        ggml_tensor * mats = ggml_new_tensor_3d(ctx, type_a, k, m, n_mats);
         ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_mats, n);
         if (v) {
             ids = ggml_view_2d(ctx, ids, n_mats/2, ids->ne[1], ids->nb[1], 0);
         }
         ggml_tensor * b = ggml_new_tensor_2d(ctx, type_b, k, n);
-        ggml_tensor * out = ggml_mul_mat_id(ctx, mats.data(), n_mats, ids, v ? id/2 : id, b);
+        ggml_tensor * out = ggml_mul_mat_id(ctx, mats, ids, v ? id/2 : id, b);
         return out;
     }
 
@@ -1477,91 +1473,6 @@ struct test_leaky_relu : public test_case {
     }
 };
 
-// Mixtral MOE
-struct test_moe : public test_case {
-    const int n_experts;
-    const int n_experts_per_tok;
-    const int n_tokens;
-    const int n_embd;
-    const int n_ff;
-
-    std::string op_desc(ggml_tensor * t) override {
-        return "MOE";
-
-        GGML_UNUSED(t);
-    }
-
-    std::string vars() override {
-        return VARS_TO_STR5(n_experts, n_experts_per_tok, n_tokens, n_embd, n_ff);
-    }
-
-    test_moe(int n_experts = 8, int n_experts_per_tok = 2, int n_tokens = 1, int n_embd = 4096, int n_ff = 14336)
-        : n_experts(n_experts), n_experts_per_tok(n_experts_per_tok), n_tokens(n_tokens), n_embd(n_embd), n_ff(n_ff) {
-    }
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * ffn_gate_inp = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_experts);
-
-        std::vector<ggml_tensor *> ffn_up_exp(n_experts);
-        std::vector<ggml_tensor *> ffn_gate_exp(n_experts);
-        std::vector<ggml_tensor *> ffn_down_exp(n_experts);
-
-        for (int i = 0; i < n_experts; ++i) {
-            ffn_up_exp[i] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
-            ffn_gate_exp[i] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
-            ffn_down_exp[i] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd);
-        }
-
-        ggml_tensor * cur = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens);
-
-        ggml_tensor * logits = ggml_mul_mat(ctx, ffn_gate_inp, cur);
-        ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, nullptr, 1.0f/sqrtf(n_embd), 0.0f);
-
-        // select experts
-        ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_experts_per_tok);
-
-        ggml_tensor * weights = ggml_get_rows(ctx,
-                ggml_reshape_3d(ctx, probs, 1, n_experts, n_tokens), selected_experts);
-
-        weights = ggml_reshape_2d(ctx, weights, n_experts_per_tok, n_tokens);
-
-        ggml_tensor * weights_sum = ggml_sum_rows(ctx, weights);
-
-        weights = ggml_div(ctx, weights, weights_sum);
-
-        // compute expert outputs
-        ggml_tensor * moe_out = nullptr;
-
-        for (int i = 0; i < n_experts_per_tok; ++i) {
-            ggml_tensor * cur_expert;
-
-            ggml_tensor * cur_up = ggml_mul_mat_id(ctx, ffn_up_exp.data(), n_experts, selected_experts, i, cur);
-
-            ggml_tensor * cur_gate = ggml_mul_mat_id(ctx, ffn_gate_exp.data(), n_experts, selected_experts, i, cur);
-
-            cur_gate = ggml_silu(ctx, cur_gate);
-
-            cur_expert = ggml_mul(ctx, cur_up, cur_gate);
-
-            cur_expert = ggml_mul_mat_id(ctx, ffn_down_exp.data(), n_experts, selected_experts, i, cur_expert);
-
-            cur_expert = ggml_mul(ctx, cur_expert,
-                    ggml_view_2d(ctx, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0]));
-
-            if (i == 0) {
-                moe_out = cur_expert;
-            } else {
-                moe_out = ggml_add(ctx, moe_out, cur_expert);
-            }
-        }
-
-        cur = moe_out;
-
-        return cur;
-    }
-};
-
-
 enum llm_norm_type {
     LLM_NORM,
     LLM_NORM_RMS,
@@ -1960,7 +1871,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
         GGML_TYPE_Q6_K,
         GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S,
-        GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S,
+        GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,
         GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS,
     };
 
@@ -2169,6 +2080,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     for (ggml_sort_order order : {GGML_SORT_ORDER_ASC, GGML_SORT_ORDER_DESC}) {
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {8, 1, 1, 1}, order));
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {16, 10, 10, 10}, order));
+        test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {60, 10, 10, 10}, order)); // qwen
     }
 
     test_cases.emplace_back(new test_sum_rows());
@@ -2182,11 +2094,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     // these tests are disabled to save execution time, but they can be handy for debugging
 #if 0
-#if !defined(__SANITIZE_THREAD__)
-    // FIXME: these tests use too much memory with thread sanitizer
-    test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 8*1024));
-    //test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
-#endif
     test_cases.emplace_back(new test_llama(1));
     test_cases.emplace_back(new test_llama(2));
     test_cases.emplace_back(new test_falcon(1));

tests/test-chat-template.cpp

@@ -33,6 +33,18 @@ int main(void) {
         "{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\\n' + message['content'] | trim + '<end_of_turn>\\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\\n'}}{% endif %}",
         // OrionStarAI/Orion-14B-Chat
         "{% for message in messages %}{% if loop.first %}{{ bos_token }}{% endif %}{% if message['role'] == 'user' %}{{ 'Human: ' + message['content'] + '\\n\\nAssistant: ' + eos_token }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token }}{% endif %}{% endfor %}",
+        // openchat/openchat-3.5-0106
+        // The included chat_template differs from the author's suggestions here: https://huggingface.co/openchat/openchat_3.5/discussions/5#65448109b4a3f3a2f486fd9d
+        // So we match against the included template but implement the suggested version.
+        "{{ bos_token }}{% for message in messages %}{{ 'GPT4 Correct ' + message['role'].title() + ': ' + message['content'] + '<|end_of_turn|>'}}{% endfor %}{% if add_generation_prompt %}{{ 'GPT4 Correct Assistant:' }}{% endif %}",
+        // deepseek-ai/deepseek-coder-33b-instruct
+        "{% if not add_generation_prompt is defined %}\n{% set add_generation_prompt = false %}\n{% endif %}\n{%- set ns = namespace(found=false) -%}\n{%- for message in messages -%}\n    {%- if message['role'] == 'system' -%}\n        {%- set ns.found = true -%}\n    {%- endif -%}\n{%- endfor -%}\n{{bos_token}}{%- if not ns.found -%}\n{{'You are an AI programming assistant, utilizing the Deepseek Coder model, developed by Deepseek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer\\n'}}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'system' %}\n{{ message['content'] }}\n    {%- else %}\n        {%- if message['role'] == 'user' %}\n{{'### Instruction:\\n' + message['content'] + '\\n'}}\n        {%- else %}\n{{'### Response:\\n' + message['content'] + '\\n<|EOT|>\\n'}}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{% if add_generation_prompt %}\n{{'### Response:'}}\n{% endif %}",
+        // eachadea/vicuna-13b-1.1
+        // No template included in tokenizer_config.json, so this template likely needs to be manually set.
+        "{%- for message in messages %}{%- if message['role'] == 'system' -%}{{- '' + message['content'] + '\n\n' -}}{%- else -%}{%- if message['role'] == 'user' -%}{{-'USER: ' + message['content'] + '\n'-}}{%- else -%}{{-'ASSISTANT: ' + message['content'] + '</s>\n' -}}{%- endif -%}{%- endif -%}{%- endfor -%}{%- if add_generation_prompt -%}{{-'ASSISTANT:'-}}{%- endif -%}",
+        // Orca-Vicuna
+        // No template included in tokenizer_config.json, so this template likely needs to be manually set.
+        "{%- for message in messages %}{%- if message['role'] == 'system' -%}{{-'SYSTEM: ' + message['content'] + '\n' -}}{%- else -%}{%- if message['role'] == 'user' -%}{{-'USER: ' + message['content'] + '\n'-}}{%- else -%}{{-'ASSISTANT: ' + message['content'] + '</s>\n' -}}{%- endif -%}{%- endif -%}{%- endfor -%}{%- if add_generation_prompt -%}{{-'ASSISTANT:'-}}{%- endif -%}",
     };
     std::vector<std::string> expected_output = {
         // teknium/OpenHermes-2.5-Mistral-7B
@@ -49,6 +61,14 @@ int main(void) {
         "<start_of_turn>user\nYou are a helpful assistant\n\nHello<end_of_turn>\n<start_of_turn>model\nHi there<end_of_turn>\n<start_of_turn>user\nWho are you<end_of_turn>\n<start_of_turn>model\nI am an assistant<end_of_turn>\n<start_of_turn>user\nAnother question<end_of_turn>\n<start_of_turn>model\n",
         // OrionStarAI/Orion-14B-Chat
         "Human: You are a helpful assistant\n\nHello\n\nAssistant: </s>Hi there</s>Human: Who are you\n\nAssistant: </s>   I am an assistant   </s>Human: Another question\n\nAssistant: </s>",
+        // openchat/openchat-3.5-0106
+        "You are a helpful assistant<|end_of_turn|>GPT4 Correct User: Hello<|end_of_turn|>GPT4 Correct Assistant: Hi there<|end_of_turn|>GPT4 Correct User: Who are you<|end_of_turn|>GPT4 Correct Assistant:    I am an assistant   <|end_of_turn|>GPT4 Correct User: Another question<|end_of_turn|>GPT4 Correct Assistant:",
+        // deepseek-ai/deepseek-coder-33b-instruct
+        "You are a helpful assistant### Instruction:\nHello\n### Response:\nHi there\n<|EOT|>\n### Instruction:\nWho are you\n### Response:\n   I am an assistant   \n<|EOT|>\n### Instruction:\nAnother question\n### Response:\n",
+        // eachadea/vicuna-13b-1.1
+        "You are a helpful assistant\n\nUSER: Hello\nASSISTANT: Hi there</s>\nUSER: Who are you\nASSISTANT:    I am an assistant   </s>\nUSER: Another question\nASSISTANT:",
+        // Orca-Vicuna
+        "SYSTEM: You are a helpful assistant\nUSER: Hello\nASSISTANT: Hi there</s>\nUSER: Who are you\nASSISTANT:    I am an assistant   </s>\nUSER: Another question\nASSISTANT:",
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;