Add OpenChat, Alpaca, Vicuna chat templates (#6397 )

* Add openchat chat template * Add chat template test for openchat * Add chat template for vicuna * Add chat template for orca-vicuna * Add EOS for vicuna templates * Combine vicuna chat templates * Add tests for openchat and vicuna chat templates * Add chat template for alpaca * Add separate template name for vicuna-orca * Remove alpaca, match deepseek with jinja output * Regenerate chat template test with add_generation_prompt * Separate deepseek bos from system message * Match openchat template with jinja output * Remove BOS token from templates, unprefix openchat
readme : update hot topics
2026-07-10 22:45:53 +02:00 · 2024-04-03 17:24:31 +02:00 · 2024-04-03 16:11:15 +03:00 · 2024-04-03 16:07:05 +03:00 · 2024-04-03 10:34:40 +08:00 · 2024-04-01 09:05:57 -07:00
63 changed files with 31423 additions and 7641 deletions
@@ -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

@@ -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

@@ -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

@@ -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/
    '';
@@ -0,0 +1,279 @@
+# Benchmark
+name: Benchmark
+
+on:
+  workflow_dispatch:
+    inputs:
+      gpu-series:
+        description: 'Azure GPU series to run with'
+        required: true
+        type: choice
+        options:
+          - Standard_NC4as_T4_v3
+          - Standard_NC24ads_A100_v4
+          - Standard_NC80adis_H100_v5
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
+      duration:
+        description: 'Duration of the bench'
+        type: string
+        default: 10m
+
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/bench.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/bench/**.*']
+  pull_request_target:
+    types: [opened, synchronize, reopened]
+    paths: ['.github/workflows/bench.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/bench/**.*']
+  schedule:
+    -  cron: '04 2 * * *'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  bench-server-baseline:
+    runs-on: Standard_NC4as_T4_v3
+    env:
+      RUNNER_LABEL: Standard_NC4as_T4_v3 # FIXME Do not find a way to not duplicate it
+      N_USERS: 8
+      DURATION: 10m
+    if: ${{ github.event.inputs.gpu-series == 'Standard_NC4as_T4_v3' || github.event.schedule || github.event.pull_request || github.head_ref == 'master' || github.ref_name == 'master' || github.event.push.ref == 'refs/heads/master' }}
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Install python env
+        id: pipenv
+        run: |
+          cd examples/server/bench
+          python3 -m venv venv
+          source venv/bin/activate
+          pip install -r requirements.txt
+
+      - name: Prometheus
+        id: install_prometheus
+        run: |
+          wget --quiet https://github.com/prometheus/prometheus/releases/download/v2.51.0/prometheus-2.51.0.linux-amd64.tar.gz
+          tar xzf prometheus*.tar.gz --strip-components=1
+          ./prometheus --config.file=examples/server/bench/prometheus.yml &
+          while ! nc -z localhost 9090; do
+            sleep 0.1
+          done
+
+      - name: Install k6
+        id: k6_installation
+        run: |
+          cd examples/server/bench
+          wget --quiet https://github.com/grafana/k6/releases/download/v0.49.0/k6-v0.49.0-linux-amd64.tar.gz
+          tar xzf k6*.tar.gz --strip-components=1
+
+      - name: Build
+        id: cmake_build
+        run: |
+          set -eux
+          mkdir build
+          cd build
+          cmake .. \
+              -DLLAMA_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
+              -DLLAMA_CUBLAS=ON \
+              -DCUDAToolkit_ROOT=/usr/local/cuda \
+              -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
+              -DCMAKE_CUDA_ARCHITECTURES=75 \
+              -DLLAMA_FATAL_WARNINGS=OFF \
+              -DLLAMA_ALL_WARNINGS=OFF \
+              -DCMAKE_BUILD_TYPE=Release;
+          cmake --build . --config Release -j $(nproc) --target server
+
+      - name: Download the dataset
+        id: download_dataset
+        run: |
+          cd examples/server/bench
+          wget --quiet https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+
+      - name: Server bench
+        id: server_bench
+        run: |
+          set -eux
+
+          cd examples/server/bench
+          source venv/bin/activate
+          BENCH_K6_BIN_PATH=./k6 python bench.py \
+              --runner-label ${{ env.RUNNER_LABEL }} \
+              --name ${{ github.job }} \
+              --branch ${{ github.head_ref || github.ref_name }} \
+              --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
+              --scenario script.js \
+              --duration ${{ github.event.inputs.duration || env.DURATION }} \
+              --hf-repo ggml-org/models	 \
+              --hf-file phi-2/ggml-model-q4_0.gguf \
+              --model-path-prefix /models \
+              --parallel ${{ env.N_USERS }} \
+              -ngl 33 \
+              --batch-size 2048 \
+              --ubatch-size	256 \
+              --ctx-size 16384 \
+              --n-prompts 1000 \
+              --max-prompt-tokens 1024 \
+              --max-tokens 2048
+
+          cat results.github.env >> $GITHUB_ENV
+
+          # Remove dataset as we do not want it in the artefact
+          rm ShareGPT_V3_unfiltered_cleaned_split.json
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: benchmark-results
+          compression-level: 9
+          path: |
+            examples/server/bench/*.jpg
+            examples/server/bench/*.json
+            examples/server/bench/*.log
+
+      - name: Commit status
+        uses: Sibz/github-status-action@v1
+        with:
+          authToken: ${{secrets.GITHUB_TOKEN}}
+          sha: ${{ inputs.sha || github.event.pull_request.head.sha || github.sha }}
+          context: bench-server-baseline
+          description: |
+            ${{ env.BENCH_RESULTS }}
+          state: 'success'
+
+      - name: Upload benchmark images
+        uses: devicons/public-upload-to-imgur@v2.2.2
+        continue-on-error: true # Important as it looks unstable: 503
+        id: imgur_step
+        with:
+          client_id: ${{secrets.IMGUR_CLIENT_ID}}
+          path: |
+            examples/server/bench/prompt_tokens_seconds.jpg
+            examples/server/bench/predicted_tokens_seconds.jpg
+            examples/server/bench/kv_cache_usage_ratio.jpg
+            examples/server/bench/requests_processing.jpg
+
+      - name: Extract mermaid
+        id: set_mermaid
+        run: |
+          set -eux
+
+          cd examples/server/bench
+          PROMPT_TOKENS_SECONDS=$(cat prompt_tokens_seconds.mermaid)
+          echo "PROMPT_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
+          echo "$PROMPT_TOKENS_SECONDS" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          PREDICTED_TOKENS_SECONDS=$(cat predicted_tokens_seconds.mermaid)
+          echo "PREDICTED_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
+          echo "$PREDICTED_TOKENS_SECONDS" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          KV_CACHE_USAGE_RATIO=$(cat kv_cache_usage_ratio.mermaid)
+          echo "KV_CACHE_USAGE_RATIO<<EOF" >> $GITHUB_ENV
+          echo "$KV_CACHE_USAGE_RATIO" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          REQUESTS_PROCESSING=$(cat requests_processing.mermaid)
+          echo "REQUESTS_PROCESSING<<EOF" >> $GITHUB_ENV
+          echo "$REQUESTS_PROCESSING" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+      - name: Extract image url
+        id: extract_image_url
+        continue-on-error: true
+        run: |
+          set -eux
+
+          echo "IMAGE_O=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[0] }}" >> $GITHUB_ENV
+          echo "IMAGE_1=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[1] }}" >> $GITHUB_ENV
+          echo "IMAGE_2=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[2] }}" >> $GITHUB_ENV
+          echo "IMAGE_3=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[3] }}" >> $GITHUB_ENV
+
+      - name: Comment PR
+        uses: mshick/add-pr-comment@v2
+        id: comment_pr
+        if: ${{ github.event.pull_request != '' }}
+        with:
+          message-id: bench-${{ github.job }}-${{ env.RUNNER_LABEL }}
+          message: |
+            📈 **llama.cpp server** for _${{ github.job }}_ on _${{ env.RUNNER_LABEL }}_: **${{ env.BENCH_ITERATIONS}} iterations** 🚀
+
+            - Concurrent users: ${{ env.N_USERS }}, duration: ${{ github.event.inputs.duration || env.DURATION }}
+            - HTTP request          : avg=${{ env.HTTP_REQ_DURATION_AVG }}ms        p(90)=${{ env.HTTP_REQ_DURATION_P_90_ }}ms fails=${{ env.HTTP_REQ_FAILED_PASSES }}, finish reason: stop=${{ env.LLAMACPP_COMPLETIONS_STOP_RATE_PASSES }} truncated=${{ env.LLAMACPP_COMPLETIONS_TRUNCATED_RATE_PASSES }}
+            - Prompt processing (pp): avg=${{ env.LLAMACPP_PROMPT_TOKENS_AVG }}tk/s p(90)=${{ env.LLAMACPP_PROMPT_TOKENS_P_90_ }}tk/s **total=${{ env.LLAMACPP_PROMPT_TOKENS_TOTAL_COUNTER_RATE }}tk/s**
+            - Token generation  (tg): avg=${{ env.LLAMACPP_TOKENS_SECOND_AVG }}tk/s p(90)=${{ env.LLAMACPP_TOKENS_SECOND_P_90_ }}tk/s **total=${{ env.LLAMACPP_COMPLETION_TOKENS_TOTAL_COUNTER_RATE }}tk/s**
+            - ${{ env.BENCH_GRAPH_XLABEL }}
+
+            <details>
+
+            <summary>Time series</summary>
+
+            <p align="center">
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_O }}" alt="prompt_tokens_seconds" />
+
+            <details>
+
+            <summary>More</summary>
+
+            ```mermaid
+            ${{ env.PROMPT_TOKENS_SECONDS }}
+            ```
+
+            </details>
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_1 }}" alt="predicted_tokens_seconds"/>
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.PREDICTED_TOKENS_SECONDS }}
+            ```
+
+            </details>
+
+            </p>
+
+            <details>
+
+            <summary>Details</summary>
+
+            <p align="center">
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_2 }}" alt="kv_cache_usage_ratio" />
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.KV_CACHE_USAGE_RATIO }}
+            ```
+
+            </details>
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_3 }}" alt="requests_processing"/>
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.REQUESTS_PROCESSING }}
+            ```
+
+            </details>
+
+            </p>
+            </details>
+            </details>
@@ -840,7 +840,7 @@ 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' }}
@@ -4,6 +4,10 @@ name: Server
 on:
  workflow_dispatch: # allows manual triggering
    inputs:
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
      slow_tests:
        description: 'Run slow tests'
        required: true
@@ -11,12 +15,12 @@ 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 }}
@@ -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@v3
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: 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: |
@@ -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
@@ -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 $@ $<

@@ -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#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

@@ -18,12 +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
 - 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

 ----

@@ -115,6 +116,7 @@ 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)

 **Multimodal models:**
@@ -148,6 +150,7 @@ 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:**

@@ -174,6 +177,9 @@ 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)
+
+*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*

 ---

@@ -631,15 +637,6 @@ Building the program with BLAS support may lead to some performance improvements

 - #### Vulkan

-> [!WARNING]
->
-> Vulkan support has been broken in https://github.com/ggerganov/llama.cpp/pull/6122
-> due to relying on `GGML_OP_GET_ROWS` which is not yet properly supported by the Vulkan backend,
-> but should be fixed relatively soon (possibly in https://github.com/ggerganov/llama.cpp/pull/6155
-> (ref: https://github.com/ggerganov/llama.cpp/pull/6122#issuecomment-2015327635)).
->
-> Meanwhile, if you want to use the Vulkan backend, you should use the commit right before the breaking change, https://github.com/ggerganov/llama.cpp/commit/55c1b2a3bbd470e9e2a3a0618b92cf64a885f806
-
  **With docker**:

  You don't need to install Vulkan SDK. It will be installed inside the container.
@@ -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
@@ -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");
@@ -2489,7 +2489,7 @@ 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);
@@ -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])
@@ -372,12 +370,8 @@ 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 = []
@@ -779,6 +773,148 @@ 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
@@ -1058,12 +1194,120 @@ 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")
@@ -1080,6 +1324,92 @@ 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):
@@ -1099,7 +1429,7 @@ 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:
@@ -1700,11 +2030,8 @@ 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
@@ -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 + "'")
@@ -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
@@ -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;
@@ -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
@@ -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);

@@ -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) {
@@ -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
 ```
@@ -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;
    }

@@ -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

@@ -116,13 +116,13 @@ static void load_imatrix(const std::string & imatrix_file, std::unordered_map<st
    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;
+        exit(1);
    }
    int n_entries;
    in.read((char *)&n_entries, sizeof(n_entries));
    if (in.fail() || n_entries < 1) {
        printf("%s: no data in file %s\n", __func__, imatrix_file.c_str());
-        return;
+        exit(1);
    }
    for (int i = 0; i < n_entries; ++i) {
        int len; in.read((char *)&len, sizeof(len));
@@ -130,11 +130,11 @@ static void load_imatrix(const std::string & imatrix_file, std::unordered_map<st
        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;
+            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));
        int nval;
@@ -142,18 +142,22 @@ static void load_imatrix(const std::string & imatrix_file, std::unordered_map<st
        if (in.fail() || nval < 1) {
            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));
        if (in.fail()) {
            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());
 }
@@ -0,0 +1,303 @@
+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
+
+
+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()
+    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]
+                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 = {
+        "req": {
+            "p90": data['metrics']["http_req_duration"]["p(90)"],
+            "avg": data['metrics']["http_req_duration"]["avg"],
+        },
+        "pp": {
+            "p90": data['metrics']["llamacpp_prompt_tokens"]["p(90)"],
+            "avg": data['metrics']["llamacpp_prompt_tokens"]["avg"],
+        },
+        "tg": {
+            "p90": data['metrics']["llamacpp_tokens_second"]["p(90)"],
+            "avg": data['metrics']["llamacpp_tokens_second"]["avg"],
+        },
+    }
+    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()
@@ -0,0 +1,9 @@
+global:
+  scrape_interval:     10s
+  external_labels:
+    llamacpp: 'server'
+
+scrape_configs:
+  - job_name: 'llama.cpp server'
+    static_configs:
+      - targets: ['localhost:8080']
@@ -0,0 +1,2 @@
+matplotlib
+requests
@@ -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;
@@ -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;
@@ -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;
@@ -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;

@@ -405,7 +405,7 @@
        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) {
@@ -3566,6 +3566,7 @@ 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;
@@ -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()}")
@@ -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": {
@@ -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;
+            };
          };
      };
 }
@@ -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
@@ -377,13 +377,20 @@ 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");

-// 1.8125 bpw
+// 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)
-    uint8_t  scales[QK_K/32]; // 4-bit block scales
+#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 {
@@ -440,10 +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
@@ -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;
    }

@@ -1962,227 +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));
@@ -2192,8 +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]);
@@ -2213,6 +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;

@@ -2224,7 +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++) {
@@ -2389,7 +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;
@@ -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);
    }
@@ -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>
@@ -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
@@ -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,
@@ -1685,37 +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;
@@ -1723,7 +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
@@ -1732,7 +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
@@ -1745,7 +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;
@@ -1774,36 +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;
@@ -1813,7 +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);
@@ -1947,8 +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
@@ -1957,75 +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_IQ1_M || 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;
@@ -2432,6 +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) {
@@ -2441,11 +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:
                    {
@@ -3481,19 +3481,30 @@ void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, in
    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);
@@ -9756,11 +9767,17 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void

    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);

@@ -9784,7 +9801,9 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
        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;
@@ -9813,7 +9832,11 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
            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);
@@ -9823,14 +9846,22 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void

        }

+#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();
@@ -9842,7 +9873,9 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
        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();
@@ -9872,8 +9905,13 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void

            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);
@@ -9887,7 +9925,11 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
            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);

@@ -9907,7 +9949,9 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
        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) {
@@ -9927,15 +9971,24 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
                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;
@@ -11986,7 +12039,9 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy

    for (int ibl = 0; ibl < nbl; ++ibl) {

-        //y[ibl].d = GGML_FP32_TO_FP16(0.f);
+#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);
@@ -12161,13 +12216,22 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
        }

        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) {
@@ -12190,10 +12254,14 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
        }
        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
    }
 }

@@ -2968,7 +2968,7 @@ 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 +8079,7 @@ 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 +9956,14 @@ 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 +9982,14 @@ 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 +10008,14 @@ 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 +10034,14 @@ 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 +10060,14 @@ 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 +10400,8 @@ 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 +10409,7 @@ 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 +10424,8 @@ 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 +10433,7 @@ 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 +10448,8 @@ 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 +10457,7 @@ 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 +10472,8 @@ 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 +10481,7 @@ 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 +10496,8 @@ 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 +10505,7 @@ 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 +10520,8 @@ 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 +10529,7 @@ 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 +10544,8 @@ 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 +10553,7 @@ 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 +10568,8 @@ 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 +10577,7 @@ 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 +10592,8 @@ 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 +10601,7 @@ 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 +10616,8 @@ 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 +10625,13 @@ 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 +10641,11 @@ 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 +10668,10 @@ 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 +10694,10 @@ 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 +10720,10 @@ 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 +10746,10 @@ 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 +12792,7 @@ 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 +12850,9 @@ 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 +14913,9 @@ 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 +14946,38 @@ 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 +15063,10 @@ 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 +15966,7 @@ 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 +16001,7 @@ 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 +16016,7 @@ 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 +16368,8 @@ 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 +16739,7 @@ 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 +16812,7 @@ 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,
@@ -17155,6 +17109,7 @@ 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 +17136,7 @@ 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 +17149,21 @@ 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 +17174,9 @@ 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;
    }
@@ -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);
@@ -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
@@ -2717,8 +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();
@@ -2938,7 +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) {
@@ -4579,45 +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;
 }
@@ -10954,10 +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

@@ -10987,13 +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 :
@@ -11053,7 +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);
@@ -11088,9 +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;
@@ -11107,14 +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
@@ -16128,20 +16111,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
        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:
            {
@@ -18484,13 +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
@@ -18617,17 +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 = {
@@ -18684,10 +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;
@@ -1164,8 +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);
@@ -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"}))
@@ -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",
@@ -606,6 +608,22 @@ 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 +668,10 @@ 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,
+    ],
 }

 #
@@ -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
@@ -280,10 +278,8 @@ 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: (
@@ -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 = [
@@ -218,6 +218,7 @@ enum llm_arch {
    LLM_ARCH_GEMMA,
    LLM_ARCH_STARCODER2,
    LLM_ARCH_MAMBA,
+    LLM_ARCH_XVERSE,
    LLM_ARCH_COMMAND_R,
    LLM_ARCH_UNKNOWN,
 };
@@ -249,6 +250,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_GEMMA,           "gemma"      },
    { LLM_ARCH_STARCODER2,      "starcoder2" },
    { LLM_ARCH_MAMBA,           "mamba"      },
+    { LLM_ARCH_XVERSE,          "xverse"     },
    { LLM_ARCH_COMMAND_R,       "command-r"  },
    { LLM_ARCH_UNKNOWN,         "(unknown)"  },
 };
@@ -424,9 +426,12 @@ enum llm_tensor {
    LLM_TENSOR_FFN_DOWN,
    LLM_TENSOR_FFN_UP,
    LLM_TENSOR_FFN_ACT,
-    LLM_TENSOR_FFN_DOWN_EXP,
+    LLM_TENSOR_FFN_DOWN_EXP,  // split experts for backward compatibility
    LLM_TENSOR_FFN_GATE_EXP,
    LLM_TENSOR_FFN_UP_EXP,
+    LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
+    LLM_TENSOR_FFN_GATE_EXPS,
+    LLM_TENSOR_FFN_UP_EXPS,
    LLM_TENSOR_ATTN_Q_NORM,
    LLM_TENSOR_ATTN_K_NORM,
    LLM_TENSOR_LAYER_OUT_NORM,
@@ -461,6 +466,9 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
        },
    },
    {
@@ -514,6 +522,9 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
        },
@@ -878,6 +889,25 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
            { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
        },
    },
+    {
+        LLM_ARCH_XVERSE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
    {
        LLM_ARCH_COMMAND_R,
        {
@@ -1843,9 +1873,9 @@ struct llama_layer {

    // ff MoE
    struct ggml_tensor * ffn_gate_inp;
-    struct ggml_tensor * ffn_gate_exp[LLAMA_MAX_EXPERTS];
-    struct ggml_tensor * ffn_down_exp[LLAMA_MAX_EXPERTS];
-    struct ggml_tensor * ffn_up_exp  [LLAMA_MAX_EXPERTS];
+    struct ggml_tensor * ffn_gate_exps;
+    struct ggml_tensor * ffn_down_exps;
+    struct ggml_tensor * ffn_up_exps ;

    // ff bias
    struct ggml_tensor * ffn_down_b; // b2
@@ -2100,10 +2130,6 @@ struct llama_context {
            ggml_backend_free(backend);
        }

-#ifdef GGML_USE_VULKAN
-        ggml_vk_free_cpu_assist();
-#endif
-
        ggml_backend_buffer_free(buf_output);
    }

@@ -2851,19 +2877,19 @@ struct llama_model_loader {

    llama_mmaps mappings;

-    // Holds information on a model weights
-    struct llama_tensor_weights {
+    // Holds information on a model weight
+    struct llama_tensor_weight {
        uint16_t  idx; // source file index
        size_t   offs; // tensor data offset in the original file

        ggml_tensor * tensor;

-        llama_tensor_weights(uint16_t idx, const char * name, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) {
+        llama_tensor_weight(uint16_t idx, const char * name, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) {
            const int tensor_idx = gguf_find_tensor(gguf_ctx, name);
            offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx);
        }
    };
-    std::vector<llama_tensor_weights> weights;
+    std::vector<llama_tensor_weight> weights;

    std::unordered_map<std::string, struct llama_model_kv_override> kv_overrides;

@@ -2903,7 +2929,7 @@ struct llama_model_loader {
        // For subsidiary files, `meta` tensor data offset must not be used,
        // so we build a unified tensors index for weights.
        for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
-            weights.emplace_back(llama_tensor_weights(0, cur->name, meta, cur));
+            weights.emplace_back(0, cur->name, meta, cur);
        }
        files.emplace_back(new llama_file(fname.c_str(), "rb"));
        contexts.emplace_back(ctx);
@@ -2943,7 +2969,7 @@ struct llama_model_loader {

                // Save tensors data offset info of the shard.
                for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
-                    weights.emplace_back(llama_tensor_weights(idx, cur->name, ctx_gguf, cur));
+                    weights.emplace_back(idx, cur->name, ctx_gguf, cur);
                }
                files.emplace_back(new llama_file(split_path, "rb"));
                contexts.emplace_back(ctx);
@@ -3147,21 +3173,37 @@ struct llama_model_loader {
        return weights.at(i).tensor->name;
    }

-    const llama_tensor_weights & get_weights(const char * name) const {
+    const llama_tensor_weight * get_weight(const char * name) const {
        for (const auto & weight : weights) {
            if (strcmp(name, weight.tensor->name) == 0) {
-                return weight;
+                return &weight;
            }
        }
-        throw std::runtime_error(format("tensor %s not found", name));
+        return nullptr;
+    }
+
+    const llama_tensor_weight & require_weight(const char * name) const {
+        const llama_tensor_weight * weight = get_weight(name);
+        if (!weight) {
+            throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name));
+        }
+        return *weight;
    }

    struct ggml_tensor * get_tensor_meta(const char * name) const {
-        try {
-            return get_weights(name).tensor;
-        } catch (const std::runtime_error & e) {
-            return NULL;
+        const auto * weight = get_weight(name);
+        if (!weight) {
+            return nullptr;
        }
+        return weight->tensor;
+    }
+
+    struct ggml_tensor * require_tensor_meta(const char * name) const {
+        struct ggml_tensor * tensor = get_tensor_meta(name);
+        if (!tensor) {
+            throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name));
+        }
+        return tensor;
    }

    struct ggml_tensor * get_tensor_meta(int i) const {
@@ -3177,7 +3219,7 @@ struct llama_model_loader {
        return tensor;
    }

-    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, bool required = true) {
+    const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const {
        const struct ggml_tensor * cur = get_tensor_meta(name.c_str());

        if (cur == NULL) {
@@ -3189,8 +3231,8 @@ struct llama_model_loader {

        {
            bool is_ok = true;
-            for (size_t i = 0; i < ne.size(); ++i) {
-                if (ne[i] != cur->ne[i]) {
+            for (size_t i = 0; i < GGML_MAX_DIMS; ++i) {
+                if ((i < ne.size() && ne[i] != cur->ne[i]) || (i >= ne.size() && cur->ne[i] != 1)) {
                    is_ok = false;
                    break;
                }
@@ -3204,9 +3246,47 @@ struct llama_model_loader {
            }
        }

+        return cur;
+    }
+
+    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, bool required = true) {
+        const struct ggml_tensor * cur = check_tensor_dims(name, ne, required);
+
+        if (cur == NULL) {
+            return NULL;
+        }
+
        return create_tensor_for(ctx, cur);
    }

+    struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::vector<int64_t> & ne, size_t offset, bool required = true) {
+        const struct ggml_tensor * cur = check_tensor_dims(name, ne, required);
+
+        if (cur == NULL) {
+            return NULL;
+        }
+
+        if (cur->type != base->type) {
+            throw std::runtime_error(format("%s: tensor '%s' has wrong type; expected %s, got %s", __func__, name.c_str(), ggml_type_name(base->type), ggml_type_name(cur->type)));
+        }
+
+        std::array<int64_t, GGML_MAX_DIMS> dims;
+        for (size_t i = 0; i < GGML_MAX_DIMS; ++i) {
+            dims[i] = i < ne.size() ? ne[i] : 1;
+        }
+
+        struct ggml_tensor * tensor = ggml_view_4d(ctx, base,
+                                        dims[0], dims[1], dims[2], dims[3],
+                                        cur->nb[1], cur->nb[2], cur->nb[3],
+                                        offset);
+
+        ggml_set_name(tensor, name.c_str());
+
+        n_created++;
+
+        return tensor;
+    }
+
    void done_getting_tensors() const {
        if (n_created != n_tensors) {
            throw std::runtime_error(format("%s: wrong number of tensors; expected %d, got %d", __func__, n_tensors, n_created));
@@ -3219,7 +3299,7 @@ struct llama_model_loader {
            mmaps_used.reserve(files.size());
            for (const auto & file : files) {
                std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa()));
-                mmaps_used.emplace_back(std::make_pair(mapping->size, 0));
+                mmaps_used.emplace_back(mapping->size, 0);
                if (mlock_mmaps) {
                    std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());
                    mlock_mmap->init(mapping->addr);
@@ -3243,18 +3323,25 @@ struct llama_model_loader {
        *last  = 0;
        *addr = mapping->addr;
        for (ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor; tensor = ggml_get_next_tensor(ctx, tensor)) {
-            const auto & w = get_weights(ggml_get_name(tensor));
-            if (w.idx != idx) {
-                continue;
+            try {
+                const auto * weight = get_weight(ggml_get_name(tensor));
+                if (!weight) {
+                    continue;
+                }
+                if (weight->idx != idx) {
+                    continue;
+                }
+                *first = std::min(*first, weight->offs);
+                *last  = std::max(*last,  weight->offs + ggml_nbytes(tensor));
+            } catch(...) {
+                // the tensor is not in the model
            }
-            *first = std::min(*first, w.offs);
-            *last  = std::max(*last,  w.offs + ggml_nbytes(tensor));
        }
    }

    // for backwards compatibility, does not support ggml-backend
    void load_data_for(struct ggml_tensor * cur) const {
-        const auto & w = get_weights(ggml_get_name(cur));
+        const auto & w = require_weight(ggml_get_name(cur));

        if (use_mmap) {
            const auto & mapping = mappings.at(w.idx);
@@ -3287,44 +3374,49 @@ struct llama_model_loader {

        std::vector<no_init<uint8_t>> read_buf;
        for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) {
+            const auto * weight = get_weight(ggml_get_name(cur));
+            if (weight == nullptr) {
+                // this can happen with split experts models
+                continue;
+            }
+
            if (progress_callback) {
                if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) {
                    return false;
                }
            }

-            const auto & w = get_weights(ggml_get_name(cur));
            size_t n_size = ggml_nbytes(cur);

            if (use_mmap) {
-                const auto & mapping = mappings.at(w.idx);
+                const auto & mapping = mappings.at(weight->idx);
                ggml_backend_buffer_t buf_mmap = nullptr;
-                if (bufs_mmap.count(w.idx)) {
-                    buf_mmap = bufs_mmap.at(w.idx);
+                if (bufs_mmap.count(weight->idx)) {
+                    buf_mmap = bufs_mmap.at(weight->idx);
                }
                GGML_ASSERT(buf_mmap || cur->data); // either we have a buffer to allocate the tensor in, or it is already allocated
                if (buf_mmap && cur->data == nullptr) {
-                    ggml_backend_tensor_alloc(buf_mmap, cur, (uint8_t *) mapping->addr + w.offs);
+                    ggml_backend_tensor_alloc(buf_mmap, cur, (uint8_t *) mapping->addr + weight->offs);
                    if (lmlocks) {
-                        const auto & lmlock = lmlocks->at(w.idx);
-                        lmlock->grow_to(w.offs + ggml_nbytes(cur));
+                        const auto & lmlock = lmlocks->at(weight->idx);
+                        lmlock->grow_to(weight->offs + ggml_nbytes(cur));
                    }

-                    auto & mmap_used = mmaps_used[w.idx];
-                    mmap_used.first  = std::min(mmap_used.first,  w.offs);
-                    mmap_used.second = std::max(mmap_used.second, w.offs + n_size);
+                    auto & mmap_used = mmaps_used[weight->idx];
+                    mmap_used.first  = std::min(mmap_used.first,  weight->offs);
+                    mmap_used.second = std::max(mmap_used.second, weight->offs + n_size);
                } else {
-                    ggml_backend_tensor_set(cur, (uint8_t *) mapping->addr + w.offs, 0, n_size);
+                    ggml_backend_tensor_set(cur, (uint8_t *) mapping->addr + weight->offs, 0, n_size);
                }
            } else {
-                GGML_ASSERT(w.idx < files.size());
-                const auto & file = files.at(w.idx);
+                GGML_ASSERT(weight->idx < files.size());
+                const auto & file = files.at(weight->idx);
                if (ggml_backend_buffer_is_host(cur->buffer)) {
-                    file->seek(w.offs, SEEK_SET);
+                    file->seek(weight->offs, SEEK_SET);
                    file->read_raw(cur->data, ggml_nbytes(cur));
                } else {
                    read_buf.resize(ggml_nbytes(cur));
-                    file->seek(w.offs, SEEK_SET);
+                    file->seek(weight->offs, SEEK_SET);
                    file->read_raw(read_buf.data(), ggml_nbytes(cur));
                    ggml_backend_tensor_set(cur, read_buf.data(), 0, n_size);
                }
@@ -3847,6 +3939,16 @@ static void llm_load_hparams(
                    default: model.type = e_model::MODEL_UNKNOWN;
                }
            } break;
+        case LLM_ARCH_XVERSE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 40: model.type = e_model::MODEL_13B; break;
+                    case 80: model.type = e_model::MODEL_65B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
        case LLM_ARCH_COMMAND_R:
            {
                ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
@@ -4243,6 +4345,7 @@ static bool llm_load_tensors(

    const int64_t n_layer     = hparams.n_layer;
    const int64_t i_gpu_start = std::max((int64_t) hparams.n_layer - n_gpu_layers, (int64_t) 0);
+    bool use_mmap_buffer = true;

    // there is very little benefit to offloading the input layer, so always keep it on the CPU
    model.buft_input = llama_default_buffer_type_cpu(true);
@@ -4331,6 +4434,10 @@ static bool llm_load_tensors(

    // create one context per buffer type
    size_t ctx_size = ggml_tensor_overhead()*(ml.n_tensors + 1); // +1 for models where tok_embd is duplicated as output
+
+    // for moe merged tensors
+    ctx_size += ggml_tensor_overhead()*hparams.n_expert*n_layer;
+
    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
    for (auto & it : buft_layer_count) {
        struct ggml_init_params params = {
@@ -4357,6 +4464,11 @@ static bool llm_load_tensors(
        const int64_t n_vocab      = hparams.n_vocab;
        const int64_t n_vocab_type = hparams.n_vocab_type;
        const int64_t n_ff         = hparams.n_ff;
+        const int64_t n_expert     = hparams.n_expert;
+
+        if (n_expert > 0 && hparams.n_expert_used == 0) {
+            throw std::runtime_error("model has expert layers but no expert layers are used");
+        }

        GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);

@@ -4411,30 +4523,50 @@ static bool llm_load_tensors(

                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});

-                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd}, false);
-
-                        if (layer.ffn_gate_inp == nullptr) {
-                            GGML_ASSERT(hparams.n_expert      == 0);
-                            GGML_ASSERT(hparams.n_expert_used == 0);
-
+                        if (n_expert == 0) {
                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
                        } else {
-                            GGML_ASSERT(hparams.n_expert      > 0);
-                            GGML_ASSERT(hparams.n_expert_used > 0);
+                            layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});

-                            // MoE branch
-                            for (uint32_t x = 0; x < hparams.n_expert; ++x) {
-                                layer.ffn_gate_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), {n_embd,   n_ff});
-                                layer.ffn_down_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), {  n_ff, n_embd});
-                                layer.ffn_up_exp[x]   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, x), {n_embd,   n_ff});
+                            layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
+                            if (layer.ffn_gate_exps) {
+                                layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
+                                layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
+                            } else {
+                                // merge split expert into a single tensor for compatibility with older models
+                                // requires disabling mmap
+                                use_mmap_buffer = false;
+
+                                ggml_type type_gate = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, 0).c_str())->type;
+                                ggml_type type_down = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, 0).c_str())->type;
+                                ggml_type type_up   = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, 0).c_str())->type;
+
+                                layer.ffn_gate_exps = ggml_new_tensor_3d(ctx_split, type_gate, n_embd,   n_ff, n_expert);
+                                layer.ffn_down_exps = ggml_new_tensor_3d(ctx_split, type_down,   n_ff, n_embd, n_expert);
+                                layer.ffn_up_exps   = ggml_new_tensor_3d(ctx_split, type_up,   n_embd,   n_ff, n_expert);
+
+                                ggml_set_name(layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i).c_str());
+                                ggml_set_name(layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i).c_str());
+                                ggml_set_name(layer.ffn_up_exps,   tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i).c_str());
+
+                                for (uint32_t x = 0; x < n_expert; ++x) {
+                                    // the individual experts are loaded into a view of the merged tensor
+                                    ml.create_tensor_as_view(ctx_split, layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), { n_embd, n_ff }, layer.ffn_gate_exps->nb[2]*x);
+                                    ml.create_tensor_as_view(ctx_split, layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), { n_ff, n_embd }, layer.ffn_down_exps->nb[2]*x);
+                                    ml.create_tensor_as_view(ctx_split, layer.ffn_up_exps,   tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, x), { n_embd, n_ff }, layer.ffn_up_exps->nb[2]*x);
+                                }
                            }
                        }
                    }
                } break;
            case LLM_ARCH_GROK:
                {
+                    if (n_expert == 0) {
+                        throw std::runtime_error("Grok model cannot have zero experts");
+                    }
+
                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});

                    // output
@@ -4466,16 +4598,35 @@ static bool llm_load_tensors(

                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});

-                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd});
+                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});

-                        GGML_ASSERT(hparams.n_expert      > 0);
-                        GGML_ASSERT(hparams.n_expert_used > 0);
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, false);
+                        if (layer.ffn_gate_exps) {
+                            layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
+                            layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
+                        } else {
+                            // merge split expert into a single tensor for compatibility with older models
+                            // requires disabling mmap
+                            use_mmap_buffer = false;

-                        // MoE branch
-                        for (uint32_t x = 0; x < hparams.n_expert; ++x) {
-                            layer.ffn_gate_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), {n_embd,   n_ff});
-                            layer.ffn_down_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), {  n_ff, n_embd});
-                            layer.ffn_up_exp[x]   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, x), {n_embd,   n_ff});
+                            ggml_type type_gate = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, 0).c_str())->type;
+                            ggml_type type_down = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, 0).c_str())->type;
+                            ggml_type type_up   = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, 0).c_str())->type;
+
+                            layer.ffn_gate_exps = ggml_new_tensor_3d(ctx_split, type_gate, n_embd,   n_ff, n_expert);
+                            layer.ffn_down_exps = ggml_new_tensor_3d(ctx_split, type_down,   n_ff, n_embd, n_expert);
+                            layer.ffn_up_exps   = ggml_new_tensor_3d(ctx_split, type_up,   n_embd,   n_ff, n_expert);
+
+                            ggml_set_name(layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i).c_str());
+                            ggml_set_name(layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i).c_str());
+                            ggml_set_name(layer.ffn_up_exps,   tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i).c_str());
+
+                            for (uint32_t x = 0; x < n_expert; ++x) {
+                                // the individual experts are loaded into a view of the merged tensor
+                                ml.create_tensor_as_view(ctx_split, layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), { n_embd, n_ff }, layer.ffn_gate_exps->nb[2]*x);
+                                ml.create_tensor_as_view(ctx_split, layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), { n_ff, n_embd }, layer.ffn_down_exps->nb[2]*x);
+                                ml.create_tensor_as_view(ctx_split, layer.ffn_up_exps,   tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, x), { n_embd, n_ff }, layer.ffn_up_exps->nb[2]*x);
+                            }
                        }

                        layer.layer_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd});
@@ -5200,6 +5351,28 @@ static bool llm_load_tensors(
                        layer.ssm_out = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd});
                    }
                } break;
+            case LLM_ARCH_XVERSE:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+                        auto & layer = model.layers[i];
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
            case LLM_ARCH_COMMAND_R:
                {
                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@@ -5238,7 +5411,7 @@ static bool llm_load_tensors(

    ml.done_getting_tensors();

-    ml.init_mappings(true, &model.mlock_mmaps);
+    ml.init_mappings(true, use_mlock ? &model.mlock_mmaps : nullptr);
    model.mappings.reserve(ml.mappings.size());

    // create the backend buffers
@@ -5259,7 +5432,7 @@ static bool llm_load_tensors(
        // only the mmap region containing the tensors in the model is mapped to the backend buffer
        // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
        // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
-        if (ml.use_mmap && buft == llama_default_buffer_type_cpu(true)) {
+        if (ml.use_mmap && use_mmap_buffer && buft == llama_default_buffer_type_cpu(true)) {
            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                void * addr = nullptr;
                size_t first, last;
@@ -5283,7 +5456,7 @@ static bool llm_load_tensors(
            }
        }
 #ifdef GGML_USE_METAL
-        else if (ml.use_mmap && buft == ggml_backend_metal_buffer_type()) {
+        else if (ml.use_mmap && use_mmap_buffer && buft == ggml_backend_metal_buffer_type()) {
            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                const size_t max_size = ggml_get_max_tensor_size(ctx);
                void * addr = nullptr;
@@ -5366,8 +5539,10 @@ static bool llm_load_tensors(
        }
    }

-    for (auto & mapping : ml.mappings) {
-        model.mappings.emplace_back(std::move(mapping));
+    if (use_mmap_buffer) {
+        for (auto & mapping : ml.mappings) {
+            model.mappings.emplace_back(std::move(mapping));
+        }
    }

    // loading time will be recalculate after the first eval, so
@@ -5523,8 +5698,8 @@ static void llm_build_kv_store(
    GGML_ASSERT(kv.size == n_ctx);

    // compute the transposed [n_tokens, n_embd] V matrix
-    struct ggml_tensor * v_cur_t = ggml_transpose(ctx, ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens));
-    //struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); // TODO: reshape above is likely not needed
+    assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens);
+    struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur);
    cb(v_cur_t, "v_cur_t", il);

    struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_k_gqa,
@@ -6235,19 +6410,19 @@ struct llm_build_context {
                for (int i = 0; i < n_expert_used; ++i) {
                    ggml_tensor * cur_expert;

-                    ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exp, n_expert, selected_experts, i, cur);
+                    ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur);
                    cb(cur_up, "ffn_moe_up", il);

-                    ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exp, n_expert, selected_experts, i, cur);
+                    ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur);
                    cb(cur_gate, "ffn_moe_gate", il);

                    cur_gate = ggml_silu(ctx0, cur_gate);
                    cb(cur_gate, "ffn_moe_silu", il);

-                    cur_expert = ggml_mul(ctx0, cur_up, cur_gate); // [n_tokens, n_embd]
+                    cur_expert = ggml_mul(ctx0, cur_up, cur_gate);
                    cb(cur_expert, "ffn_moe_gate_par", il);

-                    cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exp, n_expert, selected_experts, i, cur_expert); // [n_tokens, n_embd]
+                    cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd]
                    cb(cur_expert, "ffn_moe_down", il);

                    cur_expert = ggml_mul(ctx0, cur_expert,
@@ -6411,6 +6586,111 @@ struct llm_build_context {
        return gf;
    }

+    struct ggml_cgraph * build_xverse() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        // positions of the tokens in the KV cache
+        struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,      cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   NULL,
+                        model.layers[il].ffn_gate, NULL,
+                        model.layers[il].ffn_down, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
    struct ggml_cgraph * build_falcon() {
        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);

@@ -6664,20 +6944,20 @@ struct llm_build_context {
            for (int i = 0; i < n_expert_used; ++i) {
                ggml_tensor * cur_expert;

-                ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exp, n_expert, selected_experts, i, cur);
+                ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur);
                cb(cur_up, "ffn_moe_up", il);

-                ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exp, n_expert, selected_experts, i, cur);
+                ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur);
                cb(cur_gate, "ffn_moe_gate", il);

                //GeLU
                cur_gate = ggml_gelu(ctx0, cur_gate);
                cb(cur_gate, "ffn_moe_gelu", il);

-                cur_expert = ggml_mul(ctx0, cur_up, cur_gate); // [n_tokens, n_embd]
+                cur_expert = ggml_mul(ctx0, cur_up, cur_gate);
                cb(cur_expert, "ffn_moe_gate_par", il);

-                cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exp, n_expert, selected_experts, i, cur_expert); // [n_tokens, n_embd]
+                cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd]
                cb(cur_expert, "ffn_moe_down", il);

                cur_expert = ggml_mul(ctx0, cur_expert,
@@ -9152,8 +9432,9 @@ struct llm_build_context {
            if (il == n_layer - 1) {
                // skip computing output for unused tokens
                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+                cur     = ggml_get_rows(ctx0,     cur, inp_out_ids);
+                inpL    = ggml_get_rows(ctx0,    inpL, inp_out_ids);
+                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
            }

            struct ggml_tensor * attn_out = cur;
@@ -9388,6 +9669,10 @@ static struct ggml_cgraph * llama_build_graph(
            {
                result = llm.build_mamba();
            } break;
+        case LLM_ARCH_XVERSE:
+            {
+                result = llm.build_xverse();
+            } break;
        case LLM_ARCH_COMMAND_R:
            {
                result = llm.build_command_r();
@@ -12743,7 +13028,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
            // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
            // for getting the current layer as I initially thought, and we need to resort to parsing the
            // tensor name.
-            n_layer /= n_expert;
            if (sscanf(name, "blk.%d.", &i_layer) != 1) {
                throw std::runtime_error(format("Failed to determine layer for tensor %s", name));
            }
@@ -13105,7 +13389,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
        kv_overrides = v->data();
    }
    llama_model_loader ml(fname_inp, use_mmap, kv_overrides);
-    ml.init_mappings(false); // no prefetching?
+    ml.init_mappings(false); // no prefetching

    llama_model model;
    llm_load_arch(ml, model);
@@ -13157,20 +13441,15 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
        // TODO: avoid hardcoded tensor names - use the TN_* constants
        if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) {
            ++qs.n_attention_wv;
-        } else if (name.find("ffn_down") != std::string::npos) {
-            ++qs.n_ffn_down;
-        } else if (name.find("ffn_gate") != std::string::npos) {
-            ++qs.n_ffn_gate;
-        } else if (name.find("ffn_up") != std::string::npos) {
-            ++qs.n_ffn_up;
        } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
            qs.has_output = true;
        }
    }
-    if (qs.n_attention_wv != qs.n_ffn_down || (uint32_t) qs.n_attention_wv != model.hparams.n_layer) {
-        LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_ffn_down = %d, hparams.n_layer = %d\n",
-                __func__, qs.n_attention_wv, qs.n_ffn_down, model.hparams.n_layer);
-    }
+
+    qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
+
+    // sanity checks
+    GGML_ASSERT(qs.n_attention_wv == (int)model.hparams.n_layer && "n_attention_wv != n_layer is unexpected");

    size_t total_size_org = 0;
    size_t total_size_new = 0;
@@ -13200,6 +13479,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
    // placeholder for the meta data
    ::zeros(fout, meta_size);

+    const auto tn = LLM_TN(model.arch);
+
    for (int i = 0; i < ml.n_tensors; ++i) {
        struct ggml_tensor * tensor = ml.get_tensor_meta(i);

@@ -13222,8 +13503,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
        // This used to be a regex, but <regex> has an extreme cost to compile times.
        bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'?

-        // quantize only 2D tensors
-        quantize &= (ggml_n_dims(tensor) == 2);
+        // quantize only 2D and 3D tensors (experts)
+        quantize &= (ggml_n_dims(tensor) >= 2);
        quantize &= params->quantize_output_tensor || name != "output.weight";
        quantize &= !params->only_copy;

@@ -13278,11 +13559,20 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                if (it == imatrix_data->end()) {
                    LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
                } else {
-                    if (it->second.size() == (size_t)tensor->ne[0]) {
+                    if (it->second.size() == (size_t)tensor->ne[0]*tensor->ne[2]) {
                        imatrix = it->second.data();
                    } else {
                        LLAMA_LOG_INFO("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
-                                int(it->second.size()), int(tensor->ne[0]), tensor->name);
+                                int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name);
+
+                        // this can happen when quantizing an old mixtral model with split tensors with a new incompatible imatrix
+                        // this is a significant error and it may be good idea to abort the process if this happens,
+                        // since many people will miss the error and not realize that most of the model is being quantized without an imatrix
+                        // tok_embd should be ignored in this case, since it always causes this warning
+                        if (name != tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
+                            throw std::runtime_error(format("imatrix size %d is different from tensor size %d for %s",
+                                    int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name));
+                        }
                    }
                }
            }
@@ -13319,15 +13609,24 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
            new_data = work.data();

            const int n_per_row = tensor->ne[0];
-            const int nrows = nelements / n_per_row;
+            const int nrows = tensor->ne[1];

            static const int min_chunk_size = 32 * 512;
            const int chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row);

-            const int nchunk = (nelements + chunk_size - 1)/chunk_size;
+            const int nelements_matrix = tensor->ne[0] * tensor->ne[1];
+            const int nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
            const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1;
-            new_size = llama_tensor_quantize_internal(new_type, f32_data, new_data, chunk_size, nrows, n_per_row, imatrix, workers, nthread_use);

+            // quantize each expert separately since they have different importance matrices
+            new_size = 0;
+            for (int64_t i03 = 0; i03 < tensor->ne[2]; ++i03) {
+                const float * f32_data_03 = f32_data + i03 * nelements_matrix;
+                void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
+                const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
+
+                new_size += llama_tensor_quantize_internal(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
+            }
            LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
        }
        total_size_org += ggml_nbytes(tensor);
@@ -13968,7 +14267,20 @@ struct llama_context * llama_new_context_with_model(
            }
        }
 #elif defined(GGML_USE_VULKAN)
-        if (model->n_gpu_layers > 0) {
+        if (model->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            LLAMA_LOG_ERROR("%s: Row split not supported. Failed to initialize Vulkan backend\n", __func__);
+            llama_free(ctx);
+            return nullptr;
+        }
+        if (model->split_mode == LLAMA_SPLIT_MODE_NONE) {
+            ggml_backend_t backend = ggml_backend_vk_init(0);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        } else {
            for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
                ggml_backend_t backend = ggml_backend_vk_init(device);
                if (backend == nullptr) {
@@ -14187,6 +14499,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
        case LLM_ARCH_ORION:
        case LLM_ARCH_INTERNLM2:
        case LLM_ARCH_MINICPM:
+        case LLM_ARCH_XVERSE:
        case LLM_ARCH_COMMAND_R:
            return LLAMA_ROPE_TYPE_NORM;

@@ -15524,6 +15837,55 @@ static int32_t llama_chat_apply_template_internal(
                ss << message->content << "</s>";
            }
        }
+    } else if (tmpl == "openchat" || tmpl.find("GPT4 Correct ") != std::string::npos) {
+        // openchat/openchat-3.5-0106,
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "<|end_of_turn|>";
+            } else {
+                role[0] = toupper(role[0]);
+                ss << "GPT4 Correct " << role << ": " << message->content << "<|end_of_turn|>";
+            }
+        }
+        if (add_ass) {
+            ss << "GPT4 Correct Assistant:";
+        }
+    } else if (tmpl == "vicuna" || tmpl == "vicuna-orca" || (tmpl.find("USER: ") != std::string::npos && tmpl.find("ASSISTANT: ") != std::string::npos)) {
+        // eachadea/vicuna-13b-1.1 (and Orca variant)
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                // Orca-Vicuna variant uses a system prefix
+                if (tmpl == "vicuna-orca" || tmpl.find("SYSTEM: ") != std::string::npos) {
+                    ss << "SYSTEM: " << message->content << "\n";
+                } else {
+                    ss << message->content << "\n\n";
+                }
+            } else if (role == "user") {
+                ss << "USER: " << message->content << "\n";
+            } else if (role == "assistant") {
+                ss << "ASSISTANT: " << message->content << "</s>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "ASSISTANT:";
+        }
+    } else if (tmpl == "deepseek" || (tmpl.find("### Instruction:") != std::string::npos && tmpl.find("<|EOT|>") != std::string::npos)) {
+        // deepseek-ai/deepseek-coder-33b-instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content;
+            } else if (role == "user") {
+                ss << "### Instruction:\n" << message->content << "\n";
+            } else if (role == "assistant") {
+                ss << "### Response:\n" << message->content << "\n<|EOT|>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "### Response:\n";
+        }
    } else {
        // template not supported
        return -1;
@@ -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
@@ -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:
@@ -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' \
@@ -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
@@ -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,
@@ -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));
@@ -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;
Author	SHA1	Message	Date
kaizau	1ff4d9f3d6	Add OpenChat, Alpaca, Vicuna chat templates (#6397 ) * Add openchat chat template * Add chat template test for openchat * Add chat template for vicuna * Add chat template for orca-vicuna * Add EOS for vicuna templates * Combine vicuna chat templates * Add tests for openchat and vicuna chat templates * Add chat template for alpaca * Add separate template name for vicuna-orca * Remove alpaca, match deepseek with jinja output * Regenerate chat template test with add_generation_prompt * Separate deepseek bos from system message * Match openchat template with jinja output * Remove BOS token from templates, unprefix openchat	2024-04-03 17:24:31 +02:00
Georgi Gerganov	076b08649e	readme : update hot topics	2024-04-03 16:11:15 +03:00
slaren	08a0c02060	ggml : mul_mat_id use the same tensor for all the experts (#6387 ) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-04-03 16:07:05 +03:00
Meng, Hengyu	52604860f9	[SYCL] Disable iqx on windows as WA (#6435 ) * disable iqx on windows as WA * array instead of global_memory	2024-04-03 10:34:40 +08:00
Georgi Gerganov	f87f7b8986	flake.lock: Update (#6402 ) Flake lock file updates: • Updated input 'nixpkgs': 'github:NixOS/nixpkgs/44d0940ea560dee511026a53f0e2e2cde489b4d4' (2024-03-23) → 'github:NixOS/nixpkgs/d8fe5e6c92d0d190646fb9f1056741a229980089' (2024-03-29) Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>	2024-04-01 09:05:57 -07:00
Johannes Gäßler	33a5244806	compare-llama-bench.py: fix long hexsha args (#6424 )	2024-04-01 13:30:43 +02:00
Pierrick Hymbert	226e819371	ci: server: verify deps are coherent with the commit (#6409 ) * ci: server: verify deps are coherent with the commit * ci: server: change the ref to build as now it's a pull event target	2024-04-01 12:36:40 +02:00
Georgi Gerganov	c50a82ce0f	readme : update hot topics	2024-03-31 11:56:30 +03:00
Pierrick Hymbert	37e7854c10	ci: bench: fix Resource not accessible by integration on PR event (#6393 )	2024-03-30 12:36:07 +02:00
Mohammadreza Hendiani	c342d070c6	Fedora build update (#6388 ) * fixed deprecated address * fixed deprecated address * fixed deprecated address * Added 'Apache-2.0' SPDX license identifier due to 'kompute.cc' submodule licensing. Explanation of licensing method: https://docs.fedoraproject.org/en-US/legal/spdx/#_and_expressions * Added 'Apache-2.0' SPDX license identifier due to 'kompute.cc' submodule licensing. Explanation of licensing method: https://docs.fedoraproject.org/en-US/legal/spdx/#_and_expressions * Added 'Apache-2.0' SPDX license identifier due to 'kompute.cc' submodule licensing. Explanation of licensing method: https://docs.fedoraproject.org/en-US/legal/spdx/#_and_expressions * reverted back to only the MIT license	2024-03-29 22:59:56 +01:00
Xuan Son Nguyen	f7fc5f6c6f	split: allow --split-max-size option (#6343 ) * split by max size * clean up arg parse * split: ok * add dry run option * error on 0 tensors * be positive * remove next_metadata_size	2024-03-29 22:34:44 +01:00
0cc4m	ba0c7c70ab	Vulkan k-quant mmq and ggml-backend offload functionality (#6155 ) * Fix Vulkan no kv offload incoherence * Add k-quant mul mat mat shaders * Rework working buffer allocation, reduces vram use noticeably Clean up cpu assist code, replaced with ggml-backend offload function * Default to all dedicated GPUs * Add fallback for integrated GPUs if no dedicated GPUs are found * Add debug info which device is allocating memory * Fix Intel dequant issue Fix validation issue * Fix Vulkan GGML_OP_GET_ROWS implementation * Clean up merge artifacts * Remove Vulkan warning	2024-03-29 17:29:21 +01:00
Georgi Gerganov	d48ccf3ad4	sync : ggml (#6351 ) * sync : ggml ggml-ci * cuda : move GGML_CUDA_DMMV constants to dmmv.cuh --------- Co-authored-by: slaren <slarengh@gmail.com>	2024-03-29 17:45:46 +02:00
hxer7963	069574775c	[Model] Add support for xverse (#6301 ) * Support xverse model convert to gguf format. * 1. Convert xverse models to gguf; 2. Add LLM_ARCH_XVERSE inference in llama.cpp; 3. Add xverse item in Supported models in README.md; * * gguf-py: remove redundant logs * llama: remove the init_mapping_prefetch custom parameter * llama.cpp: Include the changes from #6122 to exclude the unused outputs of the last layers. * - Fix format issues - Remove duplicate set kqv_out to llm_build_kv * Update llama.cpp --------- Co-authored-by: willhe <willhe@xverse.cn> Co-authored-by: willhe <hexin@xverse.cn>	2024-03-29 14:37:03 +01:00
Georgi Gerganov	cfde806eb9	ci : fix BGE wget (#6383 ) ggml-ci	2024-03-29 14:34:28 +02:00
zhouwg	b910287954	readme : add project (#6356 ) * readme: add Android UI binding * Update README.md	2024-03-29 09:33:46 +02:00
Matt Clayton	8093987090	cmake : add explicit metal version options (#6370 ) * cmake: add explicit metal version options * Update CMakeLists.txt --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-29 09:27:42 +02:00
Daniel Bevenius	057400a3fd	llama : remove redundant reshape in build_kv_store (#6369 ) * llama: remove redundant reshape in build_kv_store This commit removes the reshape of the V matrix in the build_kv_store. The motivation for this is that V matrix has the shape: ```console (gdb) p v_cur $46 = {type = GGML_TYPE_F32, backend = GGML_BACKEND_TYPE_CPU, buffer = 0x0, ne = {4096, 512, 1, 1}, nb = {4, 16384, 8388608, 8388608}, op = GGML_OP_MUL_MAT, op_params = { 0 <repeats 16 times>}, flags = 0, grad = 0x0, src = {0xb496b0, 0x7ffef1c40950, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, perf_runs = 0, perf_cycles = 0, perf_time_us = 0, view_src = 0x0, view_offs = 0, data = 0x0, name = "Vcur-0", '\000' <repeats 57 times>, extra = 0x0, padding = "\000\000\000\000\000\000\000"} ``` And after reshaping this tensor we get: ```console gdb) p ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens) $44 = {type = GGML_TYPE_F32, backend = GGML_BACKEND_TYPE_CPU, buffer = 0x0, ne = {4096, 512, 1, 1}, nb = {4, 16384, 8388608, 8388608}, op = GGML_OP_RESHAPE, op_params = { 0 <repeats 16 times>}, flags = 0, grad = 0x0, src = {0x7ffef1c40e00, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, perf_runs = 0, perf_cycles = 0, perf_time_us = 0, view_src = 0x7ffef1c40e00, view_offs = 0, data = 0x0, name = "Vcur-0 (reshaped)", '\000' <repeats 46 times>, extra = 0x0, padding = "\000\000\000\000\000\000\000"} ``` I noticed that the `src` and `view_src` fields are different but that the dimensions are the same. From the code comment it seems like the reshape call is not needed and perhaps the above can motivate the removal of the reshape call. Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> * llama : add assert --------- Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-29 09:23:22 +02:00
Pedro Cuenca	b75c38166c	convert : allow conversion of Mistral HF models (#6144 ) * Allow conversion of Mistral HF models * Homogenize Llama, Mistral, Mixtral under the same entry. * Fix tokenizer, permute tensors * Use sentencepiece tokenizer, or fall back to hfft. * convert-hf : small fix for mypy * convert-hf : fix duplicated block_count * convert-hf : add vocab size to metadata --------- Co-authored-by: Jared Van Bortel <jared@nomic.ai>	2024-03-29 09:15:00 +02:00
Georgi Gerganov	bfe7dafc9c	readme : add notice for UI list	2024-03-28 22:56:03 +02:00
Ouadie EL FAROUKI	5106ef482c	[SYCL] Revisited & updated SYCL build documentation (#6141 ) * Revisited & updated SYCL build documentation * removed outdated comment * Addressed PR comments * Trimed white spaces * added new end line	2024-03-28 16:01:47 +00:00
Jared Van Bortel	be55134a53	convert : refactor vocab selection logic (#6355 )	2024-03-28 11:44:36 -04:00
Ziang Wu	66ba560256	llava : fix MobileVLM (#6364 ) * fix empty bug * Update MobileVLM-README.md added more results on devices * Update MobileVLM-README.md * Update MobileVLM-README.md * Update MobileVLM-README.md * Update MobileVLM-README.md * Update MobileVLM-README.md * Update MobileVLM-README.md * Update examples/llava/MobileVLM-README.md Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update MobileVLM-README.md remove gguf links --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-28 16:33:10 +02:00
compilade	0308f5e3d7	llama : fix command-r inference when omitting outputs (#6367 )	2024-03-28 14:05:54 +02:00
Pierrick Hymbert	28cb9a09c4	ci: bench: fix master not schedule, fix commit status failed on external repo (#6365 )	2024-03-28 11:27:56 +01:00
Ting Sun	cfc4d75df6	doc: fix outdated default value of batch size (#6336 ) * doc: fix outdated default value of batch size * doc: add doc for ubatch-size	2024-03-28 09:51:06 +01:00
Eric Zhang	6902cb7f2e	server : stop gracefully on SIGTERM (#6348 )	2024-03-28 09:50:48 +01:00
hutli	d2d8f38996	nix: removed unnessesary indentation	2024-03-28 07:48:27 +00:00
hutli	d39b308eaf	nix: moved blas availability check to package inputs so it is still overridable	2024-03-28 07:48:27 +00:00
hutli	c873976649	using blas.meta.available to check host platform	2024-03-28 07:48:27 +00:00
hutli	dbb03e2b9c	only using explicit blas if hostPlatform is allowed	2024-03-28 07:48:27 +00:00
Someone Serge	e9f17dc3bf	nix: .#windows: proper cross-compilation set-up Take all dependencies from the cross stage, rather tha only stdenv	2024-03-28 07:48:27 +00:00
Someone Serge	22a462cc1f	nix: package: don't introduce the dependency on python - The generic /usr/bin/env shebangs are good enough - Python deps are provisioned in the devShells - We need to be able to leave python out at least on windows (currently breaks eval)	2024-03-28 07:48:27 +00:00
hutli	f6a0f5c642	nix: .#widnows: init initial nix build for windows using zig mingwW64 build removes nix zig windows build removes nix zig windows build removed unnessesary glibc.static removed unnessesary import of pkgs in nix fixed missing trailing newline on non-windows nix builds overriding stdenv when building for crosscompiling to windows in nix better variables when crosscompiling windows in nix cross compile windows on macos removed trailing whitespace remove unnessesary overwrite of "CMAKE_SYSTEM_NAME" in nix windows build nix: keep file extension when copying result files during cross compile for windows nix: better checking for file extensions when using MinGW nix: using hostPlatform instead of targetPlatform when cross compiling for Windows using hostPlatform.extensions.executable to extract executable format	2024-03-28 07:48:27 +00:00
Ziang Wu	d0e2f6416b	doc: fix typo in MobileVLM-README.md (#6181 )	2024-03-28 13:03:30 +09:00
Neo Zhang Jianyu	25f4a613c4	[SYCL] fix set main gpu crash (#6339 )	2024-03-28 08:55:24 +08:00
Pierrick Hymbert	a016026a3a	server: continuous performance monitoring and PR comment (#6283 ) * server: bench: init * server: bench: reduce list of GPU nodes * server: bench: fix graph, fix output artifact * ci: bench: add mermaid in case of image cannot be uploaded * ci: bench: more resilient, more metrics * ci: bench: trigger build * ci: bench: fix duration * ci: bench: fix typo * ci: bench: fix mermaid values, markdown generated * typo on the step name Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com> * ci: bench: trailing spaces * ci: bench: move images in a details section * ci: bench: reduce bullet point size --------- Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>	2024-03-27 20:26:49 +01:00
Someone Serge	53c7ec53d5	nix: ci: dont test cuda and rocm (for now) Until https://github.com/ggerganov/llama.cpp/issues/6346 is resolved	2024-03-27 19:18:55 +00:00
slaren	e5b89a441a	ggml : fix bounds checking of zero size views (#6347 )	2024-03-27 15:07:50 +01:00
Georgi Gerganov	3a0345970e	make : whitespace	2024-03-27 15:02:49 +02:00
howlger	1e13987fba	embedding : show full embedding for single prompt (#6342 ) * embedding : show full embedding for single prompt To support the use case of creating an embedding for a given prompt, the entire embedding and not just the first part needed to be printed. Also, show cosine similarity matrix only if there is more than one prompt, as the cosine similarity matrix for a single prompt is always `1.00`. * Update examples/embedding/embedding.cpp --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-27 13:15:44 +02:00
AidanBeltonS	e82f9e2b83	[SYCL] Fix batched impl for NVidia GPU (#6164 ) * Fix batched impl * Maintain previous behaviour for igpu * retrigger CI --------- Co-authored-by: Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>	2024-03-27 13:46:40 +05:30
Kawrakow	cbc8343619	Make IQ1_M work for QK_K = 64 (#6327 ) * iq1_m: make it work for QK_K = 64 (WIP) * iq1_m: make it work for QK_K = 64 (scalar and AVX2) * iq1_m: QK_K = 64 seems to work on Metal and ARM_NEON --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>	2024-03-27 08:44:27 +01:00
Sigbjørn Skjæret	e562b9714b	common : change --no-penalize-nl to --penalize-nl (#6334 ) * Change --no-penalize-nl to --penalize-nl * Update documentation too	2024-03-27 09:23:10 +02:00
Georgi Gerganov	2ab4f00d25	llama2c : open file as binary (#6332 )	2024-03-27 09:16:02 +02:00
Mateusz Charytoniuk	1740d6dd4e	readme : add php api bindings (#6326 ) * add php bindings to readme * readme : add link to PR --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-27 09:08:59 +02:00
Eric Zhang	0642b22cd1	server: public: use relative routes for static files (#6325 ) server: public: support custom `api_url`, default to relative base path	2024-03-27 06:55:29 +01:00
Neo Zhang Jianyu	a4f569e8a3	[SYCL] fix no file in win rel (#6314 )	2024-03-27 09:47:06 +08:00