ci : multithreaded builds (#3311 )

* mac and linux threads * windows * Update build.yml * Update build.yml * Update build.yml * automatically get thread count * windows syntax * try to fix freebsd * Update build.yml * Update build.yml * Update build.yml
train : finetune LORA (#2632 )
2026-02-12 14:03:20 +02:00 · 2023-09-28 22:31:04 +03:00 · 2023-09-28 21:40:11 +03:00 · 2023-09-28 14:30:31 -04:00 · 2023-09-28 14:30:15 -04:00 · 2023-09-28 19:36:36 +03:00
108 changed files with 14637 additions and 4858 deletions
--- a/.devops/cloud-v-pipeline
+++ b/.devops/cloud-v-pipeline
@@ -0,0 +1,22 @@
+node('x86_runner1'){            // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
+    stage('Cleanup'){
+        cleanWs()               // Cleaning previous CI build in workspace
+    }
+    stage('checkout repo'){
+        retry(5){               // Retry if the cloning fails due to some reason
+            checkout scm        // Clone the repo on Runner
+        }
+    }
+    stage('Compiling llama.cpp'){
+        sh'''#!/bin/bash
+            make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
+        '''
+    }
+    stage('Running llama.cpp'){
+        sh'''#!/bin/bash
+            module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
+            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
+            cat llama_log.txt                   # Printing results
+        '''
+    }
+}
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -27,7 +27,7 @@ jobs:
    steps:
      - name: Clone
        id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3

      - name: Dependencies
        id: depends
@@ -38,13 +38,13 @@ jobs:
      - name: Build
        id: make_build
        run: |
-          CC=gcc-8 make
+          CC=gcc-8 make -j $(nproc)

      - name: Test
        id: make_test
        run: |
-          CC=gcc-8 make tests
-          make test
+          CC=gcc-8 make tests -j $(nproc)
+          make test -j $(nproc)

  ubuntu-latest-cmake:
    runs-on: ubuntu-latest
@@ -52,7 +52,7 @@ jobs:
    steps:
      - name: Clone
        id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3

      - name: Dependencies
        id: depends
@@ -66,7 +66,7 @@ jobs:
          mkdir build
          cd build
          cmake ..
-          cmake --build . --config Release
+          cmake --build . --config Release -j $(nproc)

      - name: Test
        id: cmake_test
@@ -87,7 +87,7 @@ jobs:
    steps:
      - name: Clone
        id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3

      - name: Dependencies
        id: depends
@@ -101,7 +101,7 @@ jobs:
          mkdir build
          cd build
          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-          cmake --build . --config ${{ matrix.build_type }}
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)

      - name: Test
        id: cmake_test
@@ -121,7 +121,7 @@ jobs:
    steps:
      - name: Clone
        id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3

      - name: Dependencies
        id: depends
@@ -135,7 +135,7 @@ jobs:
          mkdir build
          cd build
          cmake -DLLAMA_MPI=ON ..
-          cmake --build . --config Release
+          cmake --build . --config Release -j $(nproc)

      - name: Test
        id: cmake_test
@@ -149,7 +149,7 @@ jobs:
    steps:
      - name: Clone
        id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3

      - name: Dependencies
        id: depends
@@ -160,17 +160,46 @@ jobs:
      - name: Build
        id: make_build
        run: |
-          make
+          make -j $(sysctl -n hw.logicalcpu)

      - name: Test
        id: make_test
        run: |
-          make tests
-          make test
+          make tests -j $(sysctl -n hw.logicalcpu)
+          make test -j $(sysctl -n hw.logicalcpu)

  macOS-latest-cmake:
    runs-on: macos-latest

+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest --verbose --timeout 900
+
+  macOS-latest-cmake-ios:
+    runs-on: macos-latest
+
    steps:
      - name: Clone
        id: checkout
@@ -188,14 +217,41 @@ jobs:
          sysctl -a
          mkdir build
          cd build
-          cmake -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF ..
-          cmake --build . --config Release
+          cmake -G Xcode .. \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)

-      - name: Test
-        id: cmake_test
+  macOS-latest-cmake-tvos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
          cd build
-          ctest --verbose --timeout 900
+          cmake -G Xcode .. \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=tvOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)

  windows-latest-cmake:
    runs-on: windows-latest
@@ -209,22 +265,24 @@ jobs:
      matrix:
        include:
          - build: 'noavx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
          - build: 'avx2'
-            defines: '-DLLAMA_BUILD_SERVER=ON'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
          - build: 'avx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
          - build: 'avx512'
            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
          - build: 'clblast'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
          - build: 'openblas'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'

    steps:
      - name: Clone
        id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0

      - name: Download OpenCL SDK
        id: get_opencl
@@ -266,7 +324,7 @@ jobs:
          mkdir build
          cd build
          cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}

      - name: Add clblast.dll
        id: add_clblast_dll
@@ -334,28 +392,30 @@ jobs:

    strategy:
      matrix:
-        cuda: ['12.1.0', '11.7.1']
+        cuda: ['12.2.0', '11.7.1']
        build: ['cublas']

    steps:
      - name: Clone
        id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0

-      - uses: Jimver/cuda-toolkit@v0.2.10
+      - uses: Jimver/cuda-toolkit@v0.2.11
        id: cuda-toolkit
        with:
          cuda: ${{ matrix.cuda }}
-          # TODO(green-sky): _dev seems to fail, and non dev are not enought
-          #sub-packages: '["nvcc", "cudart", "cublas", "cudart_dev", "cublas_dev"]'
+          method: 'network'
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'

      - name: Build
        id: cmake_build
        run: |
          mkdir build
          cd build
-          cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON
-          cmake --build . --config Release
+          cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}

      - name: Determine tag name
        id: tag
@@ -384,27 +444,11 @@ jobs:
            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip

      - name: Copy and pack Cuda runtime
-        if: ${{ matrix.cuda == '12.1.0' }}
-        # TODO(green-sky): paths are cuda 12 specific
        run: |
          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
-          mkdir '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_12.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_12.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_12.dll" '.\build\bin\cudart\'
-          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\*
-
-      - name: Copy and pack Cuda runtime
-        if: ${{ matrix.cuda == '11.7.1' }}
-        # TODO(green-sky): paths are cuda 11 specific
-        run: |
-          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
-          mkdir '.\build\bin\cudart\'
-          ls "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin"
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_110.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_11.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_11.dll" '.\build\bin\cudart\'
-          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\*
+          $dst='.\build\bin\cudart\'
+          robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*

      - name: Upload Cuda runtime
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -413,6 +457,23 @@ jobs:
          path: |
            cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip

+#  freeBSD-latest:
+#    runs-on: macos-12
+#    steps:
+#    - name: Clone
+#      uses: actions/checkout@v3
+#
+#    - name: Build
+#      uses: cross-platform-actions/action@v0.19.0
+#      with:
+#        operating_system: freebsd
+#        version: '13.2'
+#        hypervisor: 'qemu'
+#        run: |
+#            sudo pkg update
+#            sudo pkg install -y gmake automake autoconf pkgconf llvm15 clinfo clover opencl clblast openblas
+#            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu`
+
  release:
    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}

@@ -429,7 +490,9 @@ jobs:
    steps:
      - name: Clone
        id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0

      - name: Determine tag name
        id: tag
@@ -487,7 +550,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -511,7 +574,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -535,7 +598,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -565,7 +628,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -604,7 +667,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -650,7 +713,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -26,8 +26,15 @@ jobs:
    strategy:
      matrix:
        config:
-          - { tag: "light", dockerfile: ".devops/main.Dockerfile" }
-          - { tag: "full", dockerfile: ".devops/full.Dockerfile" }
+          - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
+          #                     have disabled them for now until the reason why
+          #                     is understood.
+          - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
    steps:
      - name: Check out the repo
        uses: actions/checkout@v3
@@ -51,7 +58,7 @@ jobs:
        with:
          context: .
          push: true
-          platforms: linux/amd64,linux/arm64
+          platforms: ${{ matrix.config.platforms }}
          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
          file: ${{ matrix.config.dockerfile }}

@@ -60,6 +67,6 @@ jobs:
        with:
          context: .
          push: ${{ github.event_name == 'push' }}
-          platforms: linux/amd64,linux/arm64
+          platforms: ${{ matrix.config.platforms }}
          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
          file: ${{ matrix.config.dockerfile }}
--- a/.github/workflows/gguf-publish.yml
+++ b/.github/workflows/gguf-publish.yml
@@ -24,7 +24,7 @@ jobs:
    runs-on: ubuntu-latest

    steps:
-    - uses: actions/checkout@v2
+    - uses: actions/checkout@v3
    - name: Set up Python
      uses: actions/setup-python@v2
      with:
--- a/.gitignore
+++ b/.gitignore
@@ -51,7 +51,11 @@ models-mnt
 /save-load-state
 /server
 /simple
+/batched
+/export-lora
+/finetune
 /speculative
+/parallel
 /train-text-from-scratch
 /vdot
 build-info.h
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -80,6 +80,8 @@ set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kern
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
+set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
+                                             "llama: max. batch size for using peer access")
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
@@ -116,7 +118,7 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
    add_custom_command(
        OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h"
        COMMENT "Generating build details from Git"
-        COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
+        COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
        DEPENDS "${GIT_DIR}/index"
        VERBATIM
@@ -135,6 +137,7 @@ set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
+include(CheckCXXCompilerFlag)

 if (NOT MSVC)
    if (LLAMA_SANITIZE_THREAD)
@@ -159,6 +162,8 @@ if (APPLE AND LLAMA_ACCELERATE)
        message(STATUS "Accelerate framework found")

        add_compile_definitions(GGML_USE_ACCELERATE)
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${ACCELERATE_FRAMEWORK})
    else()
        message(WARNING "Accelerate framework not found")
@@ -171,8 +176,8 @@ if (LLAMA_METAL)
    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)

    message(STATUS "Metal framework found")
-
-    set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+    set(GGML_HEADERS_METAL ggml-metal.h)
+    set(GGML_SOURCES_METAL ggml-metal.m)

    add_compile_definitions(GGML_USE_METAL)
    if (LLAMA_METAL_NDEBUG)
@@ -191,7 +196,6 @@ if (LLAMA_METAL)
        ${METALKIT_FRAMEWORK}
        )
 endif()
-
 if (LLAMA_BLAS)
    if (LLAMA_STATIC)
        set(BLA_STATIC ON)
@@ -268,7 +272,8 @@ if (LLAMA_BLAS)
 endif()

 if (LLAMA_K_QUANTS)
-    set(GGML_SOURCES_EXTRA ${GGML_SOURCES_EXTRA} k_quants.c k_quants.h)
+    set(GGML_HEADERS_EXTRA k_quants.h)
+    set(GGML_SOURCES_EXTRA k_quants.c)
    add_compile_definitions(GGML_USE_K_QUANTS)
    if (LLAMA_QKK_64)
        add_compile_definitions(GGML_QKK_64)
@@ -284,7 +289,8 @@ if (LLAMA_CUBLAS)

        enable_language(CUDA)

-        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
+        set(GGML_HEADERS_CUDA ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu)

        add_compile_definitions(GGML_USE_CUBLAS)
 #        if (LLAMA_CUDA_CUBLAS)
@@ -302,6 +308,7 @@ if (LLAMA_CUBLAS)
            add_compile_definitions(GGML_CUDA_F16)
        endif()
        add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})

        if (LLAMA_STATIC)
            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -332,6 +339,7 @@ if (LLAMA_MPI)
    find_package(MPI)
    if (MPI_C_FOUND)
        message(STATUS "MPI found")
+        set(GGML_HEADERS_MPI ggml-mpi.h)
        set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
        add_compile_definitions(GGML_USE_MPI)
        add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
@@ -354,7 +362,8 @@ if (LLAMA_CLBLAST)
    if (CLBlast_FOUND)
        message(STATUS "CLBlast found")

-        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
+        set(GGML_HEADERS_OPENCL ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp)

        add_compile_definitions(GGML_USE_CLBLAST)

@@ -382,13 +391,15 @@ if (LLAMA_HIPBLAS)
        message(STATUS "HIP and hipBLAS found")
        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        if (BUILD_SHARED_LIBS)
+            set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        endif()
        if (LLAMA_CUDA_FORCE_DMMV)
            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
        endif()
        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-        target_compile_definitions(ggml-rocm PRIVATE CC_TURING=1000000000)
        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)

@@ -421,6 +432,7 @@ if (LLAMA_ALL_WARNINGS)
            -Wextra
            -Wpedantic
            -Wcast-qual
+            -Wmissing-declarations
            -Wno-unused-function
            -Wno-multichar
        )
@@ -439,7 +451,7 @@ if (LLAMA_ALL_WARNINGS)

 endif()

-if (MSVC)
+if (WIN32)
    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)

    if (BUILD_SHARED_LIBS)
@@ -461,6 +473,13 @@ endif()
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
+if (MSVC)
+  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+else ()
+  set(CMAKE_GENERATOR_PLATFORM_LWR "")
+endif ()
+
 if (NOT MSVC)
    if (LLAMA_STATIC)
        add_link_options(-static)
@@ -476,25 +495,33 @@ if (NOT MSVC)
    endif()
 endif()

-if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
+if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
    message(STATUS "ARM detected")
    if (MSVC)
-        # TODO: arm msvc?
+        add_compile_definitions(__ARM_NEON)
+        add_compile_definitions(__ARM_FEATURE_FMA)
+        add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
    else()
+        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            add_compile_options(-mfp16-format=ieee)
+        endif()
        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
            # Raspberry Pi 1, Zero
-            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access)
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access)
        endif()
        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
            # Raspberry Pi 2
-            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations)
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
        endif()
        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
            # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            add_compile_options(-mfp16-format=ieee -mno-unaligned-access)
+            add_compile_options(-mno-unaligned-access)
        endif()
    endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
    message(STATUS "x86 detected")
    if (MSVC)
        if (LLAMA_AVX512)
@@ -578,10 +605,12 @@ endif()
 # RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
 # and on macOS its availability depends on enabling Darwin extensions
 # similarly on DragonFly, enabling BSD extensions is necessary
-if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
-    add_compile_definitions(_DARWIN_C_SOURCE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
+if (
+    CMAKE_SYSTEM_NAME MATCHES "Darwin" OR
+    CMAKE_SYSTEM_NAME MATCHES "iOS" OR
+    CMAKE_SYSTEM_NAME MATCHES "tvOS" OR
+    CMAKE_SYSTEM_NAME MATCHES "DragonFly"
+)
    add_compile_definitions(_DARWIN_C_SOURCE)
 endif()

@@ -614,11 +643,11 @@ add_library(ggml OBJECT
            ggml.h
            ggml-alloc.c
            ggml-alloc.h
-            ${GGML_SOURCES_CUDA}
-            ${GGML_SOURCES_OPENCL}
-            ${GGML_SOURCES_METAL}
-            ${GGML_SOURCES_MPI}
-            ${GGML_SOURCES_EXTRA}
+            ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
+            ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
            )

 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
@@ -656,14 +685,53 @@ if (BUILD_SHARED_LIBS)
    if (LLAMA_METAL)
        set_target_properties(llama PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
    endif()
-    install(TARGETS llama LIBRARY)
 endif()

+
 #
 # install
 #

 include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
+
+set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR}
+    CACHE PATH "Location of header files")
+set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR}
+    CACHE PATH "Location of library files")
+set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR}
+    CACHE PATH "Location of binary files")
+set(LLAMA_BUILD_NUMBER ${BUILD_NUMBER})
+set(LLAMA_BUILD_COMMIT ${BUILD_COMMIT})
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/scripts/LlamaConfig.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama
+    PATH_VARS LLAMA_INCLUDE_INSTALL_DIR
+              LLAMA_LIB_INSTALL_DIR
+              LLAMA_BIN_INSTALL_DIR )
+
+write_basic_package_version_file(
+        ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake
+    VERSION ${LLAMA_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
+
+set(GGML_PUBLIC_HEADERS "ggml.h"
+        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
+
+set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
+install(TARGETS ggml PUBLIC_HEADER)
+
+set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/llama.h)
+install(TARGETS llama LIBRARY PUBLIC_HEADER)
+
 install(
    FILES convert.py
    PERMISSIONS
--- a/122
+++ b/122
@@ -1,8 +1,8 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative tests/test-c.o
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative parallel finetune export-lora tests/test-c.o

 # Binaries only useful for tests
-TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1
+TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama

 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -49,7 +49,7 @@ test: $(TEST_TARGETS)
 			./$$test_target $(CURDIR)/models/ggml-vocab-llama.gguf; \
 		elif [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \
 			continue; \
-		elif [ "$$test_target" = "tests/test-tokenizer-1" ]; then \
+		elif [ "$$test_target" = "tests/test-tokenizer-1-llama" ]; then \
 			continue; \
 		else \
 			echo "Running test $$test_target..."; \
@@ -95,65 +95,60 @@ CXXV := $(shell $(CXX) --version | head -n 1)
 #

 # keep standard at C11 and C++11
+MK_CPPFLAGS = -I. -Icommon
+MK_CFLAGS   = -std=c11   -fPIC
+MK_CXXFLAGS = -std=c++11 -fPIC
+
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
-OPT = -Ofast
+MK_CFLAGS        += -Ofast
+MK_HOST_CXXFLAGS += -Ofast
+MK_CUDA_CXXFLAGS += -O3
 else
-OPT = -O3
+MK_CFLAGS        += -O3
+MK_CXXFLAGS      += -O3
 endif
-MK_CPPFLAGS = -I. -Icommon
-MK_CFLAGS   = $(OPT) -std=c11   -fPIC
-MK_CXXFLAGS = $(OPT) -std=c++11 -fPIC
-MK_LDFLAGS  =

 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
 # posix_memalign came in POSIX.1-2001 / SUSv3
 # M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-MK_CFLAGS   += -D_XOPEN_SOURCE=600
-MK_CXXFLAGS += -D_XOPEN_SOURCE=600
+MK_CPPFLAGS += -D_XOPEN_SOURCE=600

 # Somehow in OpenBSD whenever POSIX conformance is specified
 # some string functions rely on locale_t availability,
 # which was introduced in POSIX.1-2008, forcing us to go higher
 ifeq ($(UNAME_S),OpenBSD)
-	MK_CFLAGS   += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
-	MK_CXXFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
+	MK_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
 endif

 # Data types, macros and functions related to controlling CPU affinity and
 # some memory allocation are available on Linux through GNU extensions in libc
 ifeq ($(UNAME_S),Linux)
-	MK_CFLAGS   += -D_GNU_SOURCE
-	MK_CXXFLAGS += -D_GNU_SOURCE
+	MK_CPPFLAGS += -D_GNU_SOURCE
 endif

 # RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
 # and on macOS its availability depends on enabling Darwin extensions
 # similarly on DragonFly, enabling BSD extensions is necessary
 ifeq ($(UNAME_S),Darwin)
-	MK_CFLAGS   += -D_DARWIN_C_SOURCE
-	MK_CXXFLAGS += -D_DARWIN_C_SOURCE
+	MK_CPPFLAGS += -D_DARWIN_C_SOURCE
 endif
 ifeq ($(UNAME_S),DragonFly)
-	MK_CFLAGS   += -D__BSD_VISIBLE
-	MK_CXXFLAGS += -D__BSD_VISIBLE
+	MK_CPPFLAGS += -D__BSD_VISIBLE
 endif

 # alloca is a non-standard interface that is not visible on BSDs when
 # POSIX conformance is specified, but not all of them provide a clean way
 # to enable it in such cases
 ifeq ($(UNAME_S),FreeBSD)
-	MK_CFLAGS   += -D__BSD_VISIBLE
-	MK_CXXFLAGS += -D__BSD_VISIBLE
+	MK_CPPFLAGS += -D__BSD_VISIBLE
 endif
 ifeq ($(UNAME_S),NetBSD)
-	MK_CFLAGS   += -D_NETBSD_SOURCE
-	MK_CXXFLAGS += -D_NETBSD_SOURCE
+	MK_CPPFLAGS += -D_NETBSD_SOURCE
 endif
 ifeq ($(UNAME_S),OpenBSD)
-	MK_CFLAGS   += -D_BSD_SOURCE
-	MK_CXXFLAGS += -D_BSD_SOURCE
+	MK_CPPFLAGS += -D_BSD_SOURCE
 endif

 ifdef LLAMA_DEBUG
@@ -180,9 +175,16 @@ endif # LLAMA_DISABLE_LOGS
 # warnings
 MK_CFLAGS    += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \
 				-Wmissing-prototypes -Werror=implicit-int -Wno-unused-function
-MK_CXXFLAGS  += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
+MK_CXXFLAGS  += -Wall -Wextra -Wpedantic -Wcast-qual -Wmissing-declarations -Wno-unused-function -Wno-multichar

-ifeq '' '$(findstring clang++,$(CXX))'
+# TODO(cebtenzzre): remove this once PR #2632 gets merged
+TTFS_CXXFLAGS = $(CXXFLAGS) -Wno-missing-declarations
+
+ifneq '' '$(findstring clang,$(shell $(CXX) --version))'
+	# clang++ only
+	MK_CXXFLAGS   += -Wmissing-prototypes
+	TTFS_CXXFLAGS += -Wno-missing-prototypes
+else
 	# g++ only
 	MK_CXXFLAGS += -Wno-format-truncation -Wno-array-bounds
 endif
@@ -233,7 +235,7 @@ ifndef RISCV
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
 	MK_CFLAGS   += -march=native -mtune=native
-	MK_CXXFLAGS += -march=native -mtune=native
+	MK_HOST_CXXFLAGS += -march=native -mtune=native

 	# Usage AVX-only
 	#MK_CFLAGS   += -mfma -mf16c -mavx
@@ -303,6 +305,8 @@ ifndef LLAMA_NO_ACCELERATE
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
 		MK_CPPFLAGS += -DGGML_USE_ACCELERATE
+		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
+		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
 		MK_LDFLAGS  += -framework Accelerate
 	endif
 endif # LLAMA_NO_ACCELERATE
@@ -366,6 +370,11 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
+ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+else
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 #ifdef LLAMA_CUDA_CUBLAS
 #	NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
@@ -373,7 +382,7 @@ ifdef LLAMA_CUDA_CCBIN
 	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) $(subst -Ofast,-O3,$(CXXFLAGS)) -Wno-pedantic -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS

 ifdef LLAMA_CLBLAST
@@ -408,7 +417,6 @@ ifdef LLAMA_HIPBLAS
 	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
-	HIPFLAGS    += -DCC_TURING=1000000000
 ifdef LLAMA_CUDA_FORCE_DMMV
 	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
@@ -442,23 +450,30 @@ k_quants.o: k_quants.c k_quants.h
 endif # LLAMA_NO_K_QUANTS

 # combine build flags with cmdline overrides
-override CFLAGS   := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
-override CXXFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
-override LDFLAGS  := $(MK_LDFLAGS) $(LDFLAGS)
+override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
+override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CUDA_CXXFLAGS := $(MK_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS)
+override HOST_CXXFLAGS := $(MK_HOST_CXXFLAGS) $(HOST_CXXFLAGS)
+override LDFLAGS       := $(MK_LDFLAGS) $(LDFLAGS)
+
+# save CXXFLAGS before we add host-only options
+NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
+override CXXFLAGS += $(HOST_CXXFLAGS)

 #
 # Print build information
 #

 $(info I llama.cpp build info: )
-$(info I UNAME_S:  $(UNAME_S))
-$(info I UNAME_P:  $(UNAME_P))
-$(info I UNAME_M:  $(UNAME_M))
-$(info I CFLAGS:   $(CFLAGS))
-$(info I CXXFLAGS: $(CXXFLAGS))
-$(info I LDFLAGS:  $(LDFLAGS))
-$(info I CC:       $(CCV))
-$(info I CXX:      $(CXXV))
+$(info I UNAME_S:   $(UNAME_S))
+$(info I UNAME_P:   $(UNAME_P))
+$(info I UNAME_M:   $(UNAME_M))
+$(info I CFLAGS:    $(CFLAGS))
+$(info I CXXFLAGS:  $(CXXFLAGS))
+$(info I NVCCFLAGS: $(NVCCFLAGS))
+$(info I LDFLAGS:   $(LDFLAGS))
+$(info I CC:        $(CCV))
+$(info I CXX:       $(CXXV))
 $(info )

 #
@@ -485,6 +500,9 @@ console.o: common/console.cpp common/console.h
 grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@

+train.o: common/train.cpp common/train.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)

@@ -504,6 +522,9 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

+batched: examples/batched/batched.cpp                         build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

@@ -532,8 +553,8 @@ embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-te
 gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o common.o train.o $(OBJS)
+	$(CXX) $(TTFS_CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

 convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
@@ -541,22 +562,31 @@ convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggm
 llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o $(OBJS)
+baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

 beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

+finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o common.o train.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

+parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 ifdef LLAMA_METAL
 metal: examples/metal/metal.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 endif

 build-info.h: $(wildcard .git/index) scripts/build-info.sh
-	@sh scripts/build-info.sh > $@.tmp
+	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
 	else \
@@ -606,7 +636,7 @@ tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h gg
 tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

 tests/test-c.o: tests/test-c.c llama.h
--- a/Package.swift
+++ b/Package.swift
@@ -2,8 +2,30 @@

 import PackageDescription

+#if arch(arm) || arch(arm64)
+let platforms: [SupportedPlatform]? = [
+    .macOS(.v11),
+    .iOS(.v14),
+    .watchOS(.v4),
+    .tvOS(.v14)
+]
+let exclude: [String] = []
+let additionalSources: [String] = ["ggml-metal.m"]
+let additionalSettings: [CSetting] = [
+    .unsafeFlags(["-fno-objc-arc"]),
+    .define("GGML_SWIFT"),
+    .define("GGML_USE_METAL")
+]
+#else
+let platforms: [SupportedPlatform]? = nil
+let exclude: [String] = ["ggml-metal.metal"]
+let additionalSources: [String] = []
+let additionalSettings: [CSetting] = []
+#endif
+
 let package = Package(
    name: "llama",
+    platforms: platforms,
    products: [
        .library(name: "llama", targets: ["llama"]),
    ],
@@ -11,23 +33,25 @@ let package = Package(
        .target(
            name: "llama",
            path: ".",
-            exclude: ["ggml-metal.metal"],
+            exclude: exclude,
            sources: [
                "ggml.c",
                "llama.cpp",
                "ggml-alloc.c",
-                "k_quants.c"
-            ],
+                "k_quants.c",
+            ] + additionalSources,
            publicHeadersPath: "spm-headers",
            cSettings: [
                .unsafeFlags(["-Wno-shorten-64-to-32"]),
                .define("GGML_USE_K_QUANTS"),
                .define("GGML_USE_ACCELERATE")
-            ],
+                .define("ACCELERATE_NEW_LAPACK")
+                .define("ACCELERATE_LAPACK_ILP64")
+            ] + additionalSettings,
            linkerSettings: [
                .linkedFramework("Accelerate")
            ]
-        ),
+        )
    ],
    cxxLanguageStandard: .cxx11
 )
--- a/README.md
+++ b/README.md
@@ -11,6 +11,8 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++

 ### Hot topics

+- Parallel decoding + continuous batching support incoming: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
+  **Devs should become familiar with the new API**
 - Local Falcon 180B inference on Mac Studio

  https://github.com/ggerganov/llama.cpp/assets/1991296/98abd4e8-7077-464c-ae89-aebabca7757e
@@ -90,6 +92,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
 - [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) and its derivations (such as [baichuan-7b-sft](https://huggingface.co/hiyouga/baichuan-7b-sft))
 - [X] [Aquila-7B](https://huggingface.co/BAAI/Aquila-7B) / [AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
+- [X] Mistral AI v0.1

 **Bindings:**

@@ -391,17 +394,18 @@ Building the program with BLAS support may lead to some performance improvements
 <!---
  | LLAMA_CUDA_CUBLAS       | Boolean                |   false | Use cuBLAS instead of custom CUDA kernels for prompt processing. Faster for all quantization formats except for q4_0 and q8_0, especially for k-quants. Increases VRAM usage (700 MiB for 7b, 970 MiB for 13b, 1430 MiB for 33b). |
 --->
-  | Option                  | Legal values           | Default | Description |
-  |-------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
-  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
-  | LLAMA_CUDA_F16          | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
-  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+  | Option                         | Legal values           | Default | Description |
+  |--------------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_FORCE_DMMV          | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
+  | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y               | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
+  | LLAMA_CUDA_F16                 | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
+  | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+  | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       |     128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |

 - #### hipBLAS

-  This provide BLAS acceleation on HIP supported GPU like AMD GPU.
+  This provides BLAS acceleration on HIP-supported AMD GPUs.
  Make sure to have ROCm installed.
  You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
  Windows support is coming soon...
@@ -498,7 +502,7 @@ Building the program with BLAS support may lead to some performance improvements
    ```sh
    mkdir build
    cd build
-    cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_dir=/some/path
+    cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path
    cmake --build . --config Release
    ```
  - CMake (Windows):
@@ -554,6 +558,10 @@ python3 convert.py models/7B/
 # quantize the model to 4-bits (using q4_0 method)
 ./quantize ./models/7B/ggml-model-f16.gguf ./models/7B/ggml-model-q4_0.gguf q4_0

+# update the gguf filetype to current if older version is unsupported by another application
+./quantize ./models/7B/ggml-model-q4_0.gguf ./models/7B/ggml-model-q4_0-v2.gguf COPY
+
+
 # run the inference
 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
 ```
@@ -590,6 +598,11 @@ Several quantization methods are supported. They differ in the resulting model d
 |   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
 |   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |

+- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
+- recent k-quants improvements
+  - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
+  - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
+
 ### Perplexity (measuring model quality)

 You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
@@ -844,8 +857,17 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th
 #### Images
 We have two Docker images available for this project:

-1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
-2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file.
+1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+
+Additionally, there the following images, similar to the above:
+
+- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the Gitlab Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).

 #### Usage

--- a/build.zig
+++ b/build.zig
@@ -36,17 +36,20 @@ const Maker = struct {
    }

    fn init(builder: *std.build.Builder) !Maker {
-        const commit_hash = @embedFile(".git/refs/heads/master");
+        // const commit_hash = @embedFile(".git/refs/heads/master");
+        const target = builder.standardTargetOptions(.{});
        const config_header = builder.addConfigHeader(
            .{ .style = .blank, .include_path = "build-info.h" },
            .{
                .BUILD_NUMBER = 0,
-                .BUILD_COMMIT = commit_hash[0 .. commit_hash.len - 1], // omit newline
+                .BUILD_COMMIT = "12345", // omit newline
+                .BUILD_COMPILER = "Zig 0.11.0",
+                .BUILD_TARGET = try target.allocDescription(builder.allocator),
            },
        );
        var m = Maker{
            .builder = builder,
-            .target = builder.standardTargetOptions(.{}),
+            .target = target,
            .optimize = builder.standardOptimizeOption(.{}),
            .config_header = config_header,
            .enable_lto = false,
@@ -58,7 +61,7 @@ const Maker = struct {
        try m.addCFlag("-std=c11");
        try m.addCxxFlag("-std=c++11");
        try m.addProjectInclude(&.{});
-        try m.addProjectInclude(&.{"examples"});
+        try m.addProjectInclude(&.{"common"});
        return m;
    }

@@ -71,6 +74,7 @@ const Maker = struct {
            o.addCSourceFiles(&.{src}, m.cxxflags.items);
            o.linkLibCpp();
        }
+        o.addConfigHeader(m.config_header);
        for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
        o.want_lto = m.enable_lto;
        return o;
@@ -104,15 +108,15 @@ pub fn build(b: *std.build.Builder) !void {
    const ggml = make.obj("ggml", "ggml.c");
    const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
    const llama = make.obj("llama", "llama.cpp");
-    const common = make.obj("common", "examples/common.cpp");
-    const console = make.obj("common", "examples/console.cpp");
-    const grammar_parser = make.obj("grammar-parser", "examples/grammar-parser.cpp");
+    const common = make.obj("common", "common/common.cpp");
+    const console = make.obj("common", "common/console.cpp");
+    const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");

    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, llama, common, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, llama });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, llama, common });
    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, llama, common });
    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, llama, common });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, llama });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, llama, common });

    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, llama, common, grammar_parser });
    if (server.target.isWindows()) {
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -9,6 +9,8 @@ add_library(${TARGET} OBJECT
    console.cpp
    grammar-parser.h
    grammar-parser.cpp
+    train.h
+    train.cpp
    )

 if (BUILD_SHARED_LIBS)
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -317,6 +317,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                break;
            }
            params.n_chunks = std::stoi(argv[i]);
+        } else if (arg == "-np" || arg == "--parallel") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_parallel = std::stoi(argv[i]);
+        } else if (arg == "-ns" || arg == "--sequences") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_sequences = std::stoi(argv[i]);
        } else if (arg == "-m" || arg == "--model") {
            if (++i >= argc) {
                invalid_param = true;
@@ -340,7 +352,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                invalid_param = true;
                break;
            }
-            params.lora_adapter = argv[i];
+            params.lora_adapter.push_back({argv[i], 1.0f});
+            params.use_mmap = false;
+        } else if (arg == "--lora-scaled") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            const char * lora_adapter = argv[i];
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter.push_back({lora_adapter, std::stof(argv[i])});
            params.use_mmap = false;
        } else if (arg == "--lora-base") {
            if (++i >= argc) {
@@ -360,6 +384,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
            params.multiline_input = true;
        } else if (arg == "--simple-io") {
            params.simple_io = true;
+        } else if (arg == "-cb" || arg == "--cont-batching") {
+            params.cont_batching = true;
        } else if (arg == "--color") {
            params.use_color = true;
        } else if (arg == "--mlock") {
@@ -374,6 +400,17 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #else
            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
+        } else if (arg == "--gpu-layers-draft" || arg == "-ngld" || arg == "--n-gpu-layers-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params.n_gpu_layers_draft = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
        } else if (arg == "--main-gpu" || arg == "-mg") {
            if (++i >= argc) {
@@ -423,12 +460,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #endif // GGML_USE_CUBLAS
        } else if (arg == "--no-mmap") {
            params.use_mmap = false;
-        } else if (arg == "--mtest") {
-            params.mem_test = true;
        } else if (arg == "--numa") {
            params.numa = true;
-        } else if (arg == "--export") {
-            params.export_cgraph = true;
        } else if (arg == "--verbose-prompt") {
            params.verbose_prompt = true;
        } else if (arg == "-r" || arg == "--reverse-prompt") {
@@ -447,8 +480,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
            if (params.logdir.back() != DIRECTORY_SEPARATOR) {
                params.logdir += DIRECTORY_SEPARATOR;
            }
-        } else if (arg == "--perplexity") {
-            params.perplexity = true;
+        } else if (arg == "--perplexity" || arg == "--all-logits") {
+            params.logits_all = true;
        } else if (arg == "--ppl-stride") {
            if (++i >= argc) {
                invalid_param = true;
@@ -638,20 +671,23 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
    printf("  --cfg-negative-prompt-file FNAME\n");
    printf("                        negative prompt file to use for guidance. (default: empty)\n");
    printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
-    printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
-    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
+    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
+    printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
    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("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
    printf("  --temp N              temperature (default: %.1f)\n", (double)params.temp);
-    printf("  --perplexity          compute perplexity over each ctx window of the prompt\n");
+    printf("  --logits-all          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");
    printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
    printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
    printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
    printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
+    printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
+    printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
    if (llama_mlock_supported()) {
        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
    }
@@ -664,6 +700,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
    printf("  -ngl N, --n-gpu-layers N\n");
    printf("                        number of layers to store in VRAM\n");
+    printf("  -ngld N, --n-gpu-layers-draft N\n");
+    printf("                        number of layers to store in VRAM for the draft model\n");
    printf("  -ts SPLIT --tensor-split SPLIT\n");
    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
@@ -674,11 +712,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif // GGML_USE_CUBLAS
 #endif
-    printf("  --mtest               compute maximum memory usage\n");
-    printf("  --export              export the computation graph to 'llama.ggml'\n");
    printf("  --verbose-prompt      print prompt before generation\n");
    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
    printf("  -m FNAME, --model FNAME\n");
    printf("                        model path (default: %s)\n", params.model.c_str());
@@ -728,7 +765,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
    lparams.f16_kv          = params.memory_f16;
    lparams.use_mmap        = params.use_mmap;
    lparams.use_mlock       = params.use_mlock;
-    lparams.logits_all      = params.perplexity;
+    lparams.logits_all      = params.logits_all;
    lparams.embedding       = params.embedding;
    lparams.rope_freq_base  = params.rope_freq_base;
    lparams.rope_freq_scale = params.rope_freq_scale;
@@ -752,10 +789,15 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
        return std::make_tuple(nullptr, nullptr);
    }

-    if (!params.lora_adapter.empty()) {
+    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
+        const std::string& lora_adapter = std::get<0>(params.lora_adapter[i]);
+        float lora_scale = std::get<1>(params.lora_adapter[i]);
        int err = llama_model_apply_lora_from_file(model,
-                                             params.lora_adapter.c_str(),
-                                             params.lora_base.empty() ? NULL : params.lora_base.c_str(),
+                                             lora_adapter.c_str(),
+                                             lora_scale,
+                                             ((i > 0) || params.lora_base.empty())
+                                                ? NULL
+                                                : params.lora_base.c_str(),
                                             params.n_threads);
        if (err != 0) {
            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
@@ -772,8 +814,9 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
    {
        LOG("warming up the model with an empty run\n");

-        const std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
-        llama_eval(lctx, tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, params.n_threads);
+        std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0), params.n_threads);
+        llama_kv_cache_tokens_rm(lctx, -1, -1);
        llama_reset_timings(lctx);
    }

@@ -791,10 +834,10 @@ std::vector<llama_token> llama_tokenize(
    // upper limit for the number of tokens
    int n_tokens = text.length() + add_bos;
    std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(ctx, text.c_str(), result.data(), result.size(), add_bos);
+    n_tokens = llama_tokenize(ctx, text.data(), text.length(), result.data(), result.size(), add_bos);
    if (n_tokens < 0) {
        result.resize(-n_tokens);
-        int check = llama_tokenize(ctx, text.c_str(), result.data(), result.size(), add_bos);
+        int check = llama_tokenize(ctx, text.data(), text.length(), result.data(), result.size(), add_bos);
        GGML_ASSERT(check == -n_tokens);
    } else {
        result.resize(n_tokens);
@@ -880,7 +923,7 @@ llama_token llama_sample_token(

    llama_token id = 0;

-    float * logits = llama_get_logits(ctx) + idx * n_vocab;
+    float * logits = llama_get_logits_ith(ctx, idx);

    // Apply params.logit_bias map
    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
@@ -931,11 +974,11 @@ llama_token llama_sample_token(
        if (mirostat == 1) {
            static float mirostat_mu = 2.0f * mirostat_tau;
            const int mirostat_m = 100;
-            llama_sample_temperature(ctx, &cur_p, temp);
+            llama_sample_temp(ctx, &cur_p, temp);
            id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
        } else if (mirostat == 2) {
            static float mirostat_mu = 2.0f * mirostat_tau;
-            llama_sample_temperature(ctx, &cur_p, temp);
+            llama_sample_temp(ctx, &cur_p, temp);
            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
        } else {
            // Temperature sampling
@@ -943,7 +986,7 @@ llama_token llama_sample_token(
            llama_sample_tail_free  (ctx, &cur_p, tfs_z, 1);
            llama_sample_typical    (ctx, &cur_p, typical_p, 1);
            llama_sample_top_p      (ctx, &cur_p, top_p, 1);
-            llama_sample_temperature(ctx, &cur_p, temp);
+            llama_sample_temp(ctx, &cur_p, temp);

            {
                const int n_top = 10;
@@ -1172,7 +1215,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
    fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
    fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
    fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
-    fprintf(stream, "export: %s # default: false\n", params.export_cgraph ? "true" : "false");
    fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", params.frequency_penalty);
    dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
@@ -1201,7 +1243,20 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
        fprintf(stream, "  %d: %f", lb.first, lb.second);
    }

-    fprintf(stream, "lora: %s\n", params.lora_adapter.c_str());
+    fprintf(stream, "lora:\n");
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) != 1.0f) {
+            continue;
+        }
+        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
+    }
+    fprintf(stream, "lora_scaled:\n");
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) == 1.0f) {
+            continue;
+        }
+        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
+    }
    fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
    fprintf(stream, "low_vram: %s # default: false\n", params.low_vram ? "true" : "false");
    fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
@@ -1212,7 +1267,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
    fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
    fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str());
    fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
-    fprintf(stream, "mtest: %s # default: false\n", params.mem_test ? "true" : "false");
    fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
    fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
@@ -1247,6 +1301,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
    fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
    fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
    fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
+    fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
    fprintf(stream, "temp: %f # default: 0.8\n", params.temp);

    const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);
--- a/common/common.h
+++ b/common/common.h
@@ -20,8 +20,13 @@
 #define DIRECTORY_SEPARATOR '/'
 #endif // _WIN32

-#define die(msg)          do { fputs("error: " msg "\n", stderr);                  exit(1); } while (0)
-#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", ##__VA_ARGS__); exit(1); } while (0)
+#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
+#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
+
+#define print_build_info() do {                                                             \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+} while(0)

 //
 // CLI argument parsing
@@ -37,13 +42,16 @@ struct gpt_params {
    int32_t n_keep                          = 0;    // number of tokens to keep from initial prompt
    int32_t n_draft                         = 16;   // number of tokens to draft during speculative decoding
    int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
+    int32_t n_parallel                      = 1;    // number of parallel sequences to decode
+    int32_t n_sequences                     = 1;    // number of sequences to decode
    int32_t n_gpu_layers                    = -1;   // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft              = -1;   // number of layers to store in VRAM for the draft model (-1 - use default)
    int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
    float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
    int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
    int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
-    float   rope_freq_base                  = 10000.0f; // RoPE base frequency
-    float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
+    float   rope_freq_base                  = 0.0f; // RoPE base frequency
+    float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor

    // sampling parameters
    int32_t top_k             = 40;    // <= 0 to use vocab size
@@ -77,8 +85,8 @@ struct gpt_params {
    std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
    std::string logdir            = "";  // directory in which to save YAML log files

-    std::string lora_adapter = "";  // lora adapter path
-    std::string lora_base    = "";  // base model path for the lora adapter
+    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
+    std::string lora_base  = "";                              // base model path for the lora adapter

    int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
    int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -101,17 +109,16 @@ struct gpt_params {
    bool interactive_first = false; // wait for user input immediately
    bool multiline_input   = false; // reverse the usage of `\`
    bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
+    bool cont_batching     = false; // insert new sequences for decoding on-the-fly

    bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
    bool ignore_eos        = false; // ignore generated EOS tokens
    bool instruct          = false; // instruction mode (used for Alpaca models)
    bool penalize_nl       = true;  // consider newlines as a repeatable token
-    bool perplexity        = false; // compute perplexity over the prompt
+    bool logits_all        = false; // return logits for all tokens in the batch
    bool use_mmap          = true;  // use mmap for faster loads
    bool use_mlock         = false; // use mlock to keep model in memory
-    bool mem_test          = false; // compute maximum memory usage
    bool numa              = false; // attempt optimizations that help on some NUMA systems
-    bool export_cgraph     = false; // export the computation graph
    bool verbose_prompt    = false; // print prompt tokens before generation
 };

@@ -121,6 +128,8 @@ void gpt_print_usage(int argc, char ** argv, const gpt_params & params);

 std::string gpt_random_prompt(std::mt19937 & rng);

+void process_escapes(std::string& input);
+
 //
 // Model utils
 //
@@ -176,7 +185,7 @@ std::string llama_detokenize_bpe(
 //  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
 //  - grammar:       grammar to use for sampling, ignore if NULL
 //  - last_tokens:   needed for repetition penalty, ignore if empty
-//  - idx:           sample from llama_get_logits(ctx) + idx * n_vocab
+//  - idx:           sample from llama_get_logits_ith(ctx, idx)
 //
 // returns:
 //  - token:      sampled token
--- a/common/console.cpp
+++ b/common/console.cpp
@@ -158,7 +158,7 @@ namespace console {
        }
    }

-    char32_t getchar32() {
+    static char32_t getchar32() {
 #if defined(_WIN32)
        HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
        wchar_t high_surrogate = 0;
@@ -212,7 +212,7 @@ namespace console {
 #endif
    }

-    void pop_cursor() {
+    static void pop_cursor() {
 #if defined(_WIN32)
        if (hConsole != NULL) {
            CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
@@ -233,7 +233,7 @@ namespace console {
        putc('\b', out);
    }

-    int estimateWidth(char32_t codepoint) {
+    static int estimateWidth(char32_t codepoint) {
 #if defined(_WIN32)
        (void)codepoint;
        return 1;
@@ -242,7 +242,7 @@ namespace console {
 #endif
    }

-    int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
+    static int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
 #if defined(_WIN32)
        CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
        if (!GetConsoleScreenBufferInfo(hConsole, &bufferInfo)) {
@@ -303,7 +303,7 @@ namespace console {
 #endif
    }

-    void replace_last(char ch) {
+    static void replace_last(char ch) {
 #if defined(_WIN32)
        pop_cursor();
        put_codepoint(&ch, 1, 1);
@@ -312,7 +312,7 @@ namespace console {
 #endif
    }

-    void append_utf8(char32_t ch, std::string & out) {
+    static void append_utf8(char32_t ch, std::string & out) {
        if (ch <= 0x7F) {
            out.push_back(static_cast<unsigned char>(ch));
        } else if (ch <= 0x7FF) {
@@ -333,7 +333,7 @@ namespace console {
    }

    // Helper function to remove the last UTF-8 character from a string
-    void pop_back_utf8_char(std::string & line) {
+    static void pop_back_utf8_char(std::string & line) {
        if (line.empty()) {
            return;
        }
@@ -349,7 +349,7 @@ namespace console {
        line.erase(pos);
    }

-    bool readline_advanced(std::string & line, bool multiline_input) {
+    static bool readline_advanced(std::string & line, bool multiline_input) {
        if (out != stdout) {
            fflush(stdout);
        }
@@ -452,7 +452,7 @@ namespace console {
        return has_more;
    }

-    bool readline_simple(std::string & line, bool multiline_input) {
+    static bool readline_simple(std::string & line, bool multiline_input) {
 #if defined(_WIN32)
        std::wstring wline;
        if (!std::getline(std::wcin, wline)) {
--- a/common/grammar-parser.cpp
+++ b/common/grammar-parser.cpp
@@ -9,7 +9,7 @@
 namespace grammar_parser {
    // NOTE: assumes valid utf8 (but checks for overrun)
    // copied from llama.cpp
-    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
        uint8_t  first_byte = static_cast<uint8_t>(*src);
        uint8_t  highbits   = first_byte >> 4;
@@ -24,19 +24,19 @@ namespace grammar_parser {
        return std::make_pair(value, pos);
    }

-    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
        return result.first->second;
    }

-    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
        return next_id;
    }

-    void add_rule(
+    static void add_rule(
            parse_state & state,
            uint32_t      rule_id,
            const std::vector<llama_grammar_element> & rule) {
@@ -46,11 +46,11 @@ namespace grammar_parser {
        state.rules[rule_id] = rule;
    }

-    bool is_word_char(char c) {
+    static bool is_word_char(char c) {
        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
    }

-    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
        const char * pos   = src;
        const char * end   = src + size;
        uint32_t     value = 0;
@@ -73,7 +73,7 @@ namespace grammar_parser {
        return std::make_pair(value, pos);
    }

-    const char * parse_space(const char * src, bool newline_ok) {
+    static const char * parse_space(const char * src, bool newline_ok) {
        const char * pos = src;
        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
@@ -88,7 +88,7 @@ namespace grammar_parser {
        return pos;
    }

-    const char * parse_name(const char * src) {
+    static const char * parse_name(const char * src) {
        const char * pos = src;
        while (is_word_char(*pos)) {
            pos++;
@@ -99,7 +99,7 @@ namespace grammar_parser {
        return pos;
    }

-    std::pair<uint32_t, const char *> parse_char(const char * src) {
+    static std::pair<uint32_t, const char *> parse_char(const char * src) {
        if (*src == '\\') {
            switch (src[1]) {
                case 'x': return parse_hex(src + 2, 2);
@@ -129,7 +129,7 @@ namespace grammar_parser {
            uint32_t            rule_id,
            bool                is_nested);

-    const char * parse_sequence(
+    static const char * parse_sequence(
            parse_state                        & state,
            const char                         * src,
            const std::string                  & rule_name,
@@ -247,7 +247,7 @@ namespace grammar_parser {
        return pos;
    }

-    const char * parse_rule(parse_state & state, const char * src) {
+    static const char * parse_rule(parse_state & state, const char * src) {
        const char * name_end = parse_name(src);
        const char * pos      = parse_space(name_end, false);
        size_t       name_len = name_end - src;
@@ -285,7 +285,7 @@ namespace grammar_parser {
        }
    }

-    void print_grammar_char(FILE * file, uint32_t c) {
+    static void print_grammar_char(FILE * file, uint32_t c) {
        if (0x20 <= c && c <= 0x7f) {
            fprintf(file, "%c", static_cast<char>(c));
        } else {
@@ -294,7 +294,7 @@ namespace grammar_parser {
        }
    }

-    bool is_char_element(llama_grammar_element elem) {
+    static bool is_char_element(llama_grammar_element elem) {
        switch (elem.type) {
            case LLAMA_GRETYPE_CHAR:           return true;
            case LLAMA_GRETYPE_CHAR_NOT:       return true;
@@ -304,7 +304,7 @@ namespace grammar_parser {
        }
    }

-    void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
+    static void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
        for (auto elem : rule) {
            switch (elem.type) {
                case LLAMA_GRETYPE_END:            fprintf(file, "END");            break;
@@ -334,7 +334,7 @@ namespace grammar_parser {
        fprintf(file, "\n");
    }

-    void print_rule(
+    static void print_rule(
            FILE     * file,
            uint32_t   rule_id,
            const std::vector<llama_grammar_element> & rule,
--- a/common/train.cpp
+++ b/common/train.cpp
--- a/common/train.h
+++ b/common/train.h
@@ -0,0 +1,230 @@
+// Various helper functions and utilities for training
+
+#pragma once
+
+#include <string>
+#include <random>
+#include <vector>
+
+#include "ggml.h"
+#include "llama.h"
+
+typedef std::string mt19937_state;
+
+struct train_state {
+    struct ggml_opt_context * opt;
+
+    uint64_t train_its;
+    uint64_t train_samples;
+    uint64_t train_tokens;
+    uint64_t train_epochs;
+
+    size_t        shuffle_samples_hash; // fn, sample_count, *zip(sample_begins, sample_sizes)
+    mt19937_state shuffle_rng_state_current;
+    mt19937_state shuffle_rng_state_next;
+    size_t        shuffle_sample_count;
+    size_t        shuffle_next_sample;
+};
+
+struct train_params_common {
+    const char * fn_train_data;
+    const char * fn_checkpoint_in;
+    const char * fn_checkpoint_out;
+    const char * pattern_fn_it;
+    const char * fn_latest;
+
+    bool print_usage;
+
+    int save_every;
+
+    uint32_t seed;
+
+    int n_ctx;
+    int n_threads;
+    int n_batch;
+    int n_gradient_accumulation;
+    int n_epochs;
+
+    bool custom_n_ctx;
+
+    bool use_flash;
+    bool use_checkpointing;
+
+    std::string sample_start;
+    bool include_sample_start;
+    bool escape;
+    bool overlapping_samples;
+    bool fill_with_next_samples;
+    bool separate_with_eos;
+    bool separate_with_bos;
+    bool sample_random_offsets;
+
+    bool force_reshuffle;
+
+    int   warmup;
+    int   cos_decay_steps;
+    float cos_decay_restart;
+    float cos_decay_min;
+    bool  enable_restart;
+
+    int   opt_past;
+    float opt_delta;
+    int   opt_max_no_improvement;
+
+    int   adam_n_iter;
+    float adam_alpha;
+    float adam_min_alpha;
+    float adam_decay;
+    int   adam_decay_min_ndim;
+    float adam_beta1;
+    float adam_beta2;
+    float adam_gclip;
+    float adam_eps_f;
+};
+
+typedef void (*save_train_files_callback)(void * data, struct train_state * train);
+
+struct train_opt_callback_data {
+    struct train_params_common * params;
+    struct train_state         * train;
+    save_train_files_callback    save_cb;
+    void                       * save_data;
+    struct llama_context       * lctx;
+    int                          last_save_iter;
+    llama_token                * tokens_data;
+    size_t                       tokens_size;
+    size_t                     * samples_begin;
+    size_t                     * samples_size;
+    size_t                     * shuffled_samples_offs;
+    size_t                     * shuffled_samples_begin;
+    size_t                     * shuffled_samples_size;
+    size_t                       samples_count;
+    struct ggml_tensor         * tokens_input;
+    struct ggml_tensor         * target_probs;
+    int                          first_iter;
+    int                          first_epoch;
+    int                          iter_at_last_epoch;
+    int64_t                      last_time;
+    double                       millis_per_iter;
+};
+
+struct train_state * init_train_state();
+void free_train_state(struct train_state  * state);
+
+struct train_params_common get_default_train_params_common();
+void print_common_train_usage(int /*argc*/, char ** argv, const struct train_params_common * params);
+
+bool consume_common_train_arg(int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param);
+void finish_processing_train_args(struct train_params_common * params);
+
+struct random_normal_distribution;
+struct random_uniform_distribution;
+
+struct random_normal_distribution  * init_random_normal_distribution (int seed, float mean, float std, float min, float max);
+struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max);
+
+void free_random_normal_distribution (struct random_normal_distribution  * rnd);
+void free_random_uniform_distribution(struct random_uniform_distribution * rnd);
+
+struct ggml_tensor * randomize_tensor_normal (struct ggml_tensor * tensor, struct random_normal_distribution * rnd);
+struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd);
+
+// generate random float in interval [0,1)
+float frand();
+float frand_normal (struct random_normal_distribution * rnd);
+float frand_uniform(struct random_uniform_distribution * rnd);
+
+int   clamp (const int v, const int min, const int max);
+float fclamp(const float v, const float min, const float max);
+
+void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0);
+void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1);
+void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2);
+void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3);
+
+size_t tokenize_file(
+        struct llama_context     * lctx,
+        const char               * filename,
+        const std::string        & sample_start,
+        bool                       include_sample_start,
+        bool                       overlapping_samples,
+        unsigned                   context_length,
+        std::vector<llama_token> & out_tokens,
+        std::vector<size_t>      & out_samples_begin,
+        std::vector<size_t>      & out_samples_size);
+
+int64_t get_example_targets_batch(
+        struct llama_context * lctx,
+        struct ggml_tensor   * tokens_input,
+        struct ggml_tensor   * target_probs,
+        int64_t                example_id,
+        const size_t         * samples_offs,
+        const size_t         * samples_begin,
+        const size_t         * samples_size,
+              size_t           samples_count,
+        const llama_token    * train_data,
+        size_t                 n_train_data,
+        bool                   separate_with_eos,
+        bool                   separate_with_bos,
+        bool                   fill_with_next_samples,
+        bool                   sample_random_offsets);
+
+
+void          mt19937_set_state(std::mt19937& rng, const mt19937_state& rng_state);
+mt19937_state mt19937_get_state(const std::mt19937& rng);
+mt19937_state mt19937_seed_to_state(unsigned seed);
+
+mt19937_state shuffle_samples(
+        const mt19937_state & rng_state,
+        size_t              * shuffled_offs,
+        size_t              * shuffled_begins,
+        size_t              * shuffled_sizes,
+        const size_t        * begins,
+        const size_t        * sizes,
+        size_t                count);
+
+size_t hash_combine(size_t h1, size_t h2);
+
+size_t compute_samples_hash(
+    const char* fn,
+    const size_t* samples_begin,
+    const size_t* samples_size,
+    size_t sample_count);
+
+
+std::string replace_str(const char * s, const char * needle, const char * replacement);
+
+void print_duration(double milliseconds);
+
+float cosine_decay(
+    int64_t step,
+    int64_t decay_steps,
+    float   minimum);
+
+float cosine_decay_restart(
+    int64_t step,
+    int64_t decay_steps,
+    float   minimum,
+    float   restart_step_mult);
+
+float learning_schedule(
+    int64_t step,
+    int64_t warmup_steps,
+    int64_t decay_steps,
+    float   learning_rate,
+    float   overall_minimum,
+    float   cos_decay_minimum,
+    float   cos_decay_restart_step_mult,
+    bool    enable_restart);
+
+void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name);
+
+void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt);
+void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt);
+
+bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train);
+void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train);
+
+std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration);
+
+void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel);
--- a/convert-baichuan-hf-to-gguf.py
+++ b/convert-baichuan-hf-to-gguf.py
@@ -0,0 +1,304 @@
+#!/usr/bin/env python3
+# HF baichuan --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+import itertools
+import gguf
+import numpy as np
+import torch
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+
+
+if TYPE_CHECKING:
+    from typing import TypeAlias
+
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
+
+# reverse HF permute back to original pth layout
+
+
+def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray:
+    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))
+
+def reverse_hf_permute_part(weights: NDArray, n_part: int, n_head: int, n_head_kv: int| None = None) -> NDArray:
+        r = weights.shape[0] // 3
+        return (reverse_hf_permute(weights[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
+
+def reverse_hf_part(weights: NDArray, n_part: int) -> NDArray:
+        r = weights.shape[0] // 3
+        return weights[r * n_part : r * n_part + r, ...]
+
+def count_model_parts(dir_model: str) -> int:
+    num_parts = 0
+
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+
+    return num_parts
+
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+print("hello print: ",hparams["architectures"][0])
+if hparams["architectures"][0] != "BaichuanForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+print(f"num_parts:{num_parts}\n")
+ARCH=gguf.MODEL_ARCH.BAICHUAN
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["num_hidden_layers"]
+head_count = hparams["num_attention_heads"]
+
+if "num_key_value_heads" in hparams:
+    head_count_kv = hparams["num_key_value_heads"]
+else:
+    head_count_kv = head_count
+
+if "_name_or_path" in hparams:
+    hf_repo = hparams["_name_or_path"]
+else:
+    hf_repo = ""
+
+if "max_sequence_length" in hparams:
+    ctx_length = hparams["max_sequence_length"]
+elif "max_position_embeddings" in hparams:
+    ctx_length = hparams["max_position_embeddings"]
+elif "model_max_length" in hparams:
+    ctx_length = hparams["model_max_length"]
+else:
+    print("gguf: can not find ctx length parameter.")
+
+    sys.exit()
+
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_source_hf_repo(hf_repo)
+gguf_writer.add_tensor_data_layout("Meta AI original pth")
+gguf_writer.add_context_length(ctx_length)
+gguf_writer.add_embedding_length(hparams["hidden_size"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+gguf_writer.add_head_count(head_count)
+gguf_writer.add_head_count_kv(head_count_kv)
+gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+
+if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
+    if "type" in hparams["rope_scaling"]:
+        if hparams["rope_scaling"]["type"] == "linear":
+            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
+
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytes] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+tokenizer_model_file = dir_model / 'tokenizer.model'
+if not tokenizer_model_file.is_file():
+    print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr)
+    sys.exit(1)
+
+# vocab type sentencepiece
+print("gguf: get sentencepiece tokenizer vocab, scores and token types")
+
+tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
+
+for i in range(tokenizer.vocab_size()):
+    text: bytes
+    score: float
+
+    piece = tokenizer.id_to_piece(i)
+    text = piece.encode("utf-8")
+    score = tokenizer.get_score(i)
+
+    toktype = 1  # defualt to normal token type
+    if tokenizer.is_unknown(i):
+        toktype = 2
+    if tokenizer.is_control(i):
+        toktype = 3
+
+    # toktype = 4 is user-defined = tokens from added_tokens.json
+
+    if tokenizer.is_unused(i):
+        toktype = 5
+    if tokenizer.is_byte(i):
+        toktype = 6
+
+    tokens.append(text)
+    scores.append(score)
+    toktypes.append(toktype)
+
+added_tokens_file = dir_model / 'added_tokens.json'
+if added_tokens_file.is_file():
+    with open(added_tokens_file, "r", encoding="utf-8") as f:
+        addtokens_json = json.load(f)
+
+        print("gguf: get added tokens")
+
+        for key in addtokens_json:
+            tokens.append( key.encode("utf-8") )
+            scores.append(-1000.0)
+            toktypes.append(4) # user-defined token type
+
+
+gguf_writer.add_tokenizer_model("llama")
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    tmp=model_part
+    for i in range(block_count):
+        if f"model.layers.{i}.self_attn.W_pack.weight" in model_part:
+            print(f"Unpacking and permuting layer {i}")
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],0,head_count,head_count)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],1,head_count,head_count_kv)
+            tmp[f"model.layers.{i}.self_attn.v_proj.weight"]=reverse_hf_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
+
+    for name in model_part.keys():
+        data = model_part[name]
+        # we don't need these
+        if name.endswith(".rotary_emb.inv_freq"):
+            continue
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if 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 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 ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name + " -> " +  new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")
--- a/convert-falcon-hf-to-gguf.py
+++ b/convert-falcon-hf-to-gguf.py
@@ -55,10 +55,22 @@ def count_model_parts(dir_model: Path) -> int:

 def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
-    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
-    parser.add_argument("--outfile",    type=Path,           help="path to write to; default: based on input")
-    parser.add_argument("model",        type=Path,           help="directory containing model file, or model file itself (*.bin)")
-    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
    return parser.parse_args()

 args = parse_args()
@@ -121,8 +133,6 @@ gguf_writer.add_file_type(ftype)
 print("gguf: get tokenizer metadata")

 tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []

 tokenizer_json_file = dir_model / 'tokenizer.json'
 if not tokenizer_json_file.is_file():
@@ -137,7 +147,9 @@ with open(tokenizer_json_file, "r", encoding="utf-8") as f:

 print("gguf: get gpt2 tokenizer vocab")

-vocab_size = len(tokenizer_json["model"]["vocab"])
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"])

 # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
 tokenizer = AutoTokenizer.from_pretrained(dir_model)
@@ -163,12 +175,8 @@ for i in range(vocab_size):
        text = bytearray(pad_token)

    tokens.append(text)
-    scores.append(0.0)                      # dymmy
-    toktypes.append(gguf.TokenType.NORMAL)  # dummy

 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)

 special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
 special_vocab.add_to_gguf(gguf_writer)
--- a/convert-gptneox-hf-to-gguf.py
+++ b/convert-gptneox-hf-to-gguf.py
@@ -56,10 +56,22 @@ def count_model_parts(dir_model: Path) -> int:

 def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file")
-    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
-    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.bin)")
-    parser.add_argument("ftype",     type=int, choices=[0, 1],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
    return parser.parse_args()

 args = parse_args()
--- a/convert-starcoder-hf-to-gguf.py
+++ b/convert-starcoder-hf-to-gguf.py
@@ -0,0 +1,242 @@
+#!/usr/bin/env python3
+# HF starcoder --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def bytes_to_unicode():
+    # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    return dict(zip(bs, (chr(n) for n in cs)))
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a StarCoder model to a GGML compatible file")
+    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
+    parser.add_argument("--outfile",    type=Path,           help="path to write to; default: based on input")
+    parser.add_argument("model",        type=Path,           help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTBigCodeForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.STARCODER
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("StarCoder")
+gguf_writer.add_context_length(hparams["n_positions"])
+gguf_writer.add_embedding_length(hparams["n_embd"])
+gguf_writer.add_feed_forward_length(4 * hparams["n_embd"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(hparams["n_head"])
+gguf_writer.add_head_count_kv(1)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+
+tokenizer_json_file = dir_model / 'tokenizer.json'
+if not tokenizer_json_file.is_file():
+    print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr)
+    sys.exit(1)
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"])
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
+
+for i in range(vocab_size):
+    if i in reverse_vocab:
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[ord(c)])
+                else:  # multibyte special token character
+                    text.extend(c.encode('utf-8'))
+    else:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
+
+    tokens.append(text)
+
+gguf_writer.add_token_list(tokens)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# params for qkv transform
+n_head    = hparams["n_head"]
+n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+
+head_dim = hparams["n_embd"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if 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 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 ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")
--- a/convert.py
+++ b/convert.py
@@ -145,7 +145,6 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
 class Params:
    n_vocab:    int
    n_embd:     int
-    n_mult:     int
    n_layer:    int
    n_ctx:      int
    n_ff:       int
@@ -161,15 +160,6 @@ class Params:
    # path to the directory containing the model files
    path_model: Path | None = None

-    @staticmethod
-    def find_n_mult(n_ff: int, n_embd: int) -> int:
-        # hardcoded magic range
-        for n_mult in range(8192, 1, -1):
-            calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
-            if calc_ff == n_ff:
-                return n_mult
-        raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).")
-
    @staticmethod
    def guessed(model: LazyModel) -> Params:
        # try transformer naming first
@@ -197,7 +187,6 @@ class Params:
        return Params(
            n_vocab    = n_vocab,
            n_embd     = n_embd,
-            n_mult     = n_mult,
            n_layer    = n_layer,
            n_ctx      = -1,
            n_ff       = n_ff,
@@ -225,8 +214,6 @@ class Params:
        else:
            f_rope_scale = None

-        n_mult = Params.find_n_mult(n_ff, n_embd)
-
        if "max_sequence_length" in config:
            n_ctx = config["max_sequence_length"]
        elif "max_position_embeddings" in config:
@@ -238,7 +225,6 @@ class Params:
        return Params(
            n_vocab          = n_vocab,
            n_embd           = n_embd,
-            n_mult           = n_mult,
            n_layer          = n_layer,
            n_ctx            = n_ctx,
            n_ff             = n_ff,
@@ -250,7 +236,7 @@ class Params:
        )

    # LLaMA v2 70B params.json
-    # {"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
+    # {"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))
@@ -258,7 +244,6 @@ class Params:
        n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
        n_embd           = config["dim"]
        n_layer          = config["n_layers"]
-        n_mult           = config["multiple_of"]
        n_ff             = -1
        n_head           = config["n_heads"]
        n_head_kv        = config["n_kv_heads"] if "n_kv_heads" in config else n_head
@@ -285,7 +270,6 @@ class Params:
        return Params(
            n_vocab          = n_vocab,
            n_embd           = n_embd,
-            n_mult           = n_mult,
            n_layer          = n_layer,
            n_ctx            = n_ctx,
            n_ff             = n_ff,
@@ -455,7 +439,7 @@ Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
    #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
    if n_head_kv is not None and n_head != n_head_kv:
-        n_head //= n_head_kv
+        n_head = n_head_kv
    return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
                .swapaxes(1, 2)
                .reshape(weights.shape))
--- a/docs/BLIS.md
+++ b/docs/BLIS.md
@@ -48,7 +48,7 @@ make -j
 According to the BLIS documentation, we could set the following
 environment variables to modify the behavior of openmp:

-```
+```bash
 export GOMP_GPU_AFFINITY="0-19"
 export BLIS_NUM_THREADS=14
 ```
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -21,9 +21,12 @@ else()
    add_subdirectory(benchmark)
    add_subdirectory(baby-llama)
    add_subdirectory(train-text-from-scratch)
+    add_subdirectory(finetune)
    add_subdirectory(convert-llama2c-to-ggml)
    add_subdirectory(simple)
+    add_subdirectory(batched)
    add_subdirectory(speculative)
+    add_subdirectory(parallel)
    add_subdirectory(embd-input)
    add_subdirectory(llama-bench)
    add_subdirectory(beam-search)
@@ -33,4 +36,5 @@ else()
    if (LLAMA_BUILD_SERVER)
        add_subdirectory(server)
    endif()
+    add_subdirectory(export-lora)
 endif()
--- a/examples/baby-llama/baby-llama.cpp
+++ b/examples/baby-llama/baby-llama.cpp
@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "train.h"
 #include <vector>
 #include <cassert>
 #include <random>
@@ -9,35 +10,12 @@
 #endif

 #ifdef LLAMA_DEFAULT_RMS_EPS
-static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
+constexpr float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
 #else
-static const float rms_norm_eps = 5e-6f;
+constexpr float rms_norm_eps = 5e-6f;
 #endif

-float frand() {
-    return (float)rand()/(float)RAND_MAX;
-}
-
-struct random_normal_distribution {
-    std::mt19937 gen;
-    std::normal_distribution<float> nd;
-    float min;
-    float max;
-};
-
-void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) {
-    rnd->gen = std::mt19937(seed);
-    rnd->nd = std::normal_distribution<float>{mean, std};
-    rnd->min = min;
-    rnd->max = max;
-}
-
-float frand_normal(struct random_normal_distribution * rnd) {
-    const float r = rnd->nd(rnd->gen);
-    return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r);
-}
-
-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);

    if (plan.work_size > 0) {
@@ -48,13 +26,9 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
    ggml_graph_compute(graph, &plan);
 }

-struct ggml_tensor * randomize_tensor(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        float fmin,
-        float fmax) {
-
+static struct ggml_tensor * randomize_tensor(
+    struct ggml_tensor * tensor, int ndims, const int64_t ne[], float fmin, float fmax
+) {
    switch (ndims) {
        case 1:
            for (int i0 = 0; i0 < ne[0]; i0++) {
@@ -95,56 +69,6 @@ struct ggml_tensor * randomize_tensor(
    return tensor;
 }

-struct ggml_tensor * randomize_tensor_normal(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        struct random_normal_distribution * rnd) {
-    float scale = 1.0; // xavier
-    switch (ndims) {
-        case 1:
-            scale /= sqrtf(ne[0]);
-            for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
-            }
-            break;
-        case 2:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i1 = 0; i1 < ne[1]; i1++) {
-                for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
-                }
-            }
-            break;
-        case 3:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i2 = 0; i2 < ne[2]; i2++) {
-                for (int i1 = 0; i1 < ne[1]; i1++) {
-                    for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
-                    }
-                }
-            }
-            break;
-        case 4:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i3 = 0; i3 < ne[3]; i3++) {
-                for (int i2 = 0; i2 < ne[2]; i2++) {
-                    for (int i1 = 0; i1 < ne[1]; i1++) {
-                        for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    };
-
-    return tensor;
-}
-
 struct llama_hparams {
    uint32_t n_vocab = 32000;
    uint32_t n_ctx   = 512;   // this is provided as user input?
@@ -159,7 +83,7 @@ struct llama_hparams {
    }
 };

-uint32_t get_n_ff(const struct llama_hparams* hparams) {
+static uint32_t get_n_ff(const struct llama_hparams* hparams) {
    const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
    return n_ff;
 }
@@ -260,7 +184,7 @@ struct llama_model_lora {
    std::vector<llama_layer_lora> layers;
 };

-void init_model(struct llama_model * model) {
+static void init_model(struct llama_model * model) {
    const auto & hparams = model->hparams;

    const uint32_t n_embd  = hparams.n_embd;
@@ -297,7 +221,7 @@ void init_model(struct llama_model * model) {
 }


-void init_model_lora(struct llama_model_lora * model) {
+static void init_model_lora(struct llama_model_lora * model) {
    const auto & hparams = model->hparams;

    const uint32_t n_embd  = hparams.n_embd;
@@ -340,7 +264,7 @@ void init_model_lora(struct llama_model_lora * model) {
    }
 }

-void set_param_model(struct llama_model * model) {
+static void set_param_model(struct llama_model * model) {
    const auto& hparams = model->hparams;

    const uint32_t n_layer = hparams.n_layer;
@@ -366,7 +290,7 @@ void set_param_model(struct llama_model * model) {
    }
 }

-void set_param_model_lora(struct llama_model_lora * model) {
+static void set_param_model_lora(struct llama_model_lora * model) {
    const auto& hparams = model->hparams;

    const uint32_t n_layer = hparams.n_layer;
@@ -397,69 +321,75 @@ void set_param_model_lora(struct llama_model_lora * model) {
    }
 }

-void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
+static void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
    const auto & hparams = model->hparams;

    const uint32_t n_layer = hparams.n_layer;

-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
-    randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd);
-    randomize_tensor_normal(model->norm,           model->norm->n_dims,           model->norm->ne,           &rnd);
-    randomize_tensor_normal(model->output,         model->output->n_dims,         model->output->ne,         &rnd);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(model->tok_embeddings , rnd);
+    randomize_tensor_normal(model->norm           , rnd);
+    randomize_tensor_normal(model->output         , rnd);

    for (uint32_t i = 0; i < n_layer; ++i) {
        auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);

-        randomize_tensor_normal(layer.wq, layer.wq->n_dims, layer.wq->ne, &rnd);
-        randomize_tensor_normal(layer.wk, layer.wk->n_dims, layer.wk->ne, &rnd);
-        randomize_tensor_normal(layer.wv, layer.wv->n_dims, layer.wv->ne, &rnd);
-        randomize_tensor_normal(layer.wo, layer.wo->n_dims, layer.wo->ne, &rnd);
+        randomize_tensor_normal(layer.wq, rnd);
+        randomize_tensor_normal(layer.wk, rnd);
+        randomize_tensor_normal(layer.wv, rnd);
+        randomize_tensor_normal(layer.wo, rnd);

-        randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd);
+        randomize_tensor_normal(layer.ffn_norm, rnd);

-        randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd);
-        randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd);
-        randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd);
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
    }
+
+    free_random_normal_distribution(rnd);
 }


-void randomize_model_lora(struct llama_model_lora * model, int seed, float mean, float std, float min, float max) {
+static void randomize_model_lora(
+    struct llama_model_lora * model, int seed, float mean, float std, float min, float max
+) {
    const auto & hparams = model->hparams;

    const uint32_t n_layer = hparams.n_layer;

-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
-    randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd);
-    randomize_tensor_normal(model->norm,           model->norm->n_dims,           model->norm->ne,           &rnd);
-    randomize_tensor_normal(model->outputa,        model->outputa->n_dims,        model->outputa->ne,         &rnd);
-    randomize_tensor_normal(model->outputb,        model->outputb->n_dims,        model->outputb->ne,         &rnd);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(model->tok_embeddings, rnd);
+    randomize_tensor_normal(model->norm          , rnd);
+    randomize_tensor_normal(model->outputa       , rnd);
+    randomize_tensor_normal(model->outputb       , rnd);

    for (uint32_t i = 0; i < n_layer; ++i) {
        auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);

-        randomize_tensor_normal(layer.wqa, layer.wqa->n_dims, layer.wqa->ne, &rnd);
-        randomize_tensor_normal(layer.wqb, layer.wqb->n_dims, layer.wqb->ne, &rnd);
-        randomize_tensor_normal(layer.wka, layer.wka->n_dims, layer.wka->ne, &rnd);
-        randomize_tensor_normal(layer.wkb, layer.wkb->n_dims, layer.wkb->ne, &rnd);
-        randomize_tensor_normal(layer.wva, layer.wva->n_dims, layer.wva->ne, &rnd);
-        randomize_tensor_normal(layer.wvb, layer.wvb->n_dims, layer.wvb->ne, &rnd);
-        randomize_tensor_normal(layer.woa, layer.woa->n_dims, layer.woa->ne, &rnd);
-        randomize_tensor_normal(layer.wob, layer.wob->n_dims, layer.wob->ne, &rnd);
+        randomize_tensor_normal(layer.wqa, rnd);
+        randomize_tensor_normal(layer.wqb, rnd);
+        randomize_tensor_normal(layer.wka, rnd);
+        randomize_tensor_normal(layer.wkb, rnd);
+        randomize_tensor_normal(layer.wva, rnd);
+        randomize_tensor_normal(layer.wvb, rnd);
+        randomize_tensor_normal(layer.woa, rnd);
+        randomize_tensor_normal(layer.wob, rnd);

-        randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd);
+        randomize_tensor_normal(layer.ffn_norm, rnd);

-        randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd);
-        randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd);
-        randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd);
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
    }
+
+    free_random_normal_distribution(rnd);
 }

-bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
+static bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
    const auto & hparams = model->hparams;

    const uint32_t n_ctx   = hparams.n_ctx;
@@ -495,7 +425,7 @@ bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int
    return true;
 }

-bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
+static bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
    const auto & hparams = model->hparams;

    const uint32_t n_ctx   = hparams.n_ctx;
@@ -531,15 +461,15 @@ bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora *
    return true;
 }

-struct ggml_tensor * forward(
-        struct llama_model    * model,
-        struct llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past) {
-
+static struct ggml_tensor * forward(
+    struct llama_model    * model,
+    struct llama_kv_cache * cache,
+    struct ggml_context   * ctx0,
+    struct ggml_cgraph    * gf,
+    struct ggml_tensor    * tokens_input,
+    const  int              n_tokens,
+    const  int              n_past
+) {
    const int N = n_tokens;

    struct llama_kv_cache& kv_self = *cache;
@@ -556,6 +486,14 @@ struct ggml_tensor * forward(
    struct ggml_tensor * kc = kv_self.k;
    struct ggml_tensor * vc = kv_self.v;

+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
    // inpL shape [n_embd,N,1,1]
    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
    for (int il = 0; il < n_layer; ++il) {
@@ -583,8 +521,8 @@ struct ggml_tensor * forward(
            // wk   shape [n_embd, n_embd, 1, 1]
            // Qcur shape [n_embd/n_head, n_head, N, 1]
            // Kcur shape [n_embd/n_head, n_head, N, 1]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);

            // store key and value to memory
            {
@@ -756,42 +694,16 @@ struct ggml_tensor * forward(
    return inpL;
 }

-void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
-    GGML_ASSERT(tensor->n_dims == 1);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-}
-
-void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
-    GGML_ASSERT(tensor->n_dims == 2);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-}
-
-void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
-    GGML_ASSERT(tensor->n_dims == 3);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-}
-
-void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
-    GGML_ASSERT(tensor->n_dims == 4);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-    GGML_ASSERT(tensor->ne[3] == ne3);
-}
-
-struct ggml_tensor * forward_batch(
-        struct llama_model    * model,
-        struct llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past,
-        const  int              n_batch) {
-
+static struct ggml_tensor * forward_batch(
+    struct llama_model    * model,
+    struct llama_kv_cache * cache,
+    struct ggml_context   * ctx0,
+    struct ggml_cgraph    * gf,
+    struct ggml_tensor    * tokens_input,
+    const  int              n_tokens,
+    const  int              n_past,
+    const  int              n_batch
+) {
    const int N = n_tokens;

    struct llama_kv_cache& kv_self = *cache;
@@ -810,9 +722,18 @@ struct ggml_tensor * forward_batch(
    struct ggml_tensor * kc = kv_self.k;
    struct ggml_tensor * vc = kv_self.v;

+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
    // inpL shape [n_embd,N*n_batch,1]
    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
    assert_shape_2d(inpL, n_embd, N*n_batch);
+
    for (int il = 0; il < n_layer; ++il) {
        struct ggml_tensor * inpSA = inpL;

@@ -840,8 +761,8 @@ struct ggml_tensor * forward_batch(
            // wk   shape [n_embd, n_embd, 1, 1]
            // Qcur shape [n_embd/n_head, n_head, N, n_batch]
            // Kcur shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
            assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
            assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);

@@ -1073,16 +994,15 @@ struct ggml_tensor * forward_batch(
    return inpL;
 }

-
-struct ggml_tensor * forward_lora(
-        struct llama_model_lora * model,
-        struct llama_kv_cache   * cache,
-        struct ggml_context     * ctx0,
-        struct ggml_cgraph      * gf,
-        struct ggml_tensor      * tokens_input,
-        const  int                n_tokens,
-        const  int                n_past) {
-
+static struct ggml_tensor * forward_lora(
+    struct llama_model_lora * model,
+    struct llama_kv_cache   * cache,
+    struct ggml_context     * ctx0,
+    struct ggml_cgraph      * gf,
+    struct ggml_tensor      * tokens_input,
+    const  int                n_tokens,
+    const  int                n_past
+) {
    const int N = n_tokens;

    struct llama_kv_cache& kv_self = *cache;
@@ -1100,6 +1020,14 @@ struct ggml_tensor * forward_lora(
    struct ggml_tensor * kc = kv_self.k;
    struct ggml_tensor * vc = kv_self.v;

+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
    // inpL shape [n_embd,N,1,1]
    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
    for (int il = 0; il < n_layer; ++il) {
@@ -1133,7 +1061,7 @@ struct ggml_tensor * forward_lora(
                                                        model->layers[il].wqb,
                                                        cur)),
                                                n_embd/n_head, n_head, N),
-                                            n_past, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0, 0);
            struct ggml_tensor * Kcur = ggml_rope(ctx0,
                                            ggml_reshape_3d(ctx0,
                                                ggml_mul_mat(ctx0,
@@ -1142,7 +1070,7 @@ struct ggml_tensor * forward_lora(
                                                        model->layers[il].wkb,
                                                        cur)),
                                                n_embd/n_head, n_head, N),
-                                            n_past, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0, 0);

            // store key and value to memory
            {
@@ -1328,7 +1256,7 @@ struct ggml_tensor * forward_lora(
    return inpL;
 }

-void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
+static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
    assert(logits->n_dims == 2);
    assert(probs->n_dims == 2);
    assert(best_samples->n_dims == 1);
@@ -1359,7 +1287,10 @@ void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, str
    }
 }

-void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
+static void sample_softmax_batch(
+    struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
+    struct ggml_tensor * best_samples
+) {
    GGML_ASSERT(best_samples->n_dims == 2);
    GGML_ASSERT(logits->n_dims == 3);
    GGML_ASSERT(probs->n_dims == 3);
@@ -1393,7 +1324,7 @@ void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits
    }
 }

-void print_row(struct ggml_tensor * probs, int i) {
+static void print_row(struct ggml_tensor * probs, int i) {
    for (int k = 0; k < probs->ne[0]; ++k) {
        float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
        printf(" %.2f", p);
@@ -1401,7 +1332,7 @@ void print_row(struct ggml_tensor * probs, int i) {
    printf("\n");
 }

-void print_matrix(struct ggml_tensor * probs) {
+static void print_matrix(struct ggml_tensor * probs) {
    assert(probs->n_dims == 2);
    for (int i = 0; i < probs->ne[1]; ++i) {
        for (int k = 0; k < probs->ne[0]; ++k) {
@@ -1412,7 +1343,7 @@ void print_matrix(struct ggml_tensor * probs) {
    }
 }

-void print_token(int token, int n_vocab) {
+static void print_token(int token, int n_vocab) {
    for (int k = 0; k < token; ++k) {
        printf(" ");
    }
@@ -1423,14 +1354,14 @@ void print_token(int token, int n_vocab) {
    printf("\n");
 }

-void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
+static void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
    for (int i=0; i<tokens->ne[0]; ++i) {
        int token = ggml_get_i32_1d(tokens, i);
        print_token(token, n_vocab);
    }
 }

-void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
+static void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
    int n_tokens = tokens_input->ne[0];
    int n_vocab = targets->ne[0];
    float randomness = 0.0f;
@@ -1451,7 +1382,9 @@ void get_example_targets(int example_id, struct ggml_tensor * tokens_input, stru
    }
 }

-void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
+static void get_example_targets_batch(
+    struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
+) {
    GGML_ASSERT(tokens_input->n_dims == 2);
    GGML_ASSERT(     targets->n_dims == 3);
    int n_tokens = tokens_input->ne[0];
@@ -1474,7 +1407,7 @@ void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct
    }
 }

-void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
+static void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
    int n_tokens = tokens_input->ne[0];
    int n_vocab = targets->ne[0];
    for (int i=0; i<n_tokens-n_shift; ++i) {
@@ -1485,12 +1418,16 @@ void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * tar
    }
 }

-struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+static struct ggml_tensor * square_error_loss(
+    struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
+) {
    // todo: instead of a-b: a[1:]-b[:-1]
    return ggml_sum(ctx, ggml_sqr(ctx, ggml_sub(ctx, a, b)));
 }

-struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+static struct ggml_tensor * cross_entropy_loss(
+    struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
+) {
    const float eps = 1e-3f;
    return
        ggml_sum(ctx,
--- a/examples/batched/CMakeLists.txt
+++ b/examples/batched/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET batched)
+add_executable(${TARGET} batched.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
--- a/examples/batched/README.md
+++ b/examples/batched/README.md
@@ -0,0 +1,44 @@
+# llama.cpp/example/batched
+
+The example demonstrates batched generation from a given prompt
+
+```bash
+./batched ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is" 4
+
+...
+
+main: n_len = 32, n_ctx = 2048, n_parallel = 4, n_kv_req = 113
+
+ Hello my name is
+
+main: generating 4 sequences ...
+
+main: stream 0 finished
+main: stream 1 finished
+main: stream 2 finished
+main: stream 3 finished
+
+sequence 0:
+
+Hello my name is Shirley. I am a 25-year-old female who has been working for over 5 years as a b
+
+sequence 1:
+
+Hello my name is Renee and I'm a 32 year old female from the United States. I'm looking for a man between
+
+sequence 2:
+
+Hello my name is Diana. I am looking for a housekeeping job. I have experience with children and have my own transportation. I am
+
+sequence 3:
+
+Hello my name is Cody. I am a 3 year old neutered male. I am a very friendly cat. I am very playful and
+
+main: decoded 108 tokens in 3.57 s, speed: 30.26 t/s
+
+llama_print_timings:        load time =   587.00 ms
+llama_print_timings:      sample time =     2.56 ms /   112 runs   (    0.02 ms per token, 43664.72 tokens per second)
+llama_print_timings: prompt eval time =  4089.11 ms /   118 tokens (   34.65 ms per token,    28.86 tokens per second)
+llama_print_timings:        eval time =     0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_print_timings:       total time =  4156.04 ms
+```
--- a/examples/batched/batched.cpp
+++ b/examples/batched/batched.cpp
@@ -0,0 +1,246 @@
+#include "common.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
+        return 1 ;
+    }
+
+    int n_parallel = 1;
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        params.prompt = argv[2];
+    }
+
+    if (argc >= 4) {
+        n_parallel = std::atoi(argv[3]);
+    }
+
+    if (params.prompt.empty()) {
+        params.prompt = "Hello my name is";
+    }
+
+    // total length of the sequences including the prompt
+    const int n_len = 32;
+
+    // init LLM
+
+    llama_backend_init(params.numa);
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = n_len*n_parallel; // FIXME: use n_kv_req instead (tokenize with model after #3301)
+    ctx_params.n_batch = std::max(n_len, n_parallel);
+    // ctx_params.n_gpu_layers = 99; // offload all layers to the GPU
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    // tokenize the prompt
+
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize(ctx, params.prompt, true);
+
+    const int n_ctx    = llama_n_ctx(ctx);
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
+
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+
+    // make sure the KV cache is big enough to hold all the prompt and generated tokens
+    if (n_kv_req > n_ctx) {
+        LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        return 1;
+    }
+
+    // print the prompt token-by-token
+
+    fprintf(stderr, "\n");
+
+    for (auto id : tokens_list) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+    }
+
+    fflush(stderr);
+
+    // create a llama_batch with size 512
+    // we use this object to submit token data for decoding
+
+    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0);
+
+    // evaluate the initial prompt
+    batch.n_tokens = tokens_list.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
+    }
+
+    // llama_decode will output logits only for the last token of the prompt
+    batch.logits[batch.n_tokens - 1] = true;
+
+    if (llama_decode(ctx, batch, params.n_threads) != 0) {
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        return 1;
+    }
+
+    // assign the system KV cache to all parallel sequences
+    // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
+    for (int32_t i = 1; i < n_parallel; ++i) {
+        llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens);
+    }
+
+    if (n_parallel > 1) {
+        LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
+    }
+
+    // main loop
+
+    // we will store the parallel decoded sequences in this vector
+    std::vector<std::string> streams(n_parallel);
+
+    // remember the batch index of the last token for each parallel sequence
+    // we need this to determine which logits to sample from
+    std::vector<int32_t> i_batch(n_parallel, batch.n_tokens - 1);
+
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;
+
+    const auto t_main_start = ggml_time_us();
+
+    while (n_cur <= n_len) {
+        // prepare the next batch
+        batch.n_tokens = 0;
+
+        // sample the next token for each parallel sequence / stream
+        for (int32_t i = 0; i < n_parallel; ++i) {
+            if (i_batch[i] < 0) {
+                // the stream has already finished
+                continue;
+            }
+
+            auto   n_vocab = llama_n_vocab(ctx);
+            auto * logits  = llama_get_logits_ith(ctx, i_batch[i]);
+
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            const int   top_k = 40;
+            const float top_p = 0.9f;
+            const float temp  = 0.4f;
+
+            llama_sample_top_k(ctx, &candidates_p, top_k, 1);
+            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+            llama_sample_temp (ctx, &candidates_p, temp);
+
+            const llama_token new_token_id = llama_sample_token(ctx, &candidates_p);
+
+            //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream? -> mark the stream as finished
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+                i_batch[i] = -1;
+                LOG_TEE("\n");
+                if (n_parallel > 1) {
+                    LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
+                }
+
+                continue;
+            }
+
+            // if there is only one stream, we print immediately to stdout
+            if (n_parallel == 1) {
+                LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+                fflush(stdout);
+            }
+
+            streams[i] += llama_token_to_piece(ctx, new_token_id);
+
+            // push this new token for next evaluation
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = i;
+            batch.logits[batch.n_tokens] = true;
+
+            i_batch[i] = batch.n_tokens;
+
+            batch.n_tokens += 1;
+
+            n_decode += 1;
+        }
+
+        // all streams are finished
+        if (batch.n_tokens == 0) {
+            break;
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch, params.n_threads)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            return 1;
+        }
+    }
+
+    LOG_TEE("\n");
+
+    if (n_parallel > 1) {
+        LOG_TEE("\n");
+
+        for (int32_t i = 0; i < n_parallel; ++i) {
+            LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
+        }
+    }
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}
--- a/examples/beam-search/CMakeLists.txt
+++ b/examples/beam-search/CMakeLists.txt
@@ -3,6 +3,3 @@ add_executable(${TARGET} beam-search.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
--- a/examples/beam-search/beam-search.cpp
+++ b/examples/beam-search/beam-search.cpp
@@ -1,6 +1,5 @@
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"

 #include <cassert>
 #include <cinttypes>
@@ -30,7 +29,8 @@ struct ostream_beam_view {
    llama_context * ctx;
    llama_beam_view beam_view;
 };
-std::ostream& operator<<(std::ostream& os, const ostream_beam_view & obv) {
+
+static std::ostream & operator<<(std::ostream & os, const ostream_beam_view & obv) {
    os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens(";
    for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) {
        os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]);
@@ -46,7 +46,7 @@ struct beam_search_callback_data {

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

@@ -56,7 +56,7 @@ bool is_at_eob(const beam_search_callback_data & callback_data, const llama_toke
 //  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
 //    This is also called when the stop condition is met.
 //    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
-void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
+static void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
    auto& callback_data = *static_cast<beam_search_callback_data*>(callback_data_ptr);
    // Mark beams as EOS as needed.
    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
@@ -158,8 +158,9 @@ int main(int argc, char ** argv)
    }
    std::cout << std::flush;

-    int n_past = llama_get_kv_cache_token_count(ctx);
-    if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past, params.n_threads))
+    int n_past = 0;
+
+    if (llama_decode(ctx, llama_batch_get_one(tokens_list.data(), tokens_list.size(), n_past, 0), params.n_threads))
    {
        fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
        return 1;
--- a/examples/benchmark/CMakeLists.txt
+++ b/examples/benchmark/CMakeLists.txt
@@ -1,7 +1,8 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
  add_dependencies(${TARGET} BUILD_INFO)
--- a/examples/benchmark/benchmark-matmult.cpp
+++ b/examples/benchmark/benchmark-matmult.cpp
@@ -1,5 +1,6 @@
-#include "ggml.h"
 #include "build-info.h"
+#include "common.h"
+#include "ggml.h"

 #include <locale.h>
 #include <assert.h>
@@ -20,7 +21,7 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif

-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);

    if (plan.work_size > 0) {
@@ -31,19 +32,19 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
    ggml_graph_compute(graph, &plan);
 }

-float tensor_sum_elements(const ggml_tensor * tensor) {
-    float sum = 0;
-    if (tensor->type==GGML_TYPE_F32) {
+static float tensor_sum_elements(const ggml_tensor * tensor) {
+    double sum = 0;
+    if (tensor->type == GGML_TYPE_F32) {
        for (int j = 0; j < tensor->ne[1]; j++) {
            for (int k = 0; k < tensor->ne[0]; k++) {
-                sum +=  ((float *) tensor->data)[j*tensor->ne[0]+k];
+                sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
            }
        }
    }
    return sum;
 }

-void tensor_dump(const ggml_tensor * tensor, const char * name) {
+static void tensor_dump(const ggml_tensor * tensor, const char * name) {
    printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
        tensor->type, ggml_type_name(tensor->type),
        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
@@ -58,7 +59,7 @@ struct benchmark_params_struct {
    int32_t n_iterations  = 10;
 };

-void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
+static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
    fprintf(stderr, "usage: %s [options]\n", argv[0]);
    fprintf(stderr, "\n");
    fprintf(stderr, "options:\n");
@@ -99,7 +100,7 @@ int main(int argc, char ** argv)  {
        exit(1);
    }

-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
    printf("Starting Test\n");

    // create the ggml context
@@ -125,12 +126,15 @@ int main(int argc, char ** argv)  {

    //printf("Memsize required = %i\n", sizex*sizex);

+    // TODO: perform the bench for all types or for a user specified type
+    const ggml_type qtype = GGML_TYPE_Q4_1;
+
    size_t ctx_size = 0;
    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
    ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
    ctx_size += 1024*1024*16;
@@ -163,7 +167,7 @@ int main(int argc, char ** argv)  {
    struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
    ggml_set_f32(m2, 2.0f);

-    printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n");
+    printf("\n------ Test 1 - Matrix Mult via F32 code\n");
    // printf("Creating new tensor m11xm2\n");
    struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);

@@ -181,17 +185,16 @@ int main(int argc, char ** argv)  {

    TENSOR_DUMP(gf.nodes[0]);

-    printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n");
+    printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));

    int32_t nelements = sizex*sizey;
-    int32_t ne[2] = { sizex, sizey };

    std::vector<int64_t> hist_cur(1 << 4, 0);

    // Set up a the benchmark matrices
    // printf("Creating new tensor q11 & Running quantize\n");
-    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
-    ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data());
+    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements, hist_cur.data());

    // Set up a the compute graph
    // printf("Creating new tensor q31\n");
@@ -202,8 +205,8 @@ int main(int argc, char ** argv)  {

    // Set up a second graph computation to make sure we override the CPU cache lines
    // printf("Creating new tensor q12 & Running quantize\n");
-    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
-    ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data());
+    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements, hist_cur.data());

    // printf("Creating new tensor q32\n");
    struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
@@ -220,7 +223,7 @@ int main(int argc, char ** argv)  {
    printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);


-    // Let's use the F32 result from above as a reference for the q4_0 multiplication
+    // Let's use the F32 result from above as a reference for the quantized multiplication
    float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);

    printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
@@ -250,7 +253,7 @@ int main(int argc, char ** argv)  {
        // Check that the matrix multiplication result is in the right ballpark
        // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
        float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
-        float delta = abs(sum_of_Q4_result - sum_of_F32_reference);
+        float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
        float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6

        if (delta > allowed_delta)  {
--- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
+++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
@@ -115,7 +115,7 @@ struct TransformerWeights {
    }
 };

-void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
+static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
    // we calloc instead of malloc to keep valgrind happy
    w->token_embedding_table = new float[p->vocab_size * p->dim]();
    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
@@ -158,7 +158,7 @@ void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
    }
 }

-int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
+static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
    if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
    if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
    if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
@@ -189,7 +189,7 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shar
    return 0;
 }

-void print_sample_weights(TransformerWeights *w){
+static void print_sample_weights(TransformerWeights *w){
    printf("----- Quick print of first of the weight vales of all the variables\n");
    printf("%f\n", w->token_embedding_table[0]);
    printf("%f\n", w->rms_att_weight[0]);
@@ -324,7 +324,7 @@ struct train_params {
    int mem_compute1_gb;
 };

-void print_params(struct my_llama_hparams * params) {
+static void print_params(struct my_llama_hparams * params) {
    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
@@ -335,7 +335,7 @@ void print_params(struct my_llama_hparams * params) {
    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
 }

-void init_model(struct my_llama_model * model) {
+static void init_model(struct my_llama_model * model) {
    const auto & hparams = model->hparams;

    const uint32_t n_embd  = hparams.n_embd;
@@ -408,17 +408,17 @@ void init_model(struct my_llama_model * model) {
    }
 }

-float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
    float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
    return *ptr;
 }

-int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
    int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
    return *ptr;
 }

-void print_row(struct ggml_tensor * probs, int i) {
+static void print_row(struct ggml_tensor * probs, int i) {
    for (int k = 0; k < probs->ne[0]; ++k) {
        float p = get_f32_2d(probs, k, i);
        printf(" %f", p);
@@ -426,7 +426,7 @@ void print_row(struct ggml_tensor * probs, int i) {
    printf("\n");
 }

-void print_matrix(struct ggml_tensor * probs) {
+static void print_matrix(struct ggml_tensor * probs) {
    assert(probs->n_dims == 2);
    for (int i = 0; i < probs->ne[1]; ++i) {
        for (int k = 0; k < probs->ne[0]; ++k) {
@@ -531,7 +531,7 @@ struct llama_file {
    }
 };

-bool is_ggml_file(const char *filename) {
+static bool is_ggml_file(const char * filename) {
    llama_file file(filename, "rb");
    if (file.size < 4) {
        return false;
@@ -540,7 +540,7 @@ bool is_ggml_file(const char *filename) {
    return magic == GGUF_MAGIC;
 }

-static std::string llama_escape_whitespaces(const std::string& text) {
+static std::string llama_escape_whitespaces(const std::string & text) {
    std::ostringstream out;
    for (char c : text) {
        if (c == ' ') out << "\xe2\x96\x81";
@@ -549,7 +549,7 @@ static std::string llama_escape_whitespaces(const std::string& text) {
    return out.str();
 }

-void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
    if (is_ggml_file(filename)) {
        struct ggml_context * ctx_data = NULL;

@@ -637,7 +637,7 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
    }
 }

-void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
+static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
    int ct;
    switch (gg_weights->n_dims){
        case 1:
@@ -673,7 +673,9 @@ void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * kar
    }
 }

-void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) {
+static void save_as_llama_model(
+    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+) {
    // convert AK weights into GG weights one by one.
    // w->token_embedding_table -> model->tok_embeddings
    // float*                   -> struct ggml_tensor
@@ -785,7 +787,7 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
    gguf_free(ctx);
 }

-struct train_params get_default_train_params() {
+static struct train_params get_default_train_params() {
    struct train_params params;
    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
    params.fn_llama2c_output_model = "ak_llama_model.bin";
@@ -835,7 +837,7 @@ struct train_params get_default_train_params() {
    return params;
 }

-void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
+static void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
    fprintf(stderr, "usage: %s [options]\n", argv[0]);
    fprintf(stderr, "\n");
    fprintf(stderr, "options:\n");
@@ -846,7 +848,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params)
    fprintf(stderr, "\n");
 }

-bool params_parse(int argc, char ** argv, struct train_params * params) {
+static bool params_parse(int argc, char ** argv, struct train_params * params) {
    bool invalid_param = false;
    bool reqd_param_found = false;
    std::string arg;
@@ -901,7 +903,7 @@ bool params_parse(int argc, char ** argv, struct train_params * params) {
    return true;
 }

-std::string basename(const std::string &path) {
+static std::string basename(const std::string &path) {
    size_t pos = path.find_last_of("/\\");
    if (pos == std::string::npos) {
        return path;
--- a/examples/embd-input/embd-input-lib.cpp
+++ b/examples/embd-input/embd-input-lib.cpp
@@ -1,3 +1,5 @@
+#include "build-info.h"
+#include "common.h"
 #include "embd-input.h"

 #include <cassert>
@@ -22,7 +24,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
        return nullptr;
    }

-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();

    if (params.seed == LLAMA_DEFAULT_SEED) {
        params.seed = uint32_t(time(NULL));
@@ -78,7 +80,8 @@ bool eval_float(void * model, float * input, int N){
        if (n_eval > n_batch) {
            n_eval = n_batch;
        }
-        if (llama_eval_embd(ctx, (input+i*n_emb), n_eval, n_past, params.n_threads)) {
+        llama_batch batch = {  int32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
+        if (llama_decode(ctx, batch, params.n_threads)) {
            fprintf(stderr, "%s : failed to eval\n", __func__);
            return false;
        }
@@ -99,7 +102,7 @@ bool eval_tokens(void * model, std::vector<llama_token> tokens) {
        if (n_eval > params.n_batch) {
            n_eval = params.n_batch;
        }
-        if (llama_eval(ctx, &tokens[i], n_eval, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&tokens[i], n_eval, n_past, 0), params.n_threads)) {
            fprintf(stderr, "%s : failed to eval\n", __func__);
            return false;
        }
@@ -181,11 +184,11 @@ llama_token sampling_id(struct MyModel* mymodel) {
            if (mirostat == 1) {
                static float mirostat_mu = 2.0f * mirostat_tau;
                const int mirostat_m = 100;
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
            } else if (mirostat == 2) {
                static float mirostat_mu = 2.0f * mirostat_tau;
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
            } else {
                // Temperature sampling
@@ -193,7 +196,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
                llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
                llama_sample_typical(ctx, &candidates_p, typical_p, 1);
                llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                id = llama_sample_token(ctx, &candidates_p);
            }
        }
--- a/examples/embd-input/embd-input.h
+++ b/examples/embd-input/embd-input.h
@@ -3,7 +3,6 @@

 #include "common.h"
 #include "llama.h"
-#include "build-info.h"

 extern "C" {

--- a/examples/embedding/README.md
+++ b/examples/embedding/README.md
@@ -1,3 +1,21 @@
-# embedding
+# llama.cpp/example/embedding

-TODO
+This example demonstrates generate high-dimensional embedding vector of a given text with llama.cpp.
+
+## Quick Start
+
+To get started right away, run the following command, making sure to use the correct path for the model you have:
+
+### Unix-based systems (Linux, macOS, etc.):
+
+```bash
+./embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
+```
+
+### Windows:
+
+```powershell
+embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
+```
+
+The above command will output space-separated float values.
--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -1,6 +1,6 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"

 #include <ctime>

@@ -17,7 +17,7 @@ int main(int argc, char ** argv) {

    params.embedding = true;

-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();

    if (params.seed == LLAMA_DEFAULT_SEED) {
        params.seed = time(NULL);
@@ -78,7 +78,7 @@ int main(int argc, char ** argv) {

    while (!embd_inp.empty()) {
        int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_eval(ctx, embd_inp.data(), n_tokens, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0), params.n_threads)) {
            fprintf(stderr, "%s : failed to eval\n", __func__);
            return 1;
        }
--- a/examples/export-lora/CMakeLists.txt
+++ b/examples/export-lora/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET export-lora)
+add_executable(${TARGET} export-lora.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
--- a/examples/export-lora/README.md
+++ b/examples/export-lora/README.md
@@ -0,0 +1,26 @@
+# export-lora
+
+Apply LORA adapters to base model and export the resulting model.
+
+```
+usage: export-lora [options]
+
+options:
+  -h, --help                         show this help message and exit
+  -m FNAME, --model-base FNAME       model path from which to load base model (default '')
+  -o FNAME, --model-out FNAME        path to save exported model (default '')
+  -l FNAME, --lora FNAME             apply LoRA adapter
+  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S
+  -t N, --threads N                  number of threads to use during computation (default: 4)
+```
+
+For example:
+
+```bash
+./bin/export-lora \
+    -m open-llama-3b-v2-q8_0.gguf \
+    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
+    -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin
+```
+
+Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters.
--- a/examples/export-lora/export-lora.cpp
+++ b/examples/export-lora/export-lora.cpp
@@ -0,0 +1,474 @@
+
+#include "common.h"
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#include <vector>
+#include <string>
+#include <thread>
+
+static const size_t tensor_alignment = 32;
+
+struct lora_info {
+    std::string filename;
+    float scale;
+};
+
+struct export_lora_params {
+    std::string fn_model_base;
+    std::string fn_model_out;
+    std::vector<struct lora_info> lora;
+    int n_threads;
+};
+
+struct lora_data {
+    struct lora_info     info;
+    std::vector<uint8_t> data;
+    struct ggml_context * ctx;
+
+    uint32_t lora_r;
+    uint32_t lora_alpha;
+};
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    bool eof() {
+        return tell() >= size;
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static struct export_lora_params get_default_export_lora_params() {
+    struct export_lora_params result;
+    result.fn_model_base = "";
+    result.fn_model_out  = "";
+    result.n_threads = GGML_DEFAULT_N_THREADS;
+    return result;
+}
+
+static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                         show this help message and exit\n");
+    fprintf(stderr, "  -m FNAME, --model-base FNAME       model path from which to load base model (default '%s')\n", params->fn_model_base.c_str());
+    fprintf(stderr, "  -o FNAME, --model-out FNAME        path to save exported model (default '%s')\n", params->fn_model_out.c_str());
+    fprintf(stderr, "  -l FNAME, --lora FNAME             apply LoRA adapter\n");
+    fprintf(stderr, "  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S\n");
+    fprintf(stderr, "  -t N, --threads N                  number of threads to use during computation (default: %d)\n", params->n_threads);
+}
+
+static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) {
+    bool invalid_param = false;
+    std::string arg;
+    struct export_lora_params default_params = get_default_export_lora_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "-m" || arg == "--model-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_base = argv[i];
+        } else if (arg == "-o" || arg == "--model-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_out = argv[i];
+        } else if (arg == "-l" || arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            lora.scale = 1.0f;
+            params->lora.push_back(lora);
+        } else if (arg == "-s" || arg == "--lora-scaled") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            lora.scale = std::stof(argv[i]);
+            params->lora.push_back(lora);
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_threads = std::stoi(argv[i]);
+            if (params->n_threads <= 0) {
+                params->n_threads = std::thread::hardware_concurrency();
+            }
+        } else {
+            fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str());
+            export_lora_print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+
+    if (params->fn_model_base == default_params.fn_model_base) {
+        fprintf(stderr, "error: please specify a filename for model-base.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (params->fn_model_out == default_params.fn_model_out) {
+        fprintf(stderr, "error: please specify a filename for model-out.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str());
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    return true;
+}
+
+static void free_lora(struct lora_data * lora) {
+    if (lora->ctx != NULL) {
+        ggml_free(lora->ctx);
+    }
+    delete lora;
+}
+
+static struct lora_data * load_lora(struct lora_info * info) {
+    struct lora_data * result = new struct lora_data;
+    result->info = *info;
+    result->ctx = NULL;
+    result->lora_r     = 1;
+    result->lora_alpha = 1;
+
+    struct llama_file file(info->filename.c_str(), "rb");
+    if (file.fp == NULL) {
+        fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n",
+            info->filename.c_str());
+        free_lora(result);
+        return NULL;
+    }
+
+    struct ggml_init_params params_ggml;
+    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_MAX_NODES;
+    params_ggml.mem_buffer = NULL;
+    params_ggml.no_alloc   = true;
+    result->ctx = ggml_init(params_ggml);
+
+    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
+    uint32_t magic   = file.read_u32();
+    if (magic != LLAMA_FILE_MAGIC_LORA) {
+        die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
+    }
+    uint32_t version = file.read_u32();
+    if (version != 1) {
+        die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str());
+    }
+    result->lora_r     = file.read_u32();
+    result->lora_alpha = file.read_u32();
+    // read tensor infos from file
+    std::vector<char> name_buf;
+    std::vector<struct ggml_tensor *> tensors;
+    std::vector<size_t> tensors_offset;
+    size_t total_nbytes_pad = 0;
+    while(!file.eof()) {
+        int64_t ne[4]   = {1,1,1,1};
+        uint32_t n_dims  = file.read_u32();
+        uint32_t namelen = file.read_u32();
+        uint32_t type    = file.read_u32();
+        for (uint32_t k = 0; k < n_dims; ++k) {
+            ne[k] = (int64_t)file.read_u32();
+        }
+        name_buf.clear();
+        name_buf.resize(namelen + 1, '\0');
+        file.read_raw(name_buf.data(), namelen);
+        file.seek((0-file.tell()) & 31, SEEK_CUR);
+        size_t offset = file.tell();
+        struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne);
+        ggml_set_name(tensor, name_buf.data());
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        total_nbytes_pad += nbytes_pad;
+        tensors.push_back(tensor);
+        tensors_offset.push_back(offset);
+        file.seek(nbytes, SEEK_CUR);
+    }
+    // read tensor data
+    result->data.resize(total_nbytes_pad);
+    size_t data_offset = 0;
+    for (size_t i = 0; i < tensors.size(); ++i) {
+        struct ggml_tensor * tensor = tensors[i];
+        size_t offset     = tensors_offset[i];
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        file.seek(offset, SEEK_SET);
+        tensor->data = result->data.data() + data_offset;
+        file.read_raw(tensor->data, nbytes);
+        data_offset += nbytes_pad;
+    }
+    return result;
+}
+
+
+static struct ggml_cgraph * build_graph_lora(
+    struct ggml_context * ctx,
+    struct ggml_tensor * tensor,
+    struct ggml_tensor * lora_a,
+    struct ggml_tensor * lora_b,
+    float scaling
+) {
+    struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
+    if (scaling != 1.0f) {
+        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
+    }
+    struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
+
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+    ggml_build_forward_expand (gf, res);
+    return gf;
+}
+
+static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) {
+    if (lora->ctx == NULL) {
+        return false;
+    }
+    std::string name = ggml_get_name(tensor);
+    std::string name_a = name + std::string(".loraA");
+    std::string name_b = name + std::string(".loraB");
+    struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str());
+    struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str());
+    if (lora_a == NULL || lora_b == NULL) {
+        return false;
+    }
+
+    float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r;
+
+    struct ggml_init_params params;
+    params.mem_size   = GGML_OBJECT_SIZE + GGML_GRAPH_SIZE + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
+    params.mem_buffer = NULL;
+    params.no_alloc   = true;
+    struct ggml_context * ctx = NULL;
+    struct ggml_allocr * alloc = NULL;
+    struct ggml_cgraph * gf = NULL;
+
+    ctx   = ggml_init(params);
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+    size_t alloc_size = ggml_allocr_alloc_graph(alloc, gf);
+    ggml_allocr_free(alloc);
+    ggml_free(ctx);
+
+    static std::vector<uint8_t> data_compute;
+    data_compute.resize(alloc_size + tensor_alignment);
+
+    ctx   = ggml_init(params);
+    alloc = ggml_allocr_new(data_compute.data(), data_compute.size(), tensor_alignment);
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+    ggml_allocr_alloc_graph(alloc, gf);
+    ggml_allocr_free(alloc);
+
+    struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
+    static std::vector<uint8_t> data_work;
+    data_work.resize(cplan.work_size);
+    cplan.work_data = data_work.data();
+
+    ggml_graph_compute(gf, &cplan);
+
+    ggml_free(ctx);
+    return true;
+}
+
+static void export_lora(struct export_lora_params * params) {
+    // load all loras
+    std::vector<struct lora_data *> loras;
+    for (size_t i = 0; i < params->lora.size(); ++i) {
+        struct lora_data * lora = load_lora(&params->lora[i]);
+        if (lora != NULL) {
+            loras.push_back(lora);
+        }
+    }
+    if (loras.size() == 0) {
+        fprintf(stderr, "warning: no lora adapters will be applied.\n");
+    }
+
+    // open input file
+    struct llama_file fin(params->fn_model_base.c_str(), "rb");
+    if (!fin.fp) {
+        die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str());
+    }
+
+    // open base model gguf, read tensors without their data
+    struct ggml_context * ctx_in;
+    struct gguf_init_params params_gguf;
+    params_gguf.no_alloc = true;
+    params_gguf.ctx      = &ctx_in;
+    struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf);
+
+    // create new gguf
+    struct gguf_context * gguf_out = gguf_init_empty();
+
+    // copy meta data from base model: kv and tensors
+    gguf_set_kv(gguf_out, gguf_in);
+    int n_tensors = gguf_get_n_tensors(gguf_in);
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+        gguf_add_tensor(gguf_out, tensor);
+    }
+
+    // create output file
+    struct llama_file fout(params->fn_model_out.c_str(), "wb");
+    if (!fout.fp) {
+        die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str());
+    }
+
+    // write gguf meta data
+    std::vector<uint8_t> meta;
+    meta.resize(gguf_get_meta_size(gguf_out));
+    gguf_get_meta_data(gguf_out, meta.data());
+    fout.write_raw(meta.data(), meta.size());
+
+    std::vector<uint8_t> data;
+    std::vector<uint8_t> padding;
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+
+        // read tensor data
+        data.resize(ggml_nbytes(tensor));
+        tensor->data = data.data();
+        size_t offset = gguf_get_tensor_offset(gguf_in, i);
+        fin.seek(offset + meta.size(), SEEK_SET);
+        fin.read_raw(data.data(), data.size());
+
+        // apply all loras
+        for (size_t k = 0; k < loras.size(); ++k) {
+            apply_lora(tensor, loras[k], params->n_threads);
+        }
+
+        // write tensor data + padding
+        padding.clear();
+        padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0);
+
+        GGML_ASSERT(fout.tell() == offset + meta.size());
+        // fout.seek(offset + meta.size(), SEEK_SET);
+        fout.write_raw(data.data(), data.size());
+        fout.write_raw(padding.data(), padding.size());
+
+        if (i % 2 == 0) {
+            printf(".");
+        }
+    }
+    printf("\n");
+
+    // close gguf
+    gguf_free(gguf_out);
+    gguf_free(gguf_in);
+
+    // free loras
+    for (size_t i = 0; i < loras.size(); ++i) {
+        free_lora(loras[i]);
+    }
+}
+
+int main(int argc, char ** argv) {
+    struct export_lora_params params = get_default_export_lora_params();
+
+    if (!export_lora_params_parse(argc, argv, &params)) {
+        return 1;
+    }
+
+    export_lora(&params);
+
+    return 0;
+}
--- a/examples/finetune/CMakeLists.txt
+++ b/examples/finetune/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET finetune)
+add_executable(${TARGET} finetune.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
--- a/examples/finetune/README.md
+++ b/examples/finetune/README.md
@@ -0,0 +1,90 @@
+# finetune
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
+
+# finetune LORA adapter
+./bin/finetune \
+        --model-base open-llama-3b-v2-q8_0.gguf \
+        --checkpoint-in  chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf \
+        --checkpoint-out chk-lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.gguf \
+        --lora-out lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.bin \
+        --train-data "shakespeare.txt" \
+        --save-every 10 \
+        --threads 6 --adam-iter 30 --batch 4 --ctx 64 \
+        --use-checkpointing
+
+# predict
+./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+```
+
+Finetune output files will be saved every N iterations (config with `--save-every N`).
+The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
+So in above example after 10 iterations these files will be written:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-10.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+After 10 more iterations:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-20.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-20.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+Checkpoint files (`--checkpoint-in FN`, `--checkpoint-out FN`) store the training process. When the input checkpoint file does not exist, it will begin finetuning a new randomly initialized adapter.
+
+llama.cpp compatible LORA adapters will be saved with filename specified by `--lora-out FN`.
+These LORA adapters can then be used by `main` together with the base model, like in the 'predict' example command above.
+
+In `main` you can also load multiple LORA adapters, which will then be mixed together.
+
+For example if you have two LORA adapters `lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin` and `lora-open-llama-3b-v2-q8_0-bible-LATEST.bin`, you can mix them together like this:
+
+```bash
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
+  --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin \
+  --lora lora-open-llama-3b-v2-q8_0-bible-LATEST.bin
+```
+
+You can change how strong each LORA adapter is applied to the base model by using `--lora-scaled FN SCALE` instead of `--lora FN`.
+
+For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' LORA adapter and 100% of yet another one:
+
+```bash
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin 0.4 \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-bible-LATEST.bin 0.8 \
+  --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin
+```
+
+The scale numbers don't need to add up to one, and you can also use numbers creater than 1 to further increase the influence of an adapter. But making the values to big will sometimes result in worse output. Play around to find good values.
+
+Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime.
+If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`.
+
+The default LORA rank can be specified with `--lora-r N`.
+The LORA rank can be configured for each model tensor type separately with these command line options:
+
+```bash
+  --lora-r N                 LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default 4)
+  --rank-att-norm N          LORA rank for attention norm tensor (default 1)
+  --rank-ffn-norm N          LORA rank for feed-forward norm tensor (default 1)
+  --rank-out-norm N          LORA rank for output norm tensor (default 1)
+  --rank-tok-embd N          LORA rank for token embeddings tensor (default 4)
+  --rank-out N               LORA rank for output tensor (default 4)
+  --rank-wq N                LORA rank for wq tensor (default 4)
+  --rank-wk N                LORA rank for wk tensor (default 4)
+  --rank-wv N                LORA rank for wv tensor (default 4)
+  --rank-wo N                LORA rank for wo tensor (default 4)
+  --rank-w1 N                LORA rank for w1 tensor (default 4)
+  --rank-w2 N                LORA rank for w2 tensor (default 4)
+  --rank-w3 N                LORA rank for w3 tensor (default 4)
+```
+
+The LORA rank of 'norm' tensors should always be 1.
+
+To see all available options use `finetune --help`.
--- a/examples/finetune/convert-finetune-checkpoint-to-gguf.py
+++ b/examples/finetune/convert-finetune-checkpoint-to-gguf.py
@@ -0,0 +1,489 @@
+#!/usr/bin/env python3
+# finetune checkpoint --> gguf conversion
+
+import argparse
+import gguf
+import os
+import struct
+import sys
+import numpy as np
+from pathlib import Path
+
+# gguf constants
+LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
+LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam"
+LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs"
+LLM_KV_OPTIMIZER_FILE_VERSION               = "optimizer.file_version"
+LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT     = "optimizer.convergence_past_count"
+LLM_KV_OPTIMIZER_PARAMETER_COUNT            = "optimizer.parameter_count"
+LLM_KV_OPTIMIZER_ITERATION_COUNT            = "optimizer.iteration_count"
+LLM_KV_OPTIMIZER_JUST_INITIALIZED           = "optimizer.just_initialized"
+LLM_KV_OPTIMIZER_ADAM_BEST_LOSS             = "optimizer.adam.best_loss"
+LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS         = "optimizer.adam.previous_loss"
+LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT  = "optimizer.adam.no_improvement_count"
+LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count"
+LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS            = "optimizer.lbfgs.best_loss"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP     = "optimizer.lbfgs.line_search_step"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J        = "optimizer.lbfgs.line_search_j"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K        = "optimizer.lbfgs.line_search_k"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END      = "optimizer.lbfgs.line_search_end"
+LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count"
+
+LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS    = "optimizer.adam.first_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS   = "optimizer.adam.second_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values"
+
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS  = "optimizer.lbfgs.current_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS   = "optimizer.lbfgs.current_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS  = "optimizer.lbfgs.previous_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION    = "optimizer.lbfgs.search_direction"
+LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES    = "optimizer.lbfgs.past_loss_values"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA        = "optimizer.lbfgs.memory_alpha"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y"
+
+LLM_KV_TRAINING_TYPE_TRAIN_MODEL   = "train_model"
+LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"
+LLM_KV_TRAINING_TYPE               = "training.type"
+LLM_KV_TRAINING_FILE_VERSION       = "training.file_version"
+LLM_KV_TRAINING_ITERATION_COUNT    = "training.iteration_count"
+LLM_KV_TRAINING_SAMPLE_COUNT       = "training.sample_count"
+LLM_KV_TRAINING_TOKEN_COUNT        = "training.token_count"
+
+LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD  = "training.lora.rank.token_embd"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT      = "training.lora.rank.output"
+LLM_KV_TRAINING_LORA_RANK_ATTN_NORM   = "training.lora.rank.attn_norm"
+LLM_KV_TRAINING_LORA_RANK_ATTN_Q      = "training.lora.rank.attn_q"
+LLM_KV_TRAINING_LORA_RANK_ATTN_K      = "training.lora.rank.attn_k"
+LLM_KV_TRAINING_LORA_RANK_ATTN_V      = "training.lora.rank.attn_v"
+LLM_KV_TRAINING_LORA_RANK_ATTN_OUT    = "training.lora.rank.attn_output"
+LLM_KV_TRAINING_LORA_RANK_FFN_NORM    = "training.lora.rank.ffn_norm"
+LLM_KV_TRAINING_LORA_RANK_FFN_GATE    = "training.lora.rank.ffn_gate"
+LLM_KV_TRAINING_LORA_RANK_FFN_DOWN    = "training.lora.rank.ffn_down"
+LLM_KV_TRAINING_LORA_RANK_FFN_UP      = "training.lora.rank.ffn_up"
+
+class Tensor:
+    def __init__(self, dtype='f', ne=None):
+        if ne is None:
+            ne = []
+        self.dtype = dtype
+        self.ne = ne
+        self.nbytes = 0
+        if self.dtype == 'f':
+            if len(self.ne) == 0:
+                self.nbytes = 0
+            else:
+                self.nbytes = int(np.product(self.ne)) * 4
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+    def load(self, data, offset):
+        nd = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        namelen = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        dtype = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        assert(nd == len(self.ne))
+        ne = []
+        for d in range(nd):
+            n = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+            ne.append(n)
+
+        if tuple(ne) != tuple(self.ne):
+            raise ValueError(f"Tensor.load: Expected number of elements {str(self.ne)} does not match what is read from file {str(ne)}")
+
+        if self.dtype == 'f':
+            assert(dtype == 0)
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+        self.name = bytes(data[offset:offset+namelen]); offset += namelen
+        # 32-byte alignment
+        offset += (0 - offset) & 31
+        self.data = data[offset:offset+self.nbytes]
+        offset += self.nbytes
+        return offset
+
+    def max_storage_size(self):
+        result = 0
+        result += 4 # nd
+        result += 4 # namelen
+        result += 4 # dtype
+        result += len(self.ne)*8 # ne
+        result += 48 # name (maximum as of commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9)
+        result += 31 # 32-byte alignment
+        result += self.nbytes
+        return result
+
+    def save_gguf(self, gguf_writer, name):
+        gguf_writer.add_tensor(
+            name=name,
+            tensor=self.data,
+            raw_shape=np.array(list(reversed(self.ne))),
+            raw_dtype=gguf.GGMLQuantizationType.F32)
+
+class OptimizationContext:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]
+        offset += 4
+
+        if self.version != 1:
+            raise ValueError('Invalid version of optimization context in checkpoint file')
+
+        self.past    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_m = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.nx      = struct.unpack('N',  bytes(data[offset:offset + 8]))[0];  offset += 8
+        self.iter    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]);  offset += 4
+
+        self.adam_m  = Tensor('f', [self.nx])
+        self.adam_v  = Tensor('f', [self.nx])
+        self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
+
+        self.lbfgs_x    = Tensor('f', [self.nx])
+        self.lbfgs_xp   = Tensor('f', [self.nx])
+        self.lbfgs_g    = Tensor('f', [self.nx])
+        self.lbfgs_gp   = Tensor('f', [self.nx])
+        self.lbfgs_d    = Tensor('f', [self.nx])
+        self.lbfgs_pf   = Tensor('f', [self.past] if self.past > 0 else [])
+        self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lms  = Tensor('f', [self.nx, self.lbfgs_m])
+        self.lbfgs_lmy  = Tensor('f', [self.nx, self.lbfgs_m])
+
+        # forgot to save type in version 1:
+        # guess self.type from number of remaining bytes
+        size_type_0 = 12 + sum([t.max_storage_size() for t in
+                                [self.adam_m, self.adam_v]
+                                +([self.adam_pf] if (self.past > 0) else [])])
+        size_type_1 = 24 + sum([t.max_storage_size() for t in
+                                [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g,
+                                 self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf,
+                                 self.lbfgs_lmal, self.lbfgs_lmys,
+                                 self.lbfgs_lms, self.lbfgs_lmy]
+                                 +([self.lbfgs_pf] if (self.past > 0) else [])])
+        # due to alignment padding the size might not by exact
+        # but the difference in size for both types is significant,
+        # so we can just use whichever is closest
+        remaining = len(data) - offset
+        if abs(remaining - size_type_0) < abs(remaining - size_type_1):
+            self.type = 0
+        else:
+            self.type = 1
+
+        if self.type == 0:
+            offset = self.adam_m.load(data, offset)
+            offset = self.adam_v.load(data, offset)
+            offset = self.adam_pf.load(data,offset)
+
+            self.adam_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_fx_prev          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        elif self.type == 1:
+            offset = self.lbfgs_x.load(data, offset)
+            offset = self.lbfgs_xp.load(data, offset)
+            offset = self.lbfgs_g.load(data, offset)
+            offset = self.lbfgs_gp.load(data, offset)
+            offset = self.lbfgs_d.load(data, offset)
+            offset = self.lbfgs_pf.load(data, offset)
+            offset = self.lbfgs_lmal.load(data, offset)
+            offset = self.lbfgs_lmys.load(data, offset)
+            offset = self.lbfgs_lms.load(data, offset)
+            offset = self.lbfgs_lmy.load(data, offset)
+
+            self.lbfgs_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_step             = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_j                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_k                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_end              = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        else:
+            raise ValueError(f"Invalid optimizer type '{self.type}'")
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_FILE_VERSION, 0)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, self.past)
+        gguf_writer.add_uint64(LLM_KV_OPTIMIZER_PARAMETER_COUNT, self.nx)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ITERATION_COUNT, self.iter)
+        gguf_writer.add_bool(LLM_KV_OPTIMIZER_JUST_INITIALIZED, self.just_initialized)
+
+        if self.type == 0:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, self.adam_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, self.adam_fx_prev)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, self.adam_n_no_improvement)
+
+            self.adam_m.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS)
+            self.adam_v.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS)
+            if self.past > 0:
+                self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES)
+
+        elif self.type == 1:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement)
+
+            self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS)
+            self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS)
+            self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS)
+            self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS)
+            self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION)
+            if self.past > 0:
+                self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES)
+            self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA)
+            self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS)
+            self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S)
+            self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y)
+        else:
+            raise ValueError('Unknown optimizer type')
+
+class LoraParams:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.n_rank_attention_norm  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wq              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wk              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wv              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wo              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_ffn_norm        = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w1              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w2              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w3              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_tok_embeddings  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_norm            = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_output          = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD,  self.n_rank_tok_embeddings)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM, self.n_rank_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT,      self.n_rank_output)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_NORM,   self.n_rank_attention_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_Q,      self.n_rank_wq)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_K,      self.n_rank_wk)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_V,      self.n_rank_wv)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,    self.n_rank_wo)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_NORM,    self.n_rank_ffn_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_GATE,    self.n_rank_w1)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,    self.n_rank_w2)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_UP,      self.n_rank_w3)
+
+class ModelParams:
+    def __init__(self, n_ff = None):
+        self.n_ff = n_ff
+
+    def load(self, data, offset):
+        self.n_vocab = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_embd  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_mult  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_head  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_layer = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rot   = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def get_n_ff(self):
+        if self.n_ff is None:
+            # struct my_llama_model::get_n_ff in train-text-from-scratch.cpp commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9
+            return ((2*(4*self.n_embd)//3 + self.n_mult - 1)//self.n_mult)*self.n_mult
+        else:
+            return self.n_ff
+
+    def save_gguf(self, gguf_writer):
+        # self.n_vocab not saved
+        gguf_writer.add_embedding_length(self.n_embd)
+        gguf_writer.add_head_count(self.n_head)
+        gguf_writer.add_block_count(self.n_layer)
+        gguf_writer.add_rope_dimension_count(self.n_rot)
+        gguf_writer.add_feed_forward_length(self.get_n_ff())
+
+def tensor_name(key, bid=None, suffix=".weight"):
+    return gguf.MODEL_TENSOR_NAMES[gguf.MODEL_ARCH.LLAMA][key].format(bid=bid) + suffix
+
+class Layer:
+    def __init__(self, params, lora_params, bid):
+        self.bid = bid
+        self.att_norm_a = Tensor('f', [lora_params.n_rank_attention_norm, params.n_embd])
+        self.att_norm_b = Tensor('f', [lora_params.n_rank_attention_norm, 1])
+        self.wq_a       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wq_b       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wk_a       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wk_b       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wv_a       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wv_b       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wo_a       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.wo_b       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.ffn_norm_a = Tensor('f', [lora_params.n_rank_ffn_norm, params.n_embd])
+        self.ffn_norm_b = Tensor('f', [lora_params.n_rank_ffn_norm, 1])
+        self.w1_a       = Tensor('f', [lora_params.n_rank_w1, params.n_embd])
+        self.w1_b       = Tensor('f', [lora_params.n_rank_w1, params.get_n_ff()])
+        self.w2_a       = Tensor('f', [lora_params.n_rank_w2, params.get_n_ff()])
+        self.w2_b       = Tensor('f', [lora_params.n_rank_w2, params.n_embd])
+        self.w3_a       = Tensor('f', [lora_params.n_rank_w3, params.n_embd])
+        self.w3_b       = Tensor('f', [lora_params.n_rank_w3, params.get_n_ff()])
+
+    def load(self, data, offset):
+        offset = self.att_norm_a.load(data, offset)
+        offset = self.att_norm_b.load(data, offset)
+        offset = self.wq_a.load(data, offset)
+        offset = self.wq_b.load(data, offset)
+        offset = self.wk_a.load(data, offset)
+        offset = self.wk_b.load(data, offset)
+        offset = self.wv_a.load(data, offset)
+        offset = self.wv_b.load(data, offset)
+        offset = self.wo_a.load(data, offset)
+        offset = self.wo_b.load(data, offset)
+        offset = self.ffn_norm_a.load(data, offset)
+        offset = self.ffn_norm_b.load(data, offset)
+        offset = self.w1_a.load(data, offset)
+        offset = self.w1_b.load(data, offset)
+        offset = self.w2_a.load(data, offset)
+        offset = self.w2_b.load(data, offset)
+        offset = self.w3_a.load(data, offset)
+        offset = self.w3_b.load(data, offset)
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.att_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_a"))
+        self.att_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_b"))
+        self.wq_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_a"))
+        self.wq_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_b"))
+        self.wk_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_a"))
+        self.wk_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_b"))
+        self.wv_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_a"))
+        self.wv_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_b"))
+        self.wo_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_a"))
+        self.wo_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_b"))
+        self.ffn_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_a"))
+        self.ffn_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_b"))
+        self.w1_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_a"))
+        self.w1_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_b"))
+        self.w2_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_a"))
+        self.w2_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_b"))
+        self.w3_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_a"))
+        self.w3_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_b"))
+
+class LoraModel:
+    def __init__(self, n_ff = None):
+        self.params = ModelParams(n_ff = n_ff)
+        self.lora_params = LoraParams()
+        self.layers = []
+
+    def load(self, data, offset):
+        offset = self.params.load(data, offset)
+        offset = self.lora_params.load(data, offset)
+
+        self.tok_embd_a = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_embd])
+        self.tok_embd_b = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_vocab])
+        self.norm_a     = Tensor('f', [self.lora_params.n_rank_norm, self.params.n_embd])
+        self.norm_b     = Tensor('f', [self.lora_params.n_rank_norm, 1])
+        self.output_a   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_embd])
+        self.output_b   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_vocab])
+
+        offset = self.tok_embd_a.load(data, offset)
+        offset = self.tok_embd_b.load(data, offset)
+        offset = self.norm_a.load(data, offset)
+        offset = self.norm_b.load(data, offset)
+        offset = self.output_a.load(data, offset)
+        offset = self.output_b.load(data, offset)
+
+        self.layers.clear()
+        for bid in range(self.params.n_layer):
+            layer = Layer(self.params, self.lora_params, bid)
+            offset = layer.load(data, offset)
+            self.layers.append(layer)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.params.save_gguf(gguf_writer)
+        self.lora_params.save_gguf(gguf_writer)
+
+        self.tok_embd_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_a"))
+        self.tok_embd_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_b"))
+        self.norm_a.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_a"))
+        self.norm_b.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_b"))
+        self.output_a.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_a"))
+        self.output_b.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_b"))
+
+        for layer in self.layers:
+            layer.save_gguf(gguf_writer)
+
+class LoraCheckpoint:
+    def __init__(self, n_ff = None):
+        self.model = LoraModel(n_ff = n_ff)
+        self.opt_ctx = OptimizationContext()
+
+    def load(self, data, offset):
+        magic   = bytes(reversed(data[offset:offset + 4])); offset += 4
+        if magic != b'ggcl':
+            raise ValueError(f"File header magic indicates, that this is no finetune-lora checkpoint file. Expected 'ggcl', Got '{str(magic)}'")
+
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        if self.version != 0:
+            raise ValueError('Invalid version of checkpoint file')
+
+        self.train_its     = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_samples = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_tokens  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        offset = self.model.load(data, offset)
+        offset = self.opt_ctx.load(data, offset)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_file_type(gguf.GGMLQuantizationType.F32)
+        gguf_writer.add_layer_norm_rms_eps(1e-5)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_FILE_VERSION,    0)
+        gguf_writer.add_string(LLM_KV_TRAINING_TYPE,            LLM_KV_TRAINING_TYPE_FINETUNE_LORA)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_ITERATION_COUNT, self.train_its)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_SAMPLE_COUNT,    self.train_samples)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_TOKEN_COUNT,     self.train_tokens)
+        self.model.save_gguf(gguf_writer)
+        self.opt_ctx.save_gguf(gguf_writer)
+
+def handle_args():
+    parser = argparse.ArgumentParser(description = 'Convert finetune checkpoints to GGUF')
+    parser.add_argument('--input',  '-i', type = Path, help = 'Input finetune checkpoint filename', required=True)
+    parser.add_argument('--output', '-o', type = Path, help = 'Output GGUF filename', required=True)
+    parser.add_argument('--ff', type = int, help = "Feedforward size, if not provided compute from n_mult. Provide this if you get 'ValueError: Tensor.load: Expected number of elements does not match what is read from file'", required=False)
+    return parser.parse_args()
+
+def main():
+    cfg = handle_args()
+    print(cfg)
+    data = np.memmap(cfg.input, mode = 'r')
+    chk = LoraCheckpoint(n_ff = cfg.ff)
+    offset = 0
+    offset = chk.load(data, offset)
+    # we should have read all available data
+    assert(offset == len(data))
+
+    gguf_writer = gguf.GGUFWriter(cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+    chk.save_gguf(gguf_writer)
+    print("    gguf: write header")
+    gguf_writer.write_header_to_file()
+    print("    gguf: write metadata")
+    gguf_writer.write_kv_data_to_file()
+    print("    gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+    gguf_writer.close()
+
+if __name__ == '__main__':
+    main()
--- a/examples/finetune/finetune.cpp
+++ b/examples/finetune/finetune.cpp
--- a/examples/gguf/gguf.cpp
+++ b/examples/gguf/gguf.cpp
@@ -13,14 +13,14 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))

-template<typename T>
+template <typename T>
 static std::string to_string(const T & val) {
    std::stringstream ss;
    ss << val;
    return ss.str();
 }

-bool gguf_ex_write(const std::string & fname) {
+static bool gguf_ex_write(const std::string & fname) {
    struct gguf_context * ctx = gguf_init_empty();

    gguf_set_val_u8  (ctx, "some.parameter.uint8",    0x12);
@@ -85,7 +85,7 @@ bool gguf_ex_write(const std::string & fname) {
 }

 // just read tensor info
-bool gguf_ex_read_0(const std::string & fname) {
+static bool gguf_ex_read_0(const std::string & fname) {
    struct gguf_init_params params = {
        /*.no_alloc = */ false,
        /*.ctx      = */ NULL,
@@ -143,7 +143,7 @@ bool gguf_ex_read_0(const std::string & fname) {
 }

 // read and create ggml_context containing the tensors and their data
-bool gguf_ex_read_1(const std::string & fname) {
+static bool gguf_ex_read_1(const std::string & fname) {
    struct ggml_context * ctx_data = NULL;

    struct gguf_init_params params = {
--- a/examples/gptneox-wip/falcon-main.cpp
+++ b/examples/gptneox-wip/falcon-main.cpp
@@ -367,10 +367,10 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
        keyidx = gguf_find_key(ggufctx, "general.architecture");
        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { printf("%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
+        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
    }

--- a/examples/gptneox-wip/gptneox-main.cpp
+++ b/examples/gptneox-wip/gptneox-main.cpp
@@ -380,10 +380,10 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
        keyidx = gguf_find_key(ggufctx, "general.architecture");
        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { printf("%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
+        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
    }

--- a/examples/llama-bench/README.md
+++ b/examples/llama-bench/README.md
@@ -0,0 +1,271 @@
+# llama.cpp/example/llama-bench
+
+Performance testing tool for llama.cpp.
+
+## Table of contents
+
+1. [Syntax](#syntax)
+2. [Examples](#examples)
+    1. [Text generation with different models](#text-generation-with-different-models)
+    2. [Prompt processing with different batch sizes](#prompt-processing-with-different-batch-sizes)
+    3. [Different numbers of threads](#different-numbers-of-threads)
+    4. [Different numbers of layers offloaded to the GPU](#different-numbers-of-layers-offloaded-to-the-gpu)
+3. [Output formats](#output-formats)
+    1. [Markdown](#markdown)
+    2. [CSV](#csv)
+    3. [JSON](#json)
+    4. [SQL](#sql)
+
+## Syntax
+
+```
+usage: ./llama-bench [options]
+
+options:
+  -h, --help
+  -m, --model <filename>            (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                (default: 512)
+  -n, --n-gen <n>                   (default: 128)
+  -b, --batch-size <n>              (default: 512)
+  --memory-f32 <0|1>                (default: 0)
+  -t, --threads <n>                 (default: 16)
+  -ngl N, --n-gpu-layers <n>        (default: 99)
+  -mg i, --main-gpu <i>             (default: 0)
+  -mmq, --mul-mat-q <0|1>           (default: 1)
+  -ts, --tensor_split <ts0/ts1/..>
+  -r, --repetitions <n>             (default: 5)
+  -o, --output <csv|json|md|sql>    (default: md)
+  -v, --verbose                     (default: 0)
+
+Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
+```
+
+llama-bench can perform two types of tests:
+
+- Prompt processing (pp): processing a prompt in batches (`-p`)
+- Text generation (tg): generating a sequence of tokens (`-n`)
+
+With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).
+
+Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
+
+For a description of the other options, see the [main example](../main/README.md).
+
+## Examples
+
+### Text generation with different models
+
+```sh
+$ ./llama-bench -m models/7B/ggml-model-q4_0.gguf -m models/13B/ggml-model-q4_0.gguf -p 0 -n 128,256,512
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    132.19 ± 0.55 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 256     |    129.37 ± 0.54 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 512     |    123.83 ± 0.25 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 128     |     82.17 ± 0.31 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 256     |     80.74 ± 0.23 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 512     |     78.08 ± 0.07 |
+
+### Prompt processing with different batch sizes
+
+```sh
+$ ./llama-bench -n 0 -p 1024 -b 128,256,512,1024
+```
+
+| model                          |       size |     params | backend    | ngl |    n_batch | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        128 | pp 1024    |   1436.51 ± 3.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        256 | pp 1024    |  1932.43 ± 23.48 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        512 | pp 1024    |  2254.45 ± 15.59 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |       1024 | pp 1024    |  2498.61 ± 13.58 |
+
+### Different numbers of threads
+
+```sh
+$ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32
+```
+
+| model                          |       size |     params | backend    |    threads | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | pp 64      |      6.17 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | tg 16      |      4.05 ± 0.02 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | pp 64      |     12.31 ± 0.13 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | tg 16      |      7.80 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | pp 64      |     23.18 ± 0.06 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | tg 16      |     12.22 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | pp 64      |     32.29 ± 1.21 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | tg 16      |     16.71 ± 0.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | pp 64      |     33.52 ± 0.03 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | tg 16      |     15.32 ± 0.05 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | pp 64      |     59.00 ± 1.11 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 ||
+
+### Different numbers of layers offloaded to the GPU
+
+```sh
+$ ./llama-bench -ngl 10,20,30,31,32,33,34,35
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | pp 512     |    373.36 ± 2.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | tg 128     |     13.45 ± 0.93 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | pp 512     |    472.65 ± 1.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | tg 128     |     21.36 ± 1.94 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | pp 512     |   631.87 ± 11.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | tg 128     |     40.04 ± 1.82 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | pp 512     |    657.89 ± 5.08 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | tg 128     |     48.19 ± 0.81 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | pp 512     |    688.26 ± 3.29 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | tg 128     |     54.78 ± 0.65 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | pp 512     |    704.27 ± 2.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | tg 128     |     60.62 ± 1.76 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | pp 512     |    881.34 ± 5.40 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | tg 128     |     71.76 ± 0.23 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
+
+## Output formats
+
+By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
+
+### Markdown
+
+```sh
+$ ./llama-bench -o md
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | pp 512     |  2368.80 ± 93.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    131.42 ± 0.59 |
+
+### CSV
+
+```sh
+$ ./llama-bench -o csv
+```
+
+```csv
+build_commit,build_number,cuda,opencl,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
+```
+
+### JSON
+
+```sh
+$ ./llama-bench -o json
+```
+
+```json
+[
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 512,
+    "n_gen": 0,
+    "test_time": "2023-09-23T12:09:57Z",
+    "avg_ns": 212365953,
+    "stddev_ns": 985423,
+    "avg_ts": 2410.974041,
+    "stddev_ts": 11.163766,
+    "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ],
+    "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ]
+  },
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 0,
+    "n_gen": 128,
+    "test_time": "2023-09-23T12:09:59Z",
+    "avg_ns": 977425219,
+    "stddev_ns": 9268593,
+    "avg_ts": 130.965708,
+    "stddev_ts": 1.238924,
+    "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ],
+    "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ]
+  }
+]
+```
+
+### SQL
+
+SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.
+
+```sh
+$ ./llama-bench -o sql
+```
+
+```sql
+CREATE TABLE IF NOT EXISTS test (
+  build_commit TEXT,
+  build_number INTEGER,
+  cuda INTEGER,
+  opencl INTEGER,
+  metal INTEGER,
+  gpu_blas INTEGER,
+  blas INTEGER,
+  cpu_info TEXT,
+  gpu_info TEXT,
+  model_filename TEXT,
+  model_type TEXT,
+  model_size INTEGER,
+  model_n_params INTEGER,
+  n_batch INTEGER,
+  n_threads INTEGER,
+  f16_kv INTEGER,
+  n_gpu_layers INTEGER,
+  main_gpu INTEGER,
+  mul_mat_q INTEGER,
+  tensor_split TEXT,
+  n_prompt INTEGER,
+  n_gen INTEGER,
+  test_time TEXT,
+  avg_ns INTEGER,
+  stddev_ns INTEGER,
+  avg_ts REAL,
+  stddev_ts REAL
+);
+
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
+```
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -74,14 +74,6 @@ static T stdev(const std::vector<T> & v) {
    return stdev;
 }

-static bool ggml_cpu_has_metal() {
-#if defined(GGML_USE_METAL)
-    return true;
-#else
-    return false;
-#endif
-}
-
 static std::string get_cpu_info() {
    std::string id;
 #ifdef __linux__
@@ -899,7 +891,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
    int n_processed = 0;
    while (n_processed < n_prompt) {
        int n_tokens = std::min(n_prompt - n_processed, n_batch);
-        llama_eval(ctx, tokens.data(), n_tokens, n_past + n_processed, n_threads);
+        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0), n_threads);
        n_processed += n_tokens;
    }
 }
@@ -907,11 +899,11 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
    llama_token token = llama_token_bos(ctx);
    for (int i = 0; i < n_gen; i++) {
-        llama_eval(ctx, &token, 1, n_past + i, n_threads);
+        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0), n_threads);
    }
 }

-static void llama_null_log_callback(enum llama_log_level level, const char * text, void * user_data) {
+static void llama_null_log_callback(enum ggml_log_level level, const char * text, void * user_data) {
    (void) level;
    (void) text;
    (void) user_data;
@@ -985,6 +977,8 @@ int main(int argc, char ** argv) {

        test t(inst, lmodel, ctx);

+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
        // warmup run
        if (t.n_prompt > 0) {
            test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
@@ -994,6 +988,8 @@ int main(int argc, char ** argv) {
        }

        for (int i = 0; i < params.reps; i++) {
+            llama_kv_cache_tokens_rm(ctx, -1, -1);
+
            uint64_t t_start = get_time_ns();
            if (t.n_prompt > 0) {
                test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
--- a/examples/main-cmake-pkg/.gitignore
+++ b/examples/main-cmake-pkg/.gitignore
@@ -0,0 +1,51 @@
+# Prerequisites
+*.d
+
+# Compiled Object files
+*.slo
+*.lo
+*.o
+*.obj
+
+# Precompiled Headers
+*.gch
+*.pch
+
+# Compiled Dynamic libraries
+*.so
+*.dylib
+*.dll
+
+# Fortran module files
+*.mod
+*.smod
+
+# Compiled Static libraries
+*.lai
+*.la
+*.a
+*.lib
+
+# Executables
+*.exe
+*.out
+*.app
+
+*.gguf
+
+*.log
+.DS_Store
+.build/
+.cache/
+.direnv/
+.envrc
+.swiftpm
+.venv
+.clang-tidy
+.vs/
+.vscode/
+
+build*/
+out/
+tmp/
+
--- a/examples/main-cmake-pkg/CMakeLists.txt
+++ b/examples/main-cmake-pkg/CMakeLists.txt
@@ -0,0 +1,36 @@
+cmake_minimum_required(VERSION 3.12)
+project("main-cmake-pkg" C CXX)
+set(TARGET main-cmake-pkg)
+
+find_package(Llama 0.0.1 REQUIRED)
+
+# Bake common functionality in with target. Because applications
+# using the relocatable Llama package should be outside of the
+# source tree, main-cmake-pkg pretends the dependencies are built-in.
+
+set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common")
+add_library(common OBJECT
+    ${_common_path}/common.h
+    ${_common_path}/common.cpp
+    ${_common_path}/console.h
+    ${_common_path}/console.cpp
+    ${_common_path}/grammar-parser.h
+    ${_common_path}/grammar-parser.cpp
+    )
+
+# WARNING: because build-info.h is auto-generated, it will only
+# be available after the user has built the llama.cpp sources.
+#
+configure_file(${_common_path}/../build-info.h
+    ${CMAKE_CURRENT_BINARY_DIR}/build-info.h
+    COPYONLY)
+
+target_include_directories(common PUBLIC ${LLAMA_INCLUDE_DIR}
+    ${CMAKE_CURRENT_BINARY_DIR})
+
+add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp)
+target_include_directories(${TARGET} PRIVATE ${_common_path})
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
--- a/examples/main-cmake-pkg/README.md
+++ b/examples/main-cmake-pkg/README.md
@@ -0,0 +1,37 @@
+# llama.cpp/example/main-cmake-pkg
+
+This program builds the [main](../main) application using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
+
+## Building
+
+Because this example is "outside of the source tree", it is important to first build/install llama.cpp using CMake. An example is provided here, but please see the [llama.cpp build instructions](../..) for more detailed build instructions.
+
+### Considerations
+
+When hardware acceleration libraries are used (e.g. CUBlas, Metal, CLBlast, etc.), CMake must be able to locate the associated CMake package. In the example below, when building _main-cmake-pkg_ notice the `CMAKE_PREFIX_PATH` includes the Llama CMake package location _in addition to_ the CLBlast package—which was used when compiling _llama.cpp_.
+
+### Build llama.cpp and install to C:\LlamaCPP directory
+
+In this case, CLBlast was already installed so the CMake package is referenced in `CMAKE_PREFIX_PATH`.
+
+```cmd
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+mkdir build
+cd build
+cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=C:/CLBlast/lib/cmake/CLBlast -G "Visual Studio 17 2022" -A x64
+cmake --build . --config Release
+cmake --install . --prefix C:/LlamaCPP
+```
+
+### Build main-cmake-pkg
+
+
+```cmd
+cd ..\examples\main-cmake-pkg
+mkdir build
+cd build
+cmake .. -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/CLBlast/lib/cmake/CLBlast;C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64
+cmake --build . --config Release
+cmake --install . --prefix C:/MyLlamaApp
+```
--- a/examples/main/README.md
+++ b/examples/main/README.md
@@ -144,7 +144,7 @@ The `--ctx-size` option allows you to set the size of the prompt context used by

 Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.

- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+-   `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.

 ### Keep Prompt

@@ -274,7 +274,7 @@ These options help improve the performance and memory usage of the LLaMA models.

 ### NUMA support

-   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop\_caches' as root.
+-   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.

 ### Memory Float 32

@@ -302,7 +302,6 @@ These options provide extra functionality and customization when running the LLa

 -   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
 -   `--verbose-prompt`: Print the prompt before generating text.
-   `--mtest`: Test the model's functionality by running a series of tests to ensure it's working properly.
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -41,7 +41,8 @@ static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting = false;

-void write_logfile(
+
+static void write_logfile(
    const llama_context * ctx, const gpt_params & params, const llama_model * model,
    const std::vector<llama_token> & input_tokens, const std::string & output,
    const std::vector<llama_token> & output_tokens
@@ -86,7 +87,7 @@ void write_logfile(
 }

 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-void sigint_handler(int signo) {
+static void sigint_handler(int signo) {
    if (signo == SIGINT) {
        if (!is_interacting) {
            is_interacting = true;
@@ -123,7 +124,7 @@ int main(int argc, char ** argv) {
    console::init(params.simple_io, params.use_color);
    atexit([]() { console::cleanup(); });

-    if (params.perplexity) {
+    if (params.logits_all) {
        printf("\n************\n");
        printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
        printf("************\n\n");
@@ -148,6 +149,7 @@ int main(int argc, char ** argv) {
    }

    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);

    if (params.seed == LLAMA_DEFAULT_SEED) {
        params.seed = time(NULL);
@@ -198,32 +200,6 @@ int main(int argc, char ** argv) {
                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
    }

-    // determine the maximum memory usage needed to do inference for the given n_batch and n_ctx parameters
-    // uncomment the "used_mem" line in llama.cpp to see the results
-    if (params.mem_test) {
-        {
-            LOG_TEE("%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx);
-
-            const std::vector<llama_token> tmp(params.n_batch, llama_token_bos(ctx));
-            llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads);
-        }
-
-        llama_print_timings(ctx);
-        llama_free(ctx);
-        llama_free_model(model);
-
-        return 0;
-    }
-
-    // export the cgraph and exit
-    if (params.export_cgraph) {
-        llama_eval_export(ctx, "llama.ggml");
-        llama_free(ctx);
-        llama_free_model(model);
-
-        return 0;
-    }
-
    std::string path_session = params.path_prompt_cache;
    std::vector<llama_token> session_tokens;

@@ -523,18 +499,23 @@ int main(int argc, char ** argv) {
                    break;
                }

-                const int n_left = n_past - params.n_keep;
-                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d\n", n_past, n_left, n_ctx, params.n_keep);
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;

-                // always keep the first token - BOS
-                n_past          = std::max(1, params.n_keep);
-                n_past_guidance = std::max(1, params.n_keep + guidance_offset);
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
+
+                if (ctx_guidance) {
+                    n_past_guidance -= n_discard;
+                }

                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);

-                // insert n_left/2 tokens at the start of embd from last_tokens
-                embd.insert(embd.begin(), last_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_tokens.end() - embd.size());
-
                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));

                LOG("clear session path\n");
@@ -595,7 +576,7 @@ int main(int argc, char ** argv) {

                for (int i = 0; i < input_size; i += params.n_batch) {
                    int n_eval = std::min(input_size - i, params.n_batch);
-                    if (llama_eval(ctx_guidance, input_buf + i, n_eval, n_past_guidance, params.n_threads)) {
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0), params.n_threads)) {
                        LOG_TEE("%s : failed to eval\n", __func__);
                        return 1;
                    }
@@ -612,7 +593,7 @@ int main(int argc, char ** argv) {

                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));

-                if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0), params.n_threads)) {
                    LOG_TEE("%s : failed to eval\n", __func__);
                    return 1;
                }
--- a/examples/make-ggml.py
+++ b/examples/make-ggml.py
@@ -1,22 +1,25 @@
 #!/usr/bin/env python3
 """
-This script converts Hugging Face llama models to GGML and quantizes them.
+This script converts Hugging Face Llama, StarCoder, Falcon, Baichuan, and GPT-NeoX models to GGUF and quantizes them.

 Usage:
-python make-ggml.py --model {model_dir_or_hf_repo_name} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
+python make-ggml.py {model_dir_or_hf_repo_name} --model_type {model_type} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]

 Arguments:
- --model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- --model_type: (Required) The type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.
 - --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used.
 - --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used.
 - --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'.
 - --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created.

-Quant types:
+Old quant types (some base model types require these):
 - Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M
 - Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L
 - Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M
 - Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M
+
+New quant types (recommended):
 - Q2_K: smallest, extreme quality loss - not recommended
 - Q3_K: alias for Q3_K_M
 - Q3_K_S: very small, very high quality loss
@@ -40,9 +43,7 @@ import argparse
 import os
 from huggingface_hub import snapshot_download

-def main(model, outname, outdir, quants, keep_fp16):
-    ggml_version = "v3"
-
+def main(model, model_type, outname, outdir, quants, keep_fp16):
    if not os.path.isdir(model):
        print(f"Model not found at {model}. Downloading...")
        try:
@@ -63,17 +64,20 @@ def main(model, outname, outdir, quants, keep_fp16):
    print("Building llama.cpp")
    subprocess.run(f"cd .. && make quantize", shell=True, check=True)

-    fp16 = f"{outdir}/{outname}.ggml{ggml_version}.fp16.bin"
+    fp16 = f"{outdir}/{outname}.gguf.fp16.bin"

-    print(f"Making unquantised GGML at {fp16}")
+    print(f"Making unquantised GGUF at {fp16}")
    if not os.path.isfile(fp16):
-        subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
+        if model_type != "llama":
+            subprocess.run(f"python3 ../convert-{model_type}-hf-to-gguf.py {model} 1 --outfile {fp16}", shell=True, check=True)
+        else:
+            subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
    else:
        print(f"Unquantised GGML already exists at: {fp16}")

    print("Making quants")
    for type in quants:
-        outfile = f"{outdir}/{outname}.ggml{ggml_version}.{type}.bin"
+        outfile = f"{outdir}/{outname}.gguf.{type}.bin"
        print(f"Making {type} : {outfile}")
        subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True)

@@ -81,8 +85,9 @@ def main(model, outname, outdir, quants, keep_fp16):
        os.remove(fp16)

 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description='Convert/Quantize HF to GGML. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
-    parser.add_argument('--model', required=True, help='Downloaded model dir or Hugging Face model repo name')
+    parser = argparse.ArgumentParser(description='Convert/Quantize HF models to GGUF. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
+    parser.add_argument('model', help='Downloaded model dir or Hugging Face model repo name')
+    parser.add_argument('--model_type', required=True, choices=['llama', 'starcoder', 'falcon', 'baichuan', 'gptneox'], help='Type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.')
    parser.add_argument('--outname', default=None, help='Output model(s) name')
    parser.add_argument('--outdir', default=None, help='Output directory')
    parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types')
@@ -90,4 +95,4 @@ if __name__ == "__main__":

    args = parser.parse_args()

-    main(args.model, args.outname, args.outdir, args.quants, args.keep_fp16)
+    main(args.model, args.model_type, args.outname, args.outdir, args.quants, args.keep_fp16)
--- a/examples/parallel/CMakeLists.txt
+++ b/examples/parallel/CMakeLists.txt
@@ -0,0 +1,8 @@
+set(TARGET parallel)
+add_executable(${TARGET} parallel.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()
--- a/examples/parallel/README.md
+++ b/examples/parallel/README.md
@@ -0,0 +1,3 @@
+# llama.cpp/example/parallel
+
+Simplified simluation for serving incoming requests in parallel
--- a/examples/parallel/parallel.cpp
+++ b/examples/parallel/parallel.cpp
@@ -0,0 +1,380 @@
+// A basic application simulating a server with multiple clients.
+// The clients submite requests to the server and they are processed in parallel.
+
+#include "build-info.h"
+
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+// trim whitespace from the beginning and end of a string
+static std::string trim(const std::string & str) {
+    size_t start = 0;
+    size_t end = str.size();
+
+    while (start < end && isspace(str[start])) {
+        start += 1;
+    }
+
+    while (end > start && isspace(str[end - 1])) {
+        end -= 1;
+    }
+
+    return str.substr(start, end - start);
+}
+
+static std::string k_system =
+R"(Transcript of a never ending dialog, where the User interacts with an Assistant.
+The Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.
+
+User: Recommend a nice restaurant in the area.
+Assistant: I recommend the restaurant "The Golden Duck". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.
+User: Who is Richard Feynman?
+Assistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including "Surely You're Joking, Mr. Feynman!" and "What Do You Care What Other People Think?".
+User:)";
+
+static std::vector<std::string> k_prompts = {
+    "What is the meaning of life?",
+    "Tell me an interesting fact about llamas.",
+    "What is the best way to cook a steak?",
+    "Are you familiar with the Special Theory of Relativity and can you explain it to me?",
+    "Recommend some interesting books to read.",
+    "What is the best way to learn a new language?",
+    "How to get a job at Google?",
+    "If you could have any superpower, what would it be?",
+    "I want to learn how to play the piano.",
+};
+
+struct client {
+    int32_t id = 0;
+
+    llama_seq_id seq_id = -1;
+
+    llama_token sampled;
+
+    int64_t t_start_prompt;
+    int64_t t_start_gen;
+
+    int32_t n_prompt  = 0;
+    int32_t n_decoded = 0;
+    int32_t i_batch   = -1;
+
+    std::string input;
+    std::string prompt;
+    std::string response;
+
+    std::vector<llama_token> tokens_prev;
+};
+
+int main(int argc, char ** argv) {
+    srand(1234);
+
+    gpt_params params;
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    // number of simultaneous "clients" to simulate
+    const int32_t n_clients = params.n_parallel;
+
+    // requests to simulate
+    const int32_t n_seq = params.n_sequences;
+
+    // insert new requests as soon as the previous one is done
+    const bool cont_batching = params.cont_batching;
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("parallel", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the target model
+    params.logits_all = true;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+
+    fprintf(stderr, "\n\n");
+    fflush(stderr);
+
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(ctx);
+
+    std::vector<client> clients(n_clients);
+    for (size_t i = 0; i < clients.size(); ++i) {
+        auto & client = clients[i];
+        client.id = i;
+        client.tokens_prev.resize(std::max(256, params.n_predict));
+        std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
+    }
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    std::vector<llama_token> tokens_system;
+    tokens_system = ::llama_tokenize(ctx, k_system, true);
+    const int32_t n_tokens_system = tokens_system.size();
+
+    llama_seq_id g_seq_id = 0;
+
+    // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
+    // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
+    llama_batch batch = llama_batch_init(params.n_ctx, 0);
+
+    int32_t n_total_prompt = 0;
+    int32_t n_total_gen    = 0;
+    int32_t n_cache_miss   = 0;
+
+    const auto t_main_start = ggml_time_us();
+
+    LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
+    LOG_TEE("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
+    LOG_TEE("\n");
+
+    {
+        LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
+
+        batch.n_tokens = n_tokens_system;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            batch.token[i]  = tokens_system[i];
+            batch.pos[i]    = i;
+            batch.seq_id[i] = 0;
+            batch.logits[i] = false;
+        }
+
+        if (llama_decode(ctx, batch, params.n_threads) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        // assign the system KV cache to all parallel sequences
+        for (int32_t i = 1; i < n_clients; ++i) {
+            llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system);
+        }
+
+        LOG_TEE("\n");
+    }
+
+    LOG_TEE("Processing requests ...\n\n");
+
+    while (true) {
+        batch.n_tokens = 0;
+
+        // decode any currently ongoing sequences
+        for (auto & client : clients) {
+            if (client.seq_id == -1) {
+                continue;
+            }
+
+            batch.token [batch.n_tokens] = client.sampled;
+            batch.pos   [batch.n_tokens] = n_tokens_system + client.n_prompt + client.n_decoded;
+            batch.seq_id[batch.n_tokens] = client.id;
+            batch.logits[batch.n_tokens] = true;
+
+            client.n_decoded += 1;
+            client.i_batch = batch.n_tokens;
+
+            batch.n_tokens += 1;
+        }
+
+        if (batch.n_tokens == 0) {
+            // all sequences have ended - clear the entire KV cache
+            for (int i = 0; i < n_clients; ++i) {
+                llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
+            }
+
+            LOG_TEE("%s: clearing the KV cache\n", __func__);
+        }
+
+        // insert new sequences for decoding
+        if (cont_batching || batch.n_tokens == 0) {
+            for (auto & client : clients) {
+                if (client.seq_id == -1 && g_seq_id < n_seq) {
+                    client.seq_id = g_seq_id;
+
+                    client.t_start_prompt = ggml_time_us();
+                    client.t_start_gen    = 0;
+
+                    client.input    = k_prompts[rand() % k_prompts.size()];
+                    client.prompt   = client.input + "\nAssistant:";
+                    client.response = "";
+
+                    std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
+
+                    // do not prepend BOS because we have a system prompt!
+                    std::vector<llama_token> tokens_prompt;
+                    tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
+
+                    for (size_t i = 0; i < tokens_prompt.size(); ++i) {
+                        batch.token [batch.n_tokens] = tokens_prompt[i];
+                        batch.pos   [batch.n_tokens] = i + n_tokens_system;
+                        batch.seq_id[batch.n_tokens] = client.id;
+                        batch.logits[batch.n_tokens] = false;
+                        batch.n_tokens += 1;
+                    }
+
+                    // extract the logits only for the last token
+                    if (batch.n_tokens > 0) {
+                        batch.logits[batch.n_tokens - 1] = true;
+                    }
+
+                    client.n_prompt  = tokens_prompt.size();
+                    client.n_decoded = 0;
+                    client.i_batch   = batch.n_tokens - 1;
+
+                    LOG_TEE("\033[1mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
+
+                    g_seq_id += 1;
+
+                    // insert new requests one-by-one
+                    //if (cont_batching) {
+                    //    break;
+                    //}
+                }
+            }
+        }
+
+        if (batch.n_tokens == 0) {
+            break;
+        }
+
+        // process in chunks of params.n_batch
+        int32_t n_batch = params.n_batch;
+
+        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
+            // experiment: process in powers of 2
+            //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
+            //    n_batch /= 2;
+            //    i -= n_batch;
+            //    continue;
+            //}
+
+            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+
+            llama_batch batch_view = {
+                n_tokens,
+                batch.token  + i,
+                nullptr,
+                batch.pos    + i,
+                batch.seq_id + i,
+                batch.logits + i,
+                0, 0, 0, // unused
+            };
+
+            const int ret = llama_decode(ctx, batch_view, params.n_threads);
+            if (ret != 0) {
+                if (n_batch == 1 || ret < 0) {
+                    // if you get here, it means the KV cache is full - try increasing it via the context size
+                    LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
+                    return 1;
+                }
+
+                LOG("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
+
+                n_cache_miss += 1;
+
+                // retry with half the batch size to try to find a free slot in the KV cache
+                n_batch /= 2;
+                i -= n_batch;
+
+                continue;
+            }
+
+            LOG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
+
+            for (auto & client : clients) {
+                if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
+                    continue;
+                }
+
+                //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
+                //        client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
+
+                const llama_token id = llama_sample_token(ctx, NULL, NULL, params, client.tokens_prev, candidates, client.i_batch - i);
+
+                if (client.n_decoded == 1) {
+                    // start measuring generation time after the first token to make sure all concurrent clients
+                    // have their prompt already processed
+                    client.t_start_gen = ggml_time_us();
+                }
+
+                // remember which tokens were sampled - used for repetition penalties during sampling
+                client.tokens_prev.erase(client.tokens_prev.begin());
+                client.tokens_prev.push_back(id);
+
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                client.response += token_str;
+                client.sampled = id;
+
+                //printf("client %d, seq %d, token %d, pos %d, batch %d: %s\n",
+                //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
+
+                if (client.n_decoded > 2 &&
+                        (id == llama_token_eos(ctx) ||
+                         (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
+                         client.response.find("User:") != std::string::npos ||
+                         client.response.find('\n') != std::string::npos)) {
+                    // basic reverse prompt
+                    const size_t pos = client.response.find("User:");
+                    if (pos != std::string::npos) {
+                        client.response = client.response.substr(0, pos);
+                    }
+
+                    // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
+                    llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, n_ctx);
+
+                    const auto t_main_end = ggml_time_us();
+
+                    LOG_TEE("\033[1mClient %3d, seq %4d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \n\nInput:    %s\nResponse: %s\n\n",
+                            client.id, client.seq_id, client.n_prompt, client.n_decoded,
+                            (t_main_end - client.t_start_prompt) / 1e6,
+                            (double) (client.n_prompt + client.n_decoded) / (t_main_end - client.t_start_prompt) * 1e6,
+                            n_cache_miss,
+                            ::trim(client.input).c_str(),
+                            ::trim(client.response).c_str());
+
+                    n_total_prompt += client.n_prompt;
+                    n_total_gen    += client.n_decoded;
+
+                    client.seq_id = -1;
+                }
+
+                client.i_batch = -1;
+            }
+        }
+    }
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("\n\n");
+    LOG_TEE("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt              ) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Total gen tokens:    %6d, speed: %5.2f t/s\n", n_total_gen,    (double) (n_total_gen                 ) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Total speed (AVG):   %6s  speed: %5.2f t/s\n", "",             (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Cache misses:        %6d\n", n_cache_miss);
+
+    LOG_TEE("\n\n");
+
+    llama_print_timings(ctx);
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}
--- a/examples/perplexity/README.md
+++ b/examples/perplexity/README.md
@@ -1,3 +1,21 @@
 # perplexity

 TODO
+
+## Llama 2 70B Scorechart
+Quantization | Model size (GiB) | Perplexity | Delta to fp16
+-- | -- | -- | --
+Q4_0 | 36.20 | 3.5550 | 3.61%
+Q4_1 | 40.20 | 3.5125 | 2.37%
+Q5_0 | 44.20 | 3.4744 | 1.26%
+Q2_K | 27.27 | 3.7339 | 8.82%
+Q3_K_S | 27.86 | 3.7019 | 7.89%
+Q3_K_M | 30.83 | 3.5932 | 4.72%
+Q3_K_L | 33.67 | 3.5617 | 3.80%
+Q4_K_S | 36.39 | 3.4852 | 1.57%
+Q4_K_M | 38.54 | 3.4725 | 1.20%
+Q5_K_S | 44.20 | 3.4483 | 0.50%
+Q5_K_M | 45.41 | 3.4451 | 0.40%
+Q6_K | 52.70 | 3.4367 | 0.16%
+fp16 | 128.5 | 3.4313 | -
+
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -1,6 +1,6 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"

 #include <cmath>
 #include <cstdio>
@@ -28,9 +28,10 @@ struct results_log_softmax {
    float  prob;
 };

-void write_logfile(const llama_context * ctx, const gpt_params & params,
-                   const llama_model * model, const struct results_perplexity & results) {
-
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const struct results_perplexity & results
+) {
    if (params.logdir.empty()) {
        return;
    }
@@ -76,10 +77,12 @@ void write_logfile(const llama_context * ctx, const gpt_params & params,
    fclose(logfile);
 }

-std::vector<float> softmax(const std::vector<float>& logits) {
+static std::vector<float> softmax(const std::vector<float>& logits) {
    std::vector<float> probs(logits.size());
    float max_logit = logits[0];
-    for (float v : logits) max_logit = std::max(max_logit, v);
+    for (float v : logits) {
+        max_logit = std::max(max_logit, v);
+    }
    double sum_exp = 0.0;
    for (size_t i = 0; i < logits.size(); i++) {
        // Subtract the maximum logit value from the current logit value for numerical stability
@@ -88,25 +91,33 @@ std::vector<float> softmax(const std::vector<float>& logits) {
        sum_exp += exp_logit;
        probs[i] = exp_logit;
    }
-    for (size_t i = 0; i < probs.size(); i++) probs[i] /= sum_exp;
+    for (size_t i = 0; i < probs.size(); i++) {
+        probs[i] /= sum_exp;
+    }
    return probs;
 }

-results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
+static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
    float max_logit = logits[0];
-    for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]);
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+    }
    double sum_exp = 0.0;
-    for (int i = 0; i < n_vocab; ++i) sum_exp += expf(logits[i] - max_logit);
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
    return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
 }

-void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
-        double & nll, double & nll2, float * logit_history, float * prob_history) {
-
+static void process_logits(
+    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
+    double & nll, double & nll2, float * logit_history, float * prob_history
+) {
    std::mutex mutex;
    int counter = 0;
    auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
-        double local_nll = 0, local_nll2 = 0;
+        double local_nll  = 0;
+        double local_nll2 = 0;
        while (true) {
            std::unique_lock<std::mutex> lock(mutex);
            int i = counter++;
@@ -124,13 +135,16 @@ void process_logits(int n_vocab, const float * logits, const int * tokens, int n
            prob_history[i]  = results.prob;
        }
    };
-    for (auto & w : workers) w = std::thread(compute);
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
    compute();
-    for (auto & w : workers) w.join();
-
+    for (auto & w : workers) {
+        w.join();
+    }
 }

-results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
+static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
    // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
    // Output: `perplexity: 13.5106 [114/114]`
@@ -150,8 +164,8 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
        return {std::move(tokens), 0., {}, {}};
    }

-    std::vector<float>       logit_history;
-    std::vector<float>       prob_history;
+    std::vector<float> logit_history;
+    std::vector<float> prob_history;

    logit_history.resize(tokens.size());
    prob_history.resize(tokens.size());
@@ -193,12 +207,15 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)

        const auto t_start = std::chrono::high_resolution_clock::now();

+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
        for (int j = 0; j < num_batches; ++j) {
            const int batch_start = start + j * n_batch;
            const int batch_size  = std::min(end - batch_start, n_batch);

            //fprintf(stderr, "    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0), params.n_threads)) {
                //fprintf(stderr, "%s : failed to eval\n", __func__);
                return {tokens, -1, logit_history, prob_history};
            }
@@ -260,8 +277,7 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
    return {tokens, std::exp(nll / count), logit_history, prob_history};
 }

-results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
-
+static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
    if (params.ppl_stride > 0) {
        return perplexity_v2(ctx, params);
    }
@@ -319,6 +335,9 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {

        const auto t_start = std::chrono::high_resolution_clock::now();

+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
        for (int j = 0; j < num_batches; ++j) {
            const int batch_start = start + j * n_batch;
            const int batch_size  = std::min(end - batch_start, n_batch);
@@ -331,7 +350,7 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
                tokens[batch_start] = llama_token_bos(ctx);
            }

-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0), params.n_threads)) {
                fprintf(stderr, "%s : failed to eval\n", __func__);
                return {tokens, -1, logit_history, prob_history};
            }
@@ -400,15 +419,16 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
    return {tokens, ppl, logit_history, prob_history};
 }

-std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch,
-        int n_vocab, int n_thread) {
+static std::vector<float> hellaswag_evaluate_tokens(
+    llama_context * ctx, std::vector<int> & tokens, int n_past, int n_batch, int n_vocab, int n_thread
+) {
    std::vector<float> result;
    result.reserve(tokens.size() * n_vocab);
    size_t n_chunk = (tokens.size() + n_batch - 1)/n_batch;
    for (size_t i_chunk = 0; i_chunk < n_chunk; ++i_chunk) {
        size_t n_tokens = tokens.size() - i_chunk * n_batch;
        n_tokens = std::min(n_tokens, size_t(n_batch));
-        if (llama_eval(ctx, tokens.data() + i_chunk * n_batch, n_tokens, n_past, n_thread)) {
+        if (llama_decode(ctx, llama_batch_get_one(tokens.data() + i_chunk * n_batch, n_tokens, n_past, 0), n_thread)) {
            fprintf(stderr, "%s : failed to eval\n", __func__);
            return {};
        }
@@ -421,7 +441,7 @@ std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vec
    return result;
 }

-void hellaswag_score(llama_context * ctx, const gpt_params & params) {
+static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
    // Calculates hellaswag score (acc_norm) from prompt
    //
    // Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
@@ -548,6 +568,9 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
            query_embd.resize(32);
        }

+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
        auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab, params.n_threads);
        if (logits.empty()) {
            fprintf(stderr, "%s : failed to eval\n", __func__);
@@ -659,7 +682,7 @@ int main(int argc, char ** argv) {
        return 1;
    }

-    params.perplexity = true;
+    params.logits_all = true;
    params.n_batch = std::min(params.n_batch, params.n_ctx);

    if (params.ppl_stride > 0) {
@@ -668,7 +691,7 @@ int main(int argc, char ** argv) {
        params.n_ctx += params.ppl_stride/2;
    }

-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();

    if (params.seed == LLAMA_DEFAULT_SEED) {
        params.seed = time(NULL);
--- a/examples/quantize-stats/CMakeLists.txt
+++ b/examples/quantize-stats/CMakeLists.txt
@@ -2,4 +2,5 @@ set(TARGET quantize-stats)
 add_executable(${TARGET} quantize-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
--- a/examples/quantize-stats/quantize-stats.cpp
+++ b/examples/quantize-stats/quantize-stats.cpp
@@ -1,7 +1,7 @@
-#include "ggml.h"
-#include "build-info.h"
-
 #define LLAMA_API_INTERNAL
+#include "build-info.h"
+#include "common.h"
+#include "ggml.h"
 #include "llama.h"

 #include <algorithm>
@@ -34,8 +34,8 @@ struct quantize_stats_params {
    std::vector<enum ggml_type> include_types;
 };

-const size_t HISTOGRAM_BUCKETS = 150;
-const double HISTOGRAM_RANGE = 0.03;
+constexpr size_t HISTOGRAM_BUCKETS = 150;
+constexpr double HISTOGRAM_RANGE = 0.03;

 struct error_stats {
    size_t num_samples;
@@ -44,8 +44,7 @@ struct error_stats {
    uint64_t error_histogram[HISTOGRAM_BUCKETS];
 };

-
-void quantize_stats_print_usage(int /*argc*/, char ** argv) {
+static void quantize_stats_print_usage(int /*argc*/, char ** argv) {
    quantize_stats_params params;
    fprintf(stderr, "usage: %s [options]\n", argv[0]);
    fprintf(stderr, "\n");
@@ -71,7 +70,7 @@ void quantize_stats_print_usage(int /*argc*/, char ** argv) {
 }

 // Check if a layer is included/excluded by command line
-bool layer_included(const quantize_stats_params & params, const std::string & layer) {
+static bool layer_included(const quantize_stats_params & params, const std::string & layer) {
    for (const auto& excluded : params.exclude_layers) {
        if (std::regex_search(layer, std::regex(excluded))) {
            return false;
@@ -86,7 +85,7 @@ bool layer_included(const quantize_stats_params & params, const std::string & la
 }

 // Update error statistics given vectors with the before/after result of quantization
-void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
+static void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
    for (int64_t i = 0; i < nelements; i++) {
        double diff = input[i] - output[i];
        stats.total_error += diff * diff;
@@ -96,14 +95,14 @@ void update_error_stats(int64_t nelements, const float * input, const float * ou
    stats.num_samples += nelements;
 }

-void combine_error_stats(error_stats & into, const error_stats & from) {
+static void combine_error_stats(error_stats & into, const error_stats & from) {
    into.num_samples += from.num_samples;
    into.total_error += from.total_error;
    if (from.max_error > into.max_error) into.max_error = from.max_error;
    for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];
 }

-double find_quantile(const error_stats & stats, double quantile) {
+static double find_quantile(const error_stats & stats, double quantile) {
    double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);

    double accum = 0;
@@ -116,7 +115,7 @@ double find_quantile(const error_stats & stats, double quantile) {
    return INFINITY;
 }

-void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
+static void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
    double rmse = sqrt(stats.total_error / (double) stats.num_samples);
    double median = find_quantile(stats, .5);
    double pct95 = find_quantile(stats, .95);
@@ -143,17 +142,10 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }

-void test_roundtrip_on_chunk(
-        const ggml_tensor * layer,
-        int64_t offset,
-        int64_t chunk_size,
-        const ggml_type_traits_t & qfns,
-        bool use_reference,
-        float * input_scratch,
-        char * quantized_scratch,
-        float * output_scratch,
-        error_stats & stats) {
-
+static void test_roundtrip_on_chunk(
+    const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits_t & qfns, bool use_reference,
+    float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
+) {
    if (layer->type == GGML_TYPE_F16) {
        for (int i = 0; i < chunk_size; i++) {
            input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
@@ -174,18 +166,11 @@ void test_roundtrip_on_chunk(


 // Run quantization function for a single layer and update error stats
-void test_roundtrip_on_layer(
-        std::string & name,
-        bool print_layer_stats,
-        const ggml_type_traits_t & qfns,
-        bool use_reference,
-        const ggml_tensor * layer,
-        std::vector<float> & input_scratch,
-        std::vector<char> & quantized_scratch,
-        std::vector<float> & output_scratch,
-        error_stats & total_error,
-        int max_thread = 0) {
-
+static void test_roundtrip_on_layer(
+    std::string & name, bool print_layer_stats, const ggml_type_traits_t & qfns, bool use_reference,
+    const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
+    std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
+) {
    assert(tensor_is_contiguous(layer));
    error_stats layer_error {};
    uint64_t nelements = ggml_nelements(layer);
@@ -314,7 +299,7 @@ int main(int argc, char ** argv) {
        return 1;
    }

-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();

    // load the model
    fprintf(stderr, "Loading model\n");
--- a/examples/quantize/CMakeLists.txt
+++ b/examples/quantize/CMakeLists.txt
@@ -2,6 +2,7 @@ set(TARGET quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
  add_dependencies(${TARGET} BUILD_INFO)
--- a/examples/quantize/README.md
+++ b/examples/quantize/README.md
@@ -1,3 +1,44 @@
 # quantize

 TODO
+
+## Llama 2 7B
+
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.35
+Q3_K_S | 3.50
+Q3_K_M | 3.91
+Q3_K_L | 4.27
+Q4_K_S | 4.58
+Q4_K_M | 4.84
+Q5_K_S | 5.52
+Q5_K_M | 5.68
+Q6_K | 6.56
+
+## Llama 2 13B
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.34
+Q3_K_S | 3.48
+Q3_K_M | 3.89
+Q3_K_L | 4.26
+Q4_K_S | 4.56
+Q4_K_M | 4.83
+Q5_K_S | 5.51
+Q5_K_M | 5.67
+Q6_K | 6.56
+
+# Llama 2 70B
+
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.40
+Q3_K_S | 3.47
+Q3_K_M | 3.85
+Q3_K_L | 4.19
+Q4_K_S | 4.53
+Q4_K_M | 4.80
+Q5_K_S | 5.50
+Q5_K_M | 5.65
+Q6_K | 6.56
--- a/examples/quantize/quantize.cpp
+++ b/examples/quantize/quantize.cpp
@@ -1,5 +1,5 @@
 #include "build-info.h"
-
+#include "common.h"
 #include "llama.h"

 #include <cstdio>
@@ -40,7 +40,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
 };


-bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
+static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
    std::string ftype_str;

    for (auto ch : ftype_str_in) {
@@ -72,7 +72,7 @@ bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std:
 // usage:
 //  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
-void usage(const char * executable) {
+static void usage(const char * executable) {
    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
    printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
    printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
@@ -161,7 +161,7 @@ int main(int argc, char ** argv) {
        }
    }

-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();

    fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
    if (params.nthread > 0) {
--- a/examples/save-load-state/save-load-state.cpp
+++ b/examples/save-load-state/save-load-state.cpp
@@ -1,6 +1,6 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"

 #include <vector>
 #include <cstdio>
@@ -17,7 +17,7 @@ int main(int argc, char ** argv) {
        return 1;
    }

-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();

    if (params.n_predict < 0) {
        params.n_predict = 16;
@@ -35,11 +35,11 @@ int main(int argc, char ** argv) {
    auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);

    // init
-    auto model = llama_load_model_from_file(params.model.c_str(), lparams);
+    auto * model = llama_load_model_from_file(params.model.c_str(), lparams);
    if (model == nullptr) {
        return 1;
    }
-    auto ctx = llama_new_context_with_model(model, lparams);
+    auto * ctx = llama_new_context_with_model(model, lparams);
    if (ctx == nullptr) {
        llama_free_model(model);
        return 1;
@@ -54,7 +54,7 @@ int main(int argc, char ** argv) {
    }

    // evaluate prompt
-    llama_eval(ctx, tokens.data(), n_prompt_tokens, n_past, params.n_threads);
+    llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt_tokens, n_past, 0), params.n_threads);

    last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
    n_past += n_prompt_tokens;
@@ -78,7 +78,7 @@ int main(int argc, char ** argv) {
    printf("\n%s", params.prompt.c_str());

    for (auto i = 0; i < params.n_predict; i++) {
-        auto logits = llama_get_logits(ctx);
+        auto * logits = llama_get_logits(ctx);
        auto n_vocab = llama_n_vocab(ctx);
        std::vector<llama_token_data> candidates;
        candidates.reserve(n_vocab);
@@ -91,7 +91,7 @@ int main(int argc, char ** argv) {
        last_n_tokens_data.push_back(next_token);

        printf("%s", next_token_str.c_str());
-        if (llama_eval(ctx, &next_token, 1, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0), params.n_threads)) {
            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
            llama_free(ctx);
            llama_free_model(model);
@@ -106,7 +106,7 @@ int main(int argc, char ** argv) {
    llama_free(ctx);

    // make new context
-    auto ctx2 = llama_new_context_with_model(model, lparams);
+    auto * ctx2 = llama_new_context_with_model(model, lparams);

    // Load state (rng, logits, embedding and kv_cache) from file
    {
@@ -138,7 +138,7 @@ int main(int argc, char ** argv) {

    // second run
    for (auto i = 0; i < params.n_predict; i++) {
-        auto logits = llama_get_logits(ctx2);
+        auto * logits = llama_get_logits(ctx2);
        auto n_vocab = llama_n_vocab(ctx2);
        std::vector<llama_token_data> candidates;
        candidates.reserve(n_vocab);
@@ -151,7 +151,7 @@ int main(int argc, char ** argv) {
        last_n_tokens_data.push_back(next_token);

        printf("%s", next_token_str.c_str());
-        if (llama_eval(ctx2, &next_token, 1, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0), params.n_threads)) {
            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
            llama_free(ctx2);
            llama_free_model(model);
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -381,6 +381,10 @@ struct llama_server_context

        // compare the evaluated prompt with the new prompt
        n_past = common_part(embd, prompt_tokens);
+
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, params.n_ctx);
+
        embd = prompt_tokens;
        if (n_past == num_prompt_tokens)
        {
@@ -411,19 +415,27 @@ struct llama_server_context

        if (embd.size() >= (size_t)params.n_ctx)
        {
-            // Reset context
-            const int n_left = (params.n_ctx - params.n_keep) / 2;
+            // Shift context
+
+            const int n_left    = n_past - params.n_keep - 1;
+            const int n_discard = n_left/2;
+
+            llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+            llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+            for (size_t i = params.n_keep + 1 + n_discard; i < embd.size(); i++)
+            {
+                embd[i - n_discard] = embd[i];
+            }
+            embd.resize(embd.size() - n_discard);
+
+            n_past -= n_discard;

-            std::vector<llama_token> new_tokens(embd.begin(), embd.begin() + params.n_keep);
-            new_tokens.insert(new_tokens.end(), embd.end() - n_left, embd.end());
-            embd = new_tokens;
-            n_past = params.n_keep;
            truncated = true;
            LOG_VERBOSE("input truncated", {
                                               {"n_ctx", params.n_ctx},
                                               {"n_keep", params.n_keep},
                                               {"n_left", n_left},
-                                               {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
                                           });
        }

@@ -434,7 +446,8 @@ struct llama_server_context
            {
                n_eval = params.n_batch;
            }
-            if (llama_eval(ctx, &embd[n_past], n_eval, n_past, params.n_threads))
+
+            if (llama_decode(ctx, llama_batch_get_one(&embd[n_past], n_eval, n_past, 0), params.n_threads))
            {
                LOG_ERROR("failed to eval", {
                                                {"n_eval", n_eval},
@@ -523,13 +536,13 @@ struct llama_server_context
                {
                    static float mirostat_mu = 2.0f * mirostat_tau;
                    const int mirostat_m = 100;
-                    llama_sample_temperature(ctx, &candidates_p, temp);
+                    llama_sample_temp(ctx, &candidates_p, temp);
                    result.tok = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
                }
                else if (mirostat == 2)
                {
                    static float mirostat_mu = 2.0f * mirostat_tau;
-                    llama_sample_temperature(ctx, &candidates_p, temp);
+                    llama_sample_temp(ctx, &candidates_p, temp);
                    result.tok = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
                }
                else
@@ -540,7 +553,7 @@ struct llama_server_context
                    llama_sample_tail_free(ctx, &candidates_p, tfs_z, min_keep);
                    llama_sample_typical(ctx, &candidates_p, typical_p, min_keep);
                    llama_sample_top_p(ctx, &candidates_p, top_p, min_keep);
-                    llama_sample_temperature(ctx, &candidates_p, temp);
+                    llama_sample_temp(ctx, &candidates_p, temp);
                    result.tok = llama_sample_token(ctx, &candidates_p);
                }
            }
@@ -701,8 +714,8 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
    printf("  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    printf("  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    printf("  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    printf("  --rope-freq-base N    RoPE base frequency (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency scaling factor (default: loaded from model)\n");
    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
@@ -943,7 +956,23 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                invalid_param = true;
                break;
            }
-            params.lora_adapter = argv[i];
+            params.lora_adapter.push_back({argv[i], 1.0f});
+            params.use_mmap = false;
+        }
+        else if (arg == "--lora-scaled")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            const char * lora_adapter = argv[i];
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter.push_back({lora_adapter, std::stof(argv[i])});
            params.use_mmap = false;
        }
        else if (arg == "--lora-base")
@@ -1083,8 +1112,9 @@ static json format_final_response(llama_server_context &llama, const std::string
    return res;
 }

-static json format_partial_response(llama_server_context &llama, const std::string &content, const std::vector<completion_token_output> &probs)
-{
+static json format_partial_response(
+    llama_server_context &llama, const std::string &content, const std::vector<completion_token_output> &probs
+) {
    json res = json{
        {"content", content},
        {"stop", false},
@@ -1215,7 +1245,7 @@ static void log_server_request(const Request &req, const Response &res)
                           });
 }

-bool is_at_eob(llama_server_context & server_context, const llama_token * tokens, const size_t n_tokens) {
+static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) {
    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
 }

@@ -1225,7 +1255,7 @@ bool is_at_eob(llama_server_context & server_context, const llama_token * tokens
 //  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
 //    This is also called when the stop condition is met.
 //    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
-void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
+static void beam_search_callback(void *callback_data, llama_beams_state beams_state) {
    auto & llama = *static_cast<llama_server_context*>(callback_data);
    // Mark beams as EOS as needed.
    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
@@ -1258,7 +1288,8 @@ struct token_translator {
    std::string operator()(const completion_token_output & cto) const { return (*this)(cto.tok); }
 };

-void append_to_generated_text_from_generated_token_probs(llama_server_context & llama) {
+static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama)
+{
    auto & gtps = llama.generated_token_probs;
    auto translator = token_translator{llama.ctx};
    auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
--- a/examples/simple/CMakeLists.txt
+++ b/examples/simple/CMakeLists.txt
@@ -3,6 +3,3 @@ add_executable(${TARGET} simple.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
--- a/examples/simple/README.md
+++ b/examples/simple/README.md
@@ -0,0 +1,21 @@
+# llama.cpp/example/simple
+
+The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
+
+```bash
+./simple ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is"
+
+...
+
+main: n_len = 32, n_ctx = 2048, n_parallel = 1, n_kv_req = 32
+
+ Hello my name is Shawn and I'm a 20 year old male from the United States. I'm a 20 year old
+
+main: decoded 27 tokens in 2.31 s, speed: 11.68 t/s
+
+llama_print_timings:        load time =   579.15 ms
+llama_print_timings:      sample time =     0.72 ms /    28 runs   (    0.03 ms per token, 38888.89 tokens per second)
+llama_print_timings: prompt eval time =   655.63 ms /    10 tokens (   65.56 ms per token,    15.25 tokens per second)
+llama_print_timings:        eval time =  2180.97 ms /    27 runs   (   80.78 ms per token,    12.38 tokens per second)
+llama_print_timings:       total time =  2891.13 ms
+```
--- a/examples/simple/simple.cpp
+++ b/examples/simple/simple.cpp
@@ -1,5 +1,3 @@
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"

@@ -28,12 +26,18 @@ int main(int argc, char ** argv) {
        params.prompt = "Hello my name is";
    }

+    // total length of the sequence including the prompt
+    const int n_len = 32;
+
    // init LLM

    llama_backend_init(params.numa);

    llama_context_params ctx_params = llama_context_default_params();

+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = 2048;
+
    llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);

    if (model == NULL) {
@@ -43,20 +47,31 @@ int main(int argc, char ** argv) {

    llama_context * ctx = llama_new_context_with_model(model, ctx_params);

+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
    // tokenize the prompt

    std::vector<llama_token> tokens_list;
    tokens_list = ::llama_tokenize(ctx, params.prompt, true);

-    const int max_context_size     = llama_n_ctx(ctx);
-    const int max_tokens_list_size = max_context_size - 4;
+    const int n_ctx    = llama_n_ctx(ctx);
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size());

-    if ((int) tokens_list.size() > max_tokens_list_size) {
-        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) tokens_list.size(), max_tokens_list_size);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, n_kv_req);
+
+    // make sure the KV cache is big enough to hold all the prompt and generated tokens
+    if (n_kv_req > n_ctx) {
+        LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
        return 1;
    }

-    fprintf(stderr, "\n\n");
+    // print the prompt token-by-token
+
+    fprintf(stderr, "\n");

    for (auto id : tokens_list) {
        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
@@ -64,63 +79,104 @@ int main(int argc, char ** argv) {

    fflush(stderr);

+    // create a llama_batch with size 512
+    // we use this object to submit token data for decoding
+
+    llama_batch batch = llama_batch_init(512, 0);
+
+    // evaluate the initial prompt
+    batch.n_tokens = tokens_list.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
+    }
+
+    // llama_decode will output logits only for the last token of the prompt
+    batch.logits[batch.n_tokens - 1] = true;
+
+    if (llama_decode(ctx, batch, params.n_threads) != 0) {
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        return 1;
+    }
+
    // main loop

-    // The LLM keeps a contextual cache memory of previous token evaluation.
-    // Usually, once this cache is full, it is required to recompute a compressed context based on previous
-    // tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
-    // example, we will just stop the loop once this cache is full or once an end of stream is detected.
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;

-    const int n_gen = std::min(32, max_context_size);
+    const auto t_main_start = ggml_time_us();

-    while (llama_get_kv_cache_token_count(ctx) < n_gen) {
-        // evaluate the transformer
+    while (n_cur <= n_len) {
+        // sample the next token
+        {
+            auto   n_vocab = llama_n_vocab(ctx);
+            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);

-        if (llama_eval(ctx, tokens_list.data(), int(tokens_list.size()), llama_get_kv_cache_token_count(ctx), params.n_threads)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            // sample the most likely token
+            const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream?
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+                LOG_TEE("\n");
+
+                break;
+            }
+
+            LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+            fflush(stdout);
+
+            // prepare the next batch
+            batch.n_tokens = 0;
+
+            // push this new token for next evaluation
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = 0;
+            batch.logits[batch.n_tokens] = true;
+
+            batch.n_tokens += 1;
+
+            n_decode += 1;
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch, params.n_threads)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
            return 1;
        }
-
-        tokens_list.clear();
-
-        // sample the next token
-
-        llama_token new_token_id = 0;
-
-        auto logits  = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-        }
-
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-        new_token_id = llama_sample_token_greedy(ctx , &candidates_p);
-
-        // is it an end of stream ?
-        if (new_token_id == llama_token_eos(ctx)) {
-            fprintf(stderr, " [end of text]\n");
-            break;
-        }
-
-        // print the new token :
-        printf("%s", llama_token_to_piece(ctx, new_token_id).c_str());
-        fflush(stdout);
-
-        // push this new token for next evaluation
-        tokens_list.push_back(new_token_id);
    }

+    LOG_TEE("\n");
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
    llama_free(ctx);
    llama_free_model(model);

    llama_backend_free();

-    fprintf(stderr, "\n\n");
-
    return 0;
 }
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -37,11 +37,12 @@ int main(int argc, char ** argv) {
    llama_context * ctx_dft = NULL;

    // load the target model
-    params.perplexity = true; // HACK: enable logits_all = true
+    params.logits_all = true;
    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);

    // load the draft model
    params.model = params.model_draft;
+    params.n_gpu_layers = params.n_gpu_layers_draft;
    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);

    // tokenize the prompt
@@ -69,9 +70,9 @@ int main(int argc, char ** argv) {
    const auto t_enc_start = ggml_time_us();

    // eval the prompt with both models
-    llama_eval(ctx_tgt,  inp.data(), int(inp.size() - 1), 0, params.n_threads);
-    llama_eval(ctx_tgt, &inp.back(),      1, inp.size() - 1, params.n_threads);
-    llama_eval(ctx_dft,  inp.data(),     int(inp.size()), 0, params.n_threads);
+    llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0,           0), params.n_threads);
+    llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0), params.n_threads);
+    llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input,     0,           0), params.n_threads);

    const auto t_enc_end = ggml_time_us();

@@ -81,7 +82,7 @@ int main(int argc, char ** argv) {
    //GGML_ASSERT(n_vocab == llama_n_vocab(ctx_dft));

    // how many tokens to draft each time
-    const int n_draft = params.n_draft;
+    int n_draft = params.n_draft;

    int n_predict = 0;
    int n_drafted = 0;
@@ -130,9 +131,10 @@ int main(int argc, char ** argv) {
        LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted));

        int i_dft = 0;
+
        while (true) {
            // sample from the target model
-            const llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);
+            llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);

            // remember which tokens were sampled - used for repetition penalties during sampling
            last_tokens.erase(last_tokens.begin());
@@ -170,9 +172,31 @@ int main(int argc, char ** argv) {
                LOG("out of drafted tokens\n");
            }

-            llama_eval(ctx_dft, &id, 1, n_past_dft, params.n_threads);
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, n_ctx);
+            llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0), params.n_threads);
            ++n_past_dft;

+            // heuristic for n_draft
+            {
+                const int  n_draft_cur  = (int) drafted.size();
+                const bool all_accepted = i_dft == n_draft_cur;
+
+                LOG("n_draft      = %d\n", n_draft);
+                LOG("n_draft_cur  = %d\n", n_draft_cur);
+                LOG("i_dft        = %d\n", i_dft);
+                LOG("all_accepted = %d\n", all_accepted);
+
+                if (all_accepted && n_draft == n_draft_cur) {
+                    LOG(" - max drafted tokens accepted - n_draft += 8\n");
+                    n_draft = std::min(30, n_draft + 8);
+                } else if (all_accepted) {
+                    LOG(" - partially drafted tokens accepted - no change\n");
+                } else {
+                    LOG(" - drafted token rejected - n_draft -= 1\n");
+                    n_draft = std::max(2, n_draft - 1);
+                }
+            }
+
            drafted.clear();
            drafted.push_back(id);

@@ -233,7 +257,8 @@ int main(int argc, char ** argv) {
            }

            // evaluate the drafted token on the draft model
-            llama_eval(ctx_dft, &drafted.back(), 1, n_past_cur, params.n_threads);
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, n_ctx);
+            llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0), params.n_threads);
            ++n_past_cur;

            if (grammar_dft != NULL) {
@@ -242,7 +267,8 @@ int main(int argc, char ** argv) {
        }

        // evaluate the target model on the drafted tokens
-        llama_eval(ctx_tgt, drafted.data(), drafted.size(), n_past_tgt, params.n_threads);
+        llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, n_ctx);
+        llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0), params.n_threads);
        ++n_past_tgt;

        // the first token is always proposed by the traget model before the speculation loop
--- a/examples/train-text-from-scratch/README.md
+++ b/examples/train-text-from-scratch/README.md
@@ -10,9 +10,9 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
 ./bin/train-text-from-scratch \
        --vocab-model ../models/ggml-vocab-llama.gguf \
        --ctx 64 --embd 256 --head 8 --layer 16 \
-        --checkpoint-in  chk-shakespeare-256x16.gguf \
-        --checkpoint-out chk-shakespeare-256x16.gguf \
-        --model-out ggml-shakespeare-256x16-f32.gguf \
+        --checkpoint-in  chk-shakespeare-256x16-LATEST.gguf \
+        --checkpoint-out chk-shakespeare-256x16-ITERATION.gguf \
+        --model-out ggml-shakespeare-256x16-f32-ITERATION.gguf \
        --train-data "shakespeare.txt" \
        -t 6 -b 16 --seed 1 --adam-iter 256 \
        --no-checkpointing
@@ -20,3 +20,8 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
 # predict
 ./bin/main -m ggml-shakespeare-256x16-f32.gguf
 ```
+
+Output files will be saved every N iterations (config with `--save-every N`).
+The pattern "ITERATION" in the output filenames will be replaced with the iteration number and "LATEST" for the latest output.
+
+To train GGUF models just pass them to `--checkpoint-in FN`.
--- a/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py
+++ b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py
@@ -47,10 +47,13 @@ LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys"
 LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s"
 LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y"

-LLM_KV_TRAINING_FILE_VERSION    = "training.file_version"
-LLM_KV_TRAINING_ITERATION_COUNT = "training.iteration_count"
-LLM_KV_TRAINING_SAMPLE_COUNT    = "training.sample_count"
-LLM_KV_TRAINING_TOKEN_COUNT     = "training.token_count"
+LLM_KV_TRAINING_TYPE_TRAIN_MODEL   = "train_model"
+LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"
+LLM_KV_TRAINING_TYPE               = "training.type"
+LLM_KV_TRAINING_FILE_VERSION       = "training.file_version"
+LLM_KV_TRAINING_ITERATION_COUNT    = "training.iteration_count"
+LLM_KV_TRAINING_SAMPLE_COUNT       = "training.sample_count"
+LLM_KV_TRAINING_TOKEN_COUNT        = "training.token_count"

 class Tensor:
    def __init__(self, dtype='f', ne=None):
@@ -460,6 +463,7 @@ class Checkpoint:
        gguf_writer.add_file_type(gguf.GGMLQuantizationType.F32)
        gguf_writer.add_layer_norm_rms_eps(1e-5)
        gguf_writer.add_uint32(LLM_KV_TRAINING_FILE_VERSION,    0)
+        gguf_writer.add_string(LLM_KV_TRAINING_TYPE,            LLM_KV_TRAINING_TYPE_TRAIN_MODEL)
        gguf_writer.add_uint32(LLM_KV_TRAINING_ITERATION_COUNT, self.train_its)
        gguf_writer.add_uint32(LLM_KV_TRAINING_SAMPLE_COUNT,    self.train_samples)
        gguf_writer.add_uint32(LLM_KV_TRAINING_TOKEN_COUNT,     self.train_tokens)
--- a/examples/train-text-from-scratch/train-text-from-scratch.cpp
+++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp
--- a/flake.nix
+++ b/flake.nix
@@ -34,7 +34,21 @@
            with pkgs; [ openblas ]
        );
        pkgs = import nixpkgs { inherit system; };
-        nativeBuildInputs = with pkgs; [ cmake ninja pkgconfig ];
+        nativeBuildInputs = with pkgs; [ cmake ninja pkg-config ];
+        cudatoolkit_joined = with pkgs; symlinkJoin {
+          # HACK(Green-Sky): nix currently has issues with cmake findcudatoolkit
+          # see https://github.com/NixOS/nixpkgs/issues/224291
+          # copied from jaxlib
+          name = "${cudaPackages.cudatoolkit.name}-merged";
+          paths = [
+            cudaPackages.cudatoolkit.lib
+            cudaPackages.cudatoolkit.out
+          ] ++ lib.optionals (lib.versionOlder cudaPackages.cudatoolkit.version "11") [
+            # for some reason some of the required libs are in the targets/x86_64-linux
+            # directory; not sure why but this works around it
+            "${cudaPackages.cudatoolkit}/targets/${system}"
+          ];
+        };
        llama-python =
          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
        postPatch = ''
@@ -45,12 +59,15 @@
        postInstall = ''
          mv $out/bin/main $out/bin/llama
          mv $out/bin/server $out/bin/llama-server
+          mkdir -p $out/include
+          cp ${src}/llama.h $out/include/
        '';
        cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
      in
      {
        packages.default = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs buildInputs postInstall;
+          inherit name src meta postPatch nativeBuildInputs postInstall;
+          buildInputs = osSpecific;
          cmakeFlags = cmakeFlags
            ++ (if isAarch64 && isDarwin then [
            "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
@@ -67,6 +84,13 @@
            "-DLLAMA_CLBLAST=ON"
          ];
        };
+        packages.cuda = pkgs.stdenv.mkDerivation {
+          inherit name src meta postPatch nativeBuildInputs postInstall;
+          buildInputs = with pkgs; buildInputs ++ [ cudatoolkit_joined ];
+          cmakeFlags = cmakeFlags ++ [
+            "-DLLAMA_CUBLAS=ON"
+          ];
+        };
        packages.rocm = pkgs.stdenv.mkDerivation {
          inherit name src meta postPatch nativeBuildInputs postInstall;
          buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ];
@@ -93,6 +117,10 @@
          type = "app";
          program = "${self.packages.${system}.default}/bin/quantize";
        };
+        apps.train-text-from-scratch = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/train-text-from-scratch";
+        };
        apps.default = self.apps.${system}.llama;
        devShells.default = pkgs.mkShell {
          buildInputs = [ llama-python ];
--- a/ggml-alloc.c
+++ b/ggml-alloc.c
@@ -77,7 +77,7 @@ struct free_block {
    size_t size;
 };

-#define MAX_FREE_BLOCKS 128
+#define MAX_FREE_BLOCKS 256

 struct ggml_allocr {
    void * data;
@@ -131,6 +131,10 @@ static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_ten
    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
 }

+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
 #ifdef GGML_ALLOCATOR_DEBUG
    GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
@@ -183,6 +187,7 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
    }

    tensor->data = addr;
+    AT_PRINTF("%s: allocated data at %p\n", __func__, tensor->data);

 #ifdef GGML_ALLOCATOR_DEBUG
    add_allocated_tensor(alloc, tensor);
@@ -214,7 +219,8 @@ static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tens

    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
    size = aligned_offset(NULL, size, alloc->alignment);
-    AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
+    AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
+    AT_PRINTF("%s: alloc->data = %p alloc->data+alloc->size = %p alloc->data+alloc->max_size = %p\n", __func__, alloc->data, (char*)alloc->data + alloc->size, (char*)alloc->data + alloc->max_size);

 #ifdef GGML_ALLOCATOR_DEBUG
    remove_allocated_tensor(alloc, tensor);
@@ -338,8 +344,8 @@ static void free_vmem(void * base_addr, size_t size) {

 // allocate uncommitted virtual memory to measure the size of the graph
 static void alloc_measure_vmem(void ** base_addr, size_t * size) {
-    // 1TB for 64-bit, 1GB for 32-bit
-    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<40;
+    // 128GB for 64-bit, 1GB for 32-bit
+    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
    do {
        *base_addr = alloc_vmem(*size);
        if (*base_addr != NULL) {
@@ -399,10 +405,6 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {

 //////////// compute graph allocator

-static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->view_src != NULL;
-}
-
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
    if (a->type != b->type) {
        return false;
@@ -631,3 +633,7 @@ static size_t ggml_allocr_alloc_graph_tensors_n(
 size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
 }
+
+size_t ggml_allocr_max_size(struct ggml_allocr * alloc) {
+    return alloc->max_size;
+}
--- a/ggml-alloc.h
+++ b/ggml-alloc.h
@@ -19,6 +19,7 @@ GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
 GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
 GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
 GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
+GGML_API size_t ggml_allocr_max_size(struct ggml_allocr * alloc);


 #ifdef  __cplusplus
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
--- a/ggml-cuda.h
+++ b/ggml-cuda.h
@@ -31,6 +31,7 @@ GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tens

 GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
+GGML_API void   ggml_cuda_copy_to_device(struct ggml_tensor * tensor);

 GGML_API void   ggml_cuda_set_main_device(int main_device);
 GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
--- a/ggml-metal.h
+++ b/ggml-metal.h
@@ -19,6 +19,8 @@

 #pragma once

+#include "ggml.h"
+
 #include <stddef.h>
 #include <stdbool.h>

@@ -33,6 +35,8 @@ struct ggml_cgraph;
 extern "C" {
 #endif

+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
 struct ggml_metal_context;

 // number of command buffers to use
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -11,11 +11,14 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))

-// TODO: temporary - reuse llama.cpp logging
 #ifdef GGML_METAL_NDEBUG
-#define metal_printf(...)
+#define GGML_METAL_LOG_INFO(...)
+#define GGML_METAL_LOG_WARN(...)
+#define GGML_METAL_LOG_ERROR(...)
 #else
-#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
+#define GGML_METAL_LOG_INFO(...)  ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
+#define GGML_METAL_LOG_WARN(...)  ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
+#define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
 #endif

 #define UNUSED(x) (void)(x)
@@ -63,7 +66,10 @@ struct ggml_metal_context {
    GGML_METAL_DECL_KERNEL(relu);
    GGML_METAL_DECL_KERNEL(gelu);
    GGML_METAL_DECL_KERNEL(soft_max);
+    GGML_METAL_DECL_KERNEL(soft_max_4);
    GGML_METAL_DECL_KERNEL(diag_mask_inf);
+    GGML_METAL_DECL_KERNEL(diag_mask_inf_8);
+    GGML_METAL_DECL_KERNEL(get_rows_f32);
    GGML_METAL_DECL_KERNEL(get_rows_f16);
    GGML_METAL_DECL_KERNEL(get_rows_q4_0);
    GGML_METAL_DECL_KERNEL(get_rows_q4_1);
@@ -75,8 +81,10 @@ struct ggml_metal_context {
    GGML_METAL_DECL_KERNEL(get_rows_q6_K);
    GGML_METAL_DECL_KERNEL(rms_norm);
    GGML_METAL_DECL_KERNEL(norm);
+    GGML_METAL_DECL_KERNEL(mul_mat_f32_f32);
    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_l4);
    GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
    GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
    GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
@@ -85,6 +93,7 @@ struct ggml_metal_context {
    GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
    GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
    GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
    GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
    GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
    GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
@@ -94,7 +103,8 @@ struct ggml_metal_context {
    GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
    GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
    GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
-    GGML_METAL_DECL_KERNEL(rope);
+    GGML_METAL_DECL_KERNEL(rope_f32);
+    GGML_METAL_DECL_KERNEL(rope_f16);
    GGML_METAL_DECL_KERNEL(alibi_f32);
    GGML_METAL_DECL_KERNEL(cpy_f32_f16);
    GGML_METAL_DECL_KERNEL(cpy_f32_f32);
@@ -114,22 +124,54 @@ static NSString * const msl_library_source = @"see metal.metal";
@implementation GGMLMetalClass
@end

-struct ggml_metal_context * ggml_metal_init(int n_cb) {
-    metal_printf("%s: allocating\n", __func__);
+ggml_log_callback ggml_metal_log_callback = NULL;
+void * ggml_metal_log_user_data = NULL;
+
+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
+    ggml_metal_log_callback  = log_callback;
+    ggml_metal_log_user_data = user_data;
+}
+
+static void ggml_metal_log(enum ggml_log_level level, const char* format, ...){
+    if (ggml_metal_log_callback != NULL) {
+        va_list args;
+        va_start(args, format);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            ggml_metal_log_callback(level, buffer, ggml_metal_log_user_data);
+        } else {
+            char* buffer2 = malloc(len+1);
+            vsnprintf(buffer2, len+1, format, args);
+            buffer2[len] = 0;
+            ggml_metal_log_callback(level, buffer2, ggml_metal_log_user_data);
+            free(buffer2);
+        }
+        va_end(args);
+    }
+}
+
+
+
+struct ggml_metal_context * ggml_metal_init(int n_cb) {
+    GGML_METAL_LOG_INFO("%s: allocating\n", __func__);

-    // Show all the Metal device instances in the system
-    NSArray * devices = MTLCopyAllDevices();
    id <MTLDevice> device;
    NSString * s;
+
+#if TARGET_OS_OSX
+    // Show all the Metal device instances in the system
+    NSArray * devices = MTLCopyAllDevices();
    for (device in devices) {
        s = [device name];
-        metal_printf("%s: found device: %s\n", __func__, [s UTF8String]);
+        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]);
    }
+#endif

    // Pick and show default Metal device
    device = MTLCreateSystemDefaultDevice();
    s = [device name];
-    metal_printf("%s: picking default device: %s\n", __func__, [s UTF8String]);
+    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]);

    // Configure context
    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
@@ -139,16 +181,24 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
    ctx->n_buffers = 0;
    ctx->concur_list_len = 0;

-    ctx->d_queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
+    ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);

-#if 0
-    // compile from source string and show compile log
+#ifdef GGML_SWIFT
+    // load the default.metallib file
    {
        NSError * error = nil;

-        ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
+        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+        NSString * llamaBundlePath = [bundle pathForResource:@"llama_llama" ofType:@"bundle"];
+        NSBundle * llamaBundle = [NSBundle bundleWithPath:llamaBundlePath];
+        NSString * libPath = [llamaBundle pathForResource:@"default" ofType:@"metallib"];
+        NSURL * libURL = [NSURL fileURLWithPath:libPath];
+
+        // Load the metallib file into a Metal library
+        ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
+
        if (error) {
-            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
            return NULL;
        }
    }
@@ -161,12 +211,12 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {

        //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-        NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
-        metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]);
+        NSString * path   = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+        GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path UTF8String]);

        NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
        if (error) {
-            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
            return NULL;
        }

@@ -178,7 +228,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
        ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
 #endif
        if (error) {
-            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
            return NULL;
        }
    }
@@ -190,11 +240,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #define GGML_METAL_ADD_KERNEL(name) \
        ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
-        metal_printf("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
+        GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
                (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
                (int) ctx->pipeline_##name.threadExecutionWidth); \
        if (error) { \
-            metal_printf("%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+          GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
            return NULL; \
        }

@@ -207,7 +257,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
        GGML_METAL_ADD_KERNEL(relu);
        GGML_METAL_ADD_KERNEL(gelu);
        GGML_METAL_ADD_KERNEL(soft_max);
+        GGML_METAL_ADD_KERNEL(soft_max_4);
        GGML_METAL_ADD_KERNEL(diag_mask_inf);
+        GGML_METAL_ADD_KERNEL(diag_mask_inf_8);
+        GGML_METAL_ADD_KERNEL(get_rows_f32);
        GGML_METAL_ADD_KERNEL(get_rows_f16);
        GGML_METAL_ADD_KERNEL(get_rows_q4_0);
        GGML_METAL_ADD_KERNEL(get_rows_q4_1);
@@ -219,8 +272,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
        GGML_METAL_ADD_KERNEL(get_rows_q6_K);
        GGML_METAL_ADD_KERNEL(rms_norm);
        GGML_METAL_ADD_KERNEL(norm);
+        GGML_METAL_ADD_KERNEL(mul_mat_f32_f32);
        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_l4);
        GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
        GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
        GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
@@ -229,6 +284,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
        GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
        GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
        GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
        GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
        GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
        GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
@@ -238,7 +294,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
        GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
        GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
        GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
-        GGML_METAL_ADD_KERNEL(rope);
+        GGML_METAL_ADD_KERNEL(rope_f32);
+        GGML_METAL_ADD_KERNEL(rope_f16);
        GGML_METAL_ADD_KERNEL(alibi_f32);
        GGML_METAL_ADD_KERNEL(cpy_f32_f16);
        GGML_METAL_ADD_KERNEL(cpy_f32_f32);
@@ -247,19 +304,21 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #undef GGML_METAL_ADD_KERNEL
    }

-    metal_printf("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-    metal_printf("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+#if TARGET_OS_OSX
+    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
    if (ctx->device.maxTransferRate != 0) {
-        metal_printf("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
    } else {
-        metal_printf("%s: maxTransferRate               = built-in GPU\n", __func__);
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
    }
+#endif

    return ctx;
 }

 void ggml_metal_free(struct ggml_metal_context * ctx) {
-    metal_printf("%s: deallocating\n", __func__);
+    GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 #define GGML_METAL_DEL_KERNEL(name) \
    [ctx->function_##name release]; \
    [ctx->pipeline_##name release];
@@ -273,7 +332,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
    GGML_METAL_DEL_KERNEL(relu);
    GGML_METAL_DEL_KERNEL(gelu);
    GGML_METAL_DEL_KERNEL(soft_max);
+    GGML_METAL_DEL_KERNEL(soft_max_4);
    GGML_METAL_DEL_KERNEL(diag_mask_inf);
+    GGML_METAL_DEL_KERNEL(diag_mask_inf_8);
+    GGML_METAL_DEL_KERNEL(get_rows_f32);
    GGML_METAL_DEL_KERNEL(get_rows_f16);
    GGML_METAL_DEL_KERNEL(get_rows_q4_0);
    GGML_METAL_DEL_KERNEL(get_rows_q4_1);
@@ -285,8 +347,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
    GGML_METAL_DEL_KERNEL(get_rows_q6_K);
    GGML_METAL_DEL_KERNEL(rms_norm);
    GGML_METAL_DEL_KERNEL(norm);
+    GGML_METAL_DEL_KERNEL(mul_mat_f32_f32);
    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row);
+    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_l4);
    GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
    GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
    GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
@@ -295,6 +359,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
    GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32);
    GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32);
    GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
    GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
    GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
    GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
@@ -304,7 +369,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
    GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
    GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
    GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
-    GGML_METAL_DEL_KERNEL(rope);
+    GGML_METAL_DEL_KERNEL(rope_f32);
+    GGML_METAL_DEL_KERNEL(rope_f16);
    GGML_METAL_DEL_KERNEL(alibi_f32);
    GGML_METAL_DEL_KERNEL(cpy_f32_f16);
    GGML_METAL_DEL_KERNEL(cpy_f32_f32);
@@ -329,7 +395,7 @@ void * ggml_metal_host_malloc(size_t n) {
    void * data = NULL;
    const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
    if (result != 0) {
-        metal_printf("%s: error: posix_memalign failed\n", __func__);
+        GGML_METAL_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
        return NULL;
    }

@@ -357,7 +423,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
 // Metal buffer based on the host memory pointer
 //
 static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
-    //metal_printf("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+    //GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);

    const int64_t tsize = ggml_nbytes(t);

@@ -365,16 +431,17 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
    for (int i = 0; i < ctx->n_buffers; ++i) {
        const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;

+        //metal_printf("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
        if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
            *offs = (size_t) ioffs;

-            //metal_printf("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+            //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);

            return ctx->buffers[i].metal;
        }
    }

-    metal_printf("%s: error: buffer is nil\n", __func__);
+    GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__);

    return nil;
 }
@@ -386,7 +453,7 @@ bool ggml_metal_add_buffer(
                         size_t   size,
                         size_t   max_size) {
    if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
-        metal_printf("%s: too many buffers\n", __func__);
+        GGML_METAL_LOG_ERROR("%s: error: too many buffers\n", __func__);
        return false;
    }

@@ -396,7 +463,7 @@ bool ggml_metal_add_buffer(
            const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;

            if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
-                metal_printf("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                GGML_METAL_LOG_ERROR("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
                return false;
            }
        }
@@ -417,11 +484,11 @@ bool ggml_metal_add_buffer(
            ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];

            if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
+                GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
                return false;
            }

-            metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+            GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);

            ++ctx->n_buffers;
        } else {
@@ -441,28 +508,32 @@ bool ggml_metal_add_buffer(
                ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];

                if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                    metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
+                    GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
                    return false;
                }

-                metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
                if (i + size_step < size) {
-                    metal_printf("\n");
+                    GGML_METAL_LOG_INFO("\n");
                }

                ++ctx->n_buffers;
            }
        }

-        metal_printf(", (%8.2f / %8.2f)",
+#if TARGET_OS_OSX
+        GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
                ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
                ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);

        if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
-            metal_printf(", warning: current allocated size is greater than the recommended max working set size\n");
+            GGML_METAL_LOG_WARN(", warning: current allocated size is greater than the recommended max working set size\n", __func__);
        } else {
-            metal_printf("\n");
+            GGML_METAL_LOG_INFO("\n");
        }
+#else
+        GGML_METAL_LOG_INFO(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
    }

    return true;
@@ -574,7 +645,7 @@ void ggml_metal_graph_find_concurrency(
    }

    if (ctx->concur_list_len > GGML_MAX_CONCUR) {
-        metal_printf("%s: too many elements for metal ctx->concur_list!\n", __func__);
+        GGML_METAL_LOG_WARN("%s: too many elements for metal ctx->concur_list!\n", __func__);
    }
 }

@@ -628,7 +699,7 @@ void ggml_metal_graph_compute(
                    continue;
                }

-                //metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+                //GGML_METAL_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));

                struct ggml_tensor * src0 = gf->nodes[i]->src[0];
                struct ggml_tensor * src1 = gf->nodes[i]->src[1];
@@ -672,17 +743,17 @@ void ggml_metal_graph_compute(
                id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
                id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;

-                //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+                //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
                //if (src0) {
-                //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+                //    GGML_METAL_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
                //            ggml_is_contiguous(src0), src0->name);
                //}
                //if (src1) {
-                //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+                //    GGML_METAL_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
                //            ggml_is_contiguous(src1), src1->name);
                //}
                //if (dst) {
-                //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+                //    GGML_METAL_LOG_INFO("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
                //            dst->name);
                //}

@@ -698,29 +769,66 @@ void ggml_metal_graph_compute(
                    case GGML_OP_ADD:
                        {
                            GGML_ASSERT(ggml_is_contiguous(src0));
+                            GGML_ASSERT(ggml_is_contiguous(src1));

-                            // utilize float4
-                            GGML_ASSERT(ne00 % 4 == 0);
-                            const int64_t nb = ne00/4;
+                            bool bcast_row = false;

-                            if (ggml_nelements(src1) == ne10) {
+                            int64_t nb = ne00;
+
+                            if (ggml_nelements(src1) == ne10 && ne00 % 4 == 0) {
                                // src1 is a row
+                                GGML_ASSERT(ne11 == 1);
+
+                                nb = ne00 / 4;
                                [encoder setComputePipelineState:ctx->pipeline_add_row];
+
+                                bcast_row = true;
                            } else {
                                [encoder setComputePipelineState:ctx->pipeline_add];
                            }
                            [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:&nb     length:sizeof(nb) atIndex:3];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) 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:&nb13 length:sizeof(nb13) atIndex:18];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];

-                            const int64_t n = ggml_nelements(dst)/4;
+                            if (bcast_row) {
+                                const int64_t n = ggml_nelements(dst)/4;

-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            } else {
+                                const int nth = MIN(1024, ne0);
+
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                            }
                        } break;
                    case GGML_OP_MUL:
                        {
                            GGML_ASSERT(ggml_is_contiguous(src0));
+                            GGML_ASSERT(ggml_is_contiguous(src1));

                            // utilize float4
                            GGML_ASSERT(ne00 % 4 == 0);
@@ -728,6 +836,7 @@ void ggml_metal_graph_compute(

                            if (ggml_nelements(src1) == ne10) {
                                // src1 is a row
+                                GGML_ASSERT(ne11 == 1);
                                [encoder setComputePipelineState:ctx->pipeline_mul_row];
                            } else {
                                [encoder setComputePipelineState:ctx->pipeline_mul];
@@ -743,6 +852,8 @@ void ggml_metal_graph_compute(
                        } break;
                    case GGML_OP_SCALE:
                        {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+
                            const float scale = *(const float *) src1->data;

                            [encoder setComputePipelineState:ctx->pipeline_scale];
@@ -750,7 +861,7 @@ void ggml_metal_graph_compute(
                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                            [encoder setBytes:&scale length:sizeof(scale) atIndex:2];

-                            const int64_t n = ggml_nelements(dst);
+                            const int64_t n = ggml_nelements(dst)/4;

                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                        } break;
@@ -762,7 +873,7 @@ void ggml_metal_graph_compute(
                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];

-                                    const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst)/4;

                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                } break;
@@ -782,27 +893,30 @@ void ggml_metal_graph_compute(
                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];

-                                    const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst)/4;

                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                } break;
                            default:
                                {
-                                    metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                                    GGML_ASSERT(false);
                                }
                        } break;
                    case GGML_OP_SOFT_MAX:
                        {
-                            const int nth = 32;
+                            const int nth = MIN(32, ne00);

-                            [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            if (ne00%4 == 0) {
+                                [encoder setComputePipelineState:ctx->pipeline_soft_max_4];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            }
                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];

                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                        } break;
@@ -810,14 +924,23 @@ void ggml_metal_graph_compute(
                        {
                            const int n_past = ((int32_t *)(dst->op_params))[0];

-                            [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            if (ne00%8 == 0) {
+                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf_8];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            }
                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                            [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
                            [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
                            [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];

-                            [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            if (ne00%8 == 0) {
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            }
+                            else {
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            }
                        } break;
                    case GGML_OP_MUL_MAT:
                        {
@@ -830,13 +953,14 @@ void ggml_metal_graph_compute(

                            // 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
-                            if (ggml_is_contiguous(src0) &&
-                                ggml_is_contiguous(src1) &&
+                            if (!ggml_is_transposed(src0) &&
+                                !ggml_is_transposed(src1) &&
                                src1t == GGML_TYPE_F32 &&
                                [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
                                ne00%32 == 0 &&
-                                ne11 > 1) {
+                                ne11 > 2) {
                                switch (src0->type) {
+                                    case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f32_f32];  break;
                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
                                    case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
                                    case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
@@ -856,25 +980,38 @@ void ggml_metal_graph_compute(
                                [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
                                [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
                                [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
-                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:8];
-                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:9];
-                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:10];
+                                [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:8];
+                                [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:9];
+                                [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:10];
+                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:11];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
+                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:13];
                                [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                            } else {
                                int nth0 = 32;
                                int nth1 = 1;
+                                int nrows = 1;

                                // use custom matrix x vector kernel
                                switch (src0t) {
+                                    case GGML_TYPE_F32:
+                                        {
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f32_f32];
+                                            nrows = 4;
+                                        } break;
                                    case GGML_TYPE_F16:
                                        {
                                            nth0 = 32;
                                            nth1 = 1;
                                            if (ne11 * ne12 < 4) {
                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row];
+                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_l4];
+                                                nrows = ne11;
                                            } else {
                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                                nrows = 4;
                                            }
                                        } break;
                                    case GGML_TYPE_Q4_0:
@@ -951,7 +1088,7 @@ void ggml_metal_graph_compute(
                                        } break;
                                    default:
                                        {
-                                            metal_printf("Asserting on type %d\n",(int)src0t);
+                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
                                            GGML_ASSERT(false && "not implemented");
                                        }
                                };
@@ -995,7 +1132,7 @@ void ggml_metal_graph_compute(
                                else if (src0t == GGML_TYPE_Q6_K) {
                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                } else {
-                                    int64_t ny = (ne11 + 3)/4;
+                                    int64_t ny = (ne11 + nrows - 1)/nrows;
                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                }
                            }
@@ -1003,6 +1140,7 @@ void ggml_metal_graph_compute(
                    case GGML_OP_GET_ROWS:
                        {
                            switch (src0->type) {
+                                case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f32];  break;
                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
                                case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
                                case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
@@ -1018,9 +1156,9 @@ void 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:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
-                            [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:5];

                            const int64_t n = ggml_nelements(src1);

@@ -1031,7 +1169,7 @@ void ggml_metal_graph_compute(
                            float eps;
                            memcpy(&eps, dst->op_params, sizeof(float));

-                            const int nth = 512;
+                            const int nth = MIN(512, ne00);

                            [encoder setComputePipelineState:ctx->pipeline_rms_norm];
                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -1050,7 +1188,7 @@ void ggml_metal_graph_compute(
                            float eps;
                            memcpy(&eps, dst->op_params, sizeof(float));

-                            const int nth = 256;
+                            const int nth = MIN(256, ne00);

                            [encoder setComputePipelineState:ctx->pipeline_norm];
                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
@@ -1068,6 +1206,8 @@ void ggml_metal_graph_compute(
                        {
                            GGML_ASSERT((src0t == GGML_TYPE_F32));

+                            const int nth = MIN(1024, ne00);
+
                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
                            const int n_head = ((int32_t *) dst->op_params)[1];
                            float max_bias;
@@ -1101,12 +1241,14 @@ void ggml_metal_graph_compute(
                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
                            [encoder setBytes:&m0  length:sizeof(    float) atIndex:18];

-                            const int nth = 32;
-
                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                        } break;
                    case GGML_OP_ROPE:
                        {
+                            GGML_ASSERT(ne10 == ne02);
+
+                            const int nth = MIN(1024, ne00);
+
                            const int n_past = ((int32_t *) dst->op_params)[0];
                            const int n_dims = ((int32_t *) dst->op_params)[1];
                            const int mode   = ((int32_t *) dst->op_params)[2];
@@ -1116,38 +1258,44 @@ void ggml_metal_graph_compute(
                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));

-                            [encoder setComputePipelineState:ctx->pipeline_rope];
-                            [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:&ne01    length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:18];
-                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:20];
-                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:21];
-                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
+                            switch (src0->type) {
+                                case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_rope_f32]; break;
+                                case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_rope_f16]; break;
+                                default: GGML_ASSERT(false);
+                            };

-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
+                            [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:&ne00    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:6];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:10];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:14];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:18];
+                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:21];
+                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:22];
+                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:23];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                        } break;
                    case GGML_OP_DUP:
                    case GGML_OP_CPY:
                    case GGML_OP_CONT:
                        {
-                            const int nth = 32;
+                            const int nth = MIN(1024, ne00);

                            switch (src0t) {
                                case GGML_TYPE_F32:
@@ -1192,7 +1340,7 @@ void ggml_metal_graph_compute(
                        } break;
                    default:
                        {
-                            metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                            GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                            GGML_ASSERT(false);
                        }
                }
@@ -1217,7 +1365,7 @@ void ggml_metal_graph_compute(

        MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
        if (status != MTLCommandBufferStatusCompleted) {
-            metal_printf("%s: command buffer %d failed with status %lu\n", __func__, i, status);
+            GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
            GGML_ASSERT(false);
        }
    }
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
--- a/ggml-opencl.cpp
+++ b/ggml-opencl.cpp
@@ -847,7 +847,7 @@ std::array<std::string, 2> mul_str_values = {
    "mul_f32", "float"
 };

-std::string& replace(std::string& s, const std::string& from, const std::string& to) {
+static std::string& replace(std::string& s, const std::string& from, const std::string& to) {
    size_t pos = 0;
    while ((pos = s.find(from, pos)) != std::string::npos) {
         s.replace(pos, from.length(), to);
@@ -856,7 +856,7 @@ std::string& replace(std::string& s, const std::string& from, const std::string&
    return s;
 }

-std::string generate_kernels() {
+static std::string generate_kernels() {
    std::stringstream src;
    src << program_source << '\n';
    src << k_quants_source << '\n';
@@ -1788,7 +1788,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
    return false;
 }

-bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
+static bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
    // If device doesn't support FP16
    if (!fp16_support) {
        return false;
--- a/ggml.c
+++ b/ggml.c
--- a/ggml.h
+++ b/ggml.h
@@ -195,6 +195,14 @@
 #    define GGML_DEPRECATED(func, hint) func
 #endif

+#ifndef __GNUC__
+#    define GGML_ATTRIBUTE_FORMAT(...)
+#elif defined(__MINGW32__)
+#    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+
 #include <stdint.h>
 #include <stddef.h>
 #include <stdbool.h>
@@ -206,8 +214,8 @@
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this

 #define GGML_MAX_DIMS          4
-#define GGML_MAX_NODES         4096
-#define GGML_MAX_PARAMS        256
+#define GGML_MAX_NODES         16384
+#define GGML_MAX_PARAMS        1024
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_SRC           6
 #define GGML_MAX_NAME          64
@@ -437,6 +445,12 @@ extern "C" {
        GGML_OBJECT_WORK_BUFFER
    };

+    enum ggml_log_level {
+        GGML_LOG_LEVEL_ERROR = 2,
+        GGML_LOG_LEVEL_WARN = 3,
+        GGML_LOG_LEVEL_INFO = 4
+    };
+
    // ggml object
    struct ggml_object {
        size_t offs;
@@ -512,7 +526,15 @@ extern "C" {
    // next prime after GGML_MAX_NODES
    // #define GGML_GRAPH_HASHTABLE_SIZE 4099
    // next prime after GGML_MAX_NODES * 2 (nodes + leafs)
-    #define GGML_GRAPH_HASHTABLE_SIZE 8273
+    // #define GGML_GRAPH_HASHTABLE_SIZE 8273
+    // #define GGML_GRAPH_HASHTABLE_SIZE 16411
+    #define GGML_GRAPH_HASHTABLE_SIZE 32771
+
+    enum ggml_cgraph_eval_order {
+        GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
+        GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
+        GGML_CGRAPH_EVAL_ORDER_COUNT
+    };

    // computation graph
    struct ggml_cgraph {
@@ -525,6 +547,8 @@ extern "C" {

        void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE];

+        enum ggml_cgraph_eval_order order;
+
        // performance
        int     perf_runs;
        int64_t perf_cycles;
@@ -672,12 +696,21 @@ extern "C" {
    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);

+    // Converts a flat index into coordinates
+    GGML_API void    ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
+
    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);

+    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);

+    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
    GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
    GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);

@@ -685,6 +718,7 @@ extern "C" {

    GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
    GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
+    GGML_ATTRIBUTE_FORMAT(2, 3)
    GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);

    //
@@ -710,6 +744,12 @@ extern "C" {
            struct ggml_tensor  * a,
            struct ggml_tensor  * b);

+    GGML_API struct ggml_tensor * ggml_add_cast(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            enum   ggml_type      type);
+
    GGML_API struct ggml_tensor * ggml_add1(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
@@ -819,6 +859,7 @@ extern "C" {
            struct ggml_tensor  * a,
            struct ggml_tensor  * b);

+    // sums repetitions in a into shape of b
    GGML_API struct ggml_tensor * ggml_repeat_back(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
@@ -1040,7 +1081,6 @@ extern "C" {
            size_t                nb1,
            size_t                offset);

-
    // a -> b, return view(b)
    GGML_API struct ggml_tensor * ggml_cpy(
            struct ggml_context * ctx,
@@ -1063,6 +1103,33 @@ extern "C" {
            struct ggml_context * ctx,
            struct ggml_tensor  * a);

+    // make contiguous, with new shape
+    GGML_API struct ggml_tensor * ggml_cont_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0);
+
+    GGML_API struct ggml_tensor * ggml_cont_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1);
+
+    GGML_API struct ggml_tensor * ggml_cont_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2);
+
+    GGML_API struct ggml_tensor * ggml_cont_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3);
+
    // return view(a), b specifies the new shape
    // TODO: when we start computing gradient, make a copy instead of view
    GGML_API struct ggml_tensor * ggml_reshape(
@@ -1210,14 +1277,15 @@ extern "C" {
            struct ggml_tensor  * b);

    // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements
+    // if mode & 1 == 1, skip n_past elements (DEPRECATED)
    // if mode & 2 == 1, GPT-NeoX style
    // if mode & 4 == 1, ChatGLM style
-    // TODO: avoid creating a new tensor every time
+    //
+    // b is an int32 vector with size a->ne[2], it contains the positions
    GGML_API struct ggml_tensor * ggml_rope(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
            int                   n_dims,
            int                   mode,
            int                   n_ctx);
@@ -1226,7 +1294,7 @@ extern "C" {
    GGML_API struct ggml_tensor * ggml_rope_inplace(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
            int                   n_dims,
            int                   mode,
            int                   n_ctx);
@@ -1235,7 +1303,7 @@ extern "C" {
    GGML_API struct ggml_tensor * ggml_rope_custom(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
            int                   n_dims,
            int                   mode,
            int                   n_ctx,
@@ -1246,7 +1314,7 @@ extern "C" {
    GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
            int                   n_dims,
            int                   mode,
            int                   n_ctx,
@@ -1257,7 +1325,7 @@ extern "C" {
    GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
            int                   n_dims,
            float                 base,
            bool                  down);
@@ -1267,7 +1335,7 @@ extern "C" {
    GGML_API struct ggml_tensor * ggml_rope_back(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
            int                   n_dims,
            int                   mode,
            int                   n_ctx,
@@ -1647,6 +1715,16 @@ extern "C" {
    // dump the graph into a file using the dot format
    GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);

+    // build gradient checkpointing backward graph gb for gf using provided checkpoints
+    // gb_tmp will contain original backward graph with rewritten backward process nodes,
+    // but without the second forward pass nodes.
+    GGML_API void ggml_build_backward_gradient_checkpointing(
+            struct ggml_context   * ctx,
+            struct ggml_cgraph    * gf,
+            struct ggml_cgraph    * gb,
+            struct ggml_cgraph    * gb_tmp,
+            struct ggml_tensor  * * checkpoints,
+            int                     n_checkpoints);
    //
    // optimization
    //
@@ -1681,7 +1759,8 @@ extern "C" {
        GGML_LINESEARCH_INVALID_PARAMETERS,
    };

-    typedef void (*ggml_opt_callback)(void * data, float * sched);
+    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);

    // optimization parameters
    //
@@ -1712,6 +1791,8 @@ extern "C" {
        bool print_forward_graph;
        bool print_backward_graph;

+        int n_gradient_accumulation;
+
        // ADAM parameters
        struct {
            int n_iter;
@@ -1757,6 +1838,7 @@ extern "C" {
        float loss_after;

        struct {
+            struct ggml_tensor * g;  // current gradient
            struct ggml_tensor * m;  // first moment
            struct ggml_tensor * v;  // second moment
            struct ggml_tensor * pf; // past function values
@@ -1866,39 +1948,39 @@ extern "C" {

    GGML_API const char * gguf_type_name(enum gguf_type type);

-    GGML_API int    gguf_get_version    (struct gguf_context * ctx);
-    GGML_API size_t gguf_get_alignment  (struct gguf_context * ctx);
-    GGML_API size_t gguf_get_data_offset(struct gguf_context * ctx);
-    GGML_API void * gguf_get_data       (struct gguf_context * ctx);
+    GGML_API int    gguf_get_version    (const struct gguf_context * ctx);
+    GGML_API size_t gguf_get_alignment  (const struct gguf_context * ctx);
+    GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx);
+    GGML_API void * gguf_get_data       (const struct gguf_context * ctx);

-    GGML_API int          gguf_get_n_kv(struct gguf_context * ctx);
-    GGML_API int          gguf_find_key(struct gguf_context * ctx, const char * key);
-    GGML_API const char * gguf_get_key (struct gguf_context * ctx, int i);
+    GGML_API int          gguf_get_n_kv(const struct gguf_context * ctx);
+    GGML_API int          gguf_find_key(const struct gguf_context * ctx, const char * key);
+    GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int key_id);

-    GGML_API enum gguf_type gguf_get_kv_type (struct gguf_context * ctx, int i);
-    GGML_API enum gguf_type gguf_get_arr_type(struct gguf_context * ctx, int i);
+    GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int key_id);
+    GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id);

-    // results are undefined if the wrong type is used for the key
-    GGML_API uint8_t      gguf_get_val_u8  (struct gguf_context * ctx, int i);
-    GGML_API int8_t       gguf_get_val_i8  (struct gguf_context * ctx, int i);
-    GGML_API uint16_t     gguf_get_val_u16 (struct gguf_context * ctx, int i);
-    GGML_API int16_t      gguf_get_val_i16 (struct gguf_context * ctx, int i);
-    GGML_API uint32_t     gguf_get_val_u32 (struct gguf_context * ctx, int i);
-    GGML_API int32_t      gguf_get_val_i32 (struct gguf_context * ctx, int i);
-    GGML_API float        gguf_get_val_f32 (struct gguf_context * ctx, int i);
-    GGML_API uint64_t     gguf_get_val_u64 (struct gguf_context * ctx, int i);
-    GGML_API int64_t      gguf_get_val_i64 (struct gguf_context * ctx, int i);
-    GGML_API double       gguf_get_val_f64 (struct gguf_context * ctx, int i);
-    GGML_API bool         gguf_get_val_bool(struct gguf_context * ctx, int i);
-    GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i);
-    GGML_API int          gguf_get_arr_n   (struct gguf_context * ctx, int i);
-    GGML_API const void * gguf_get_arr_data(struct gguf_context * ctx, int i);
-    GGML_API const char * gguf_get_arr_str (struct gguf_context * ctx, int key_id, int i);
+    // will abort if the wrong type is used for the key
+    GGML_API uint8_t      gguf_get_val_u8  (const struct gguf_context * ctx, int key_id);
+    GGML_API int8_t       gguf_get_val_i8  (const struct gguf_context * ctx, int key_id);
+    GGML_API uint16_t     gguf_get_val_u16 (const struct gguf_context * ctx, int key_id);
+    GGML_API int16_t      gguf_get_val_i16 (const struct gguf_context * ctx, int key_id);
+    GGML_API uint32_t     gguf_get_val_u32 (const struct gguf_context * ctx, int key_id);
+    GGML_API int32_t      gguf_get_val_i32 (const struct gguf_context * ctx, int key_id);
+    GGML_API float        gguf_get_val_f32 (const struct gguf_context * ctx, int key_id);
+    GGML_API uint64_t     gguf_get_val_u64 (const struct gguf_context * ctx, int key_id);
+    GGML_API int64_t      gguf_get_val_i64 (const struct gguf_context * ctx, int key_id);
+    GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
+    GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
+    GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
+    GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int key_id);
+    GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
+    GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);

-    GGML_API int    gguf_get_n_tensors    (struct gguf_context * ctx);
-    GGML_API int    gguf_find_tensor      (struct gguf_context * ctx, const char * name);
-    GGML_API size_t gguf_get_tensor_offset(struct gguf_context * ctx, int i);
-    GGML_API char * gguf_get_tensor_name  (struct gguf_context * ctx, int i);
+    GGML_API int    gguf_get_n_tensors    (const struct gguf_context * ctx);
+    GGML_API int    gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
+    GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
+    GGML_API char * gguf_get_tensor_name  (const struct gguf_context * ctx, int i);

    // overrides existing values or adds a new one
    GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
@@ -1943,11 +2025,11 @@ extern "C" {
    //

    // write the entire context to a binary file
-    GGML_API void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only_meta);
+    GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta);

    // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
-    GGML_API size_t gguf_get_meta_size(struct gguf_context * ctx);
-    GGML_API void   gguf_get_meta_data(struct gguf_context * ctx, void * data);
+    GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
+    GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);

    //
    // system info
@@ -1961,6 +2043,7 @@ extern "C" {
    GGML_API int ggml_cpu_has_fma        (void);
    GGML_API int ggml_cpu_has_neon       (void);
    GGML_API int ggml_cpu_has_arm_fma    (void);
+    GGML_API int ggml_cpu_has_metal      (void);
    GGML_API int ggml_cpu_has_f16c       (void);
    GGML_API int ggml_cpu_has_fp16_va    (void);
    GGML_API int ggml_cpu_has_wasm_simd  (void);
--- a/gguf-py/gguf/gguf.py
+++ b/gguf-py/gguf/gguf.py
@@ -32,7 +32,7 @@ KEY_GENERAL_URL                  = "general.url"
 KEY_GENERAL_DESCRIPTION          = "general.description"
 KEY_GENERAL_LICENSE              = "general.license"
 KEY_GENERAL_SOURCE_URL           = "general.source.url"
-KEY_GENERAL_SOURCE_HF_REPO       = "general.source.hugginface.repository"
+KEY_GENERAL_SOURCE_HF_REPO       = "general.source.huggingface.repository"
 KEY_GENERAL_FILE_TYPE            = "general.file_type"

 # LLM
@@ -77,12 +77,14 @@ KEY_TOKENIZER_RWKV       = "tokenizer.rwkv.world"


 class MODEL_ARCH(IntEnum):
-    LLAMA  : int = auto()
-    FALCON : int = auto()
-    GPT2   : int = auto()
-    GPTJ   : int = auto()
-    GPTNEOX: int = auto()
-    MPT    : int = auto()
+    LLAMA         : int = auto()
+    FALCON        : int = auto()
+    BAICHUAN      : int = auto()
+    GPT2          : int = auto()
+    GPTJ          : int = auto()
+    GPTNEOX       : int = auto()
+    MPT           : int = auto()
+    STARCODER     : int = auto()


 class MODEL_TENSOR(IntEnum):
@@ -106,12 +108,14 @@ class MODEL_TENSOR(IntEnum):


 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
-    MODEL_ARCH.LLAMA:   "llama",
-    MODEL_ARCH.FALCON:  "falcon",
-    MODEL_ARCH.GPT2:    "gpt2",
-    MODEL_ARCH.GPTJ:    "gptj",
-    MODEL_ARCH.GPTNEOX: "gptneox",
-    MODEL_ARCH.MPT:     "mpt",
+    MODEL_ARCH.LLAMA:          "llama",
+    MODEL_ARCH.FALCON:         "falcon",
+    MODEL_ARCH.BAICHUAN:       "baichuan",
+    MODEL_ARCH.GPT2:           "gpt2",
+    MODEL_ARCH.GPTJ:           "gptj",
+    MODEL_ARCH.GPTNEOX:        "gptneox",
+    MODEL_ARCH.MPT:            "mpt",
+    MODEL_ARCH.STARCODER:      "starcoder",
 }

 MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = {
@@ -153,6 +157,34 @@ MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = {
        MODEL_TENSOR.FFN_DOWN:    "blk.{bid}.ffn_down",
        MODEL_TENSOR.FFN_UP:      "blk.{bid}.ffn_up",
    },
+    MODEL_ARCH.BAICHUAN: {
+        MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
+        MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
+        MODEL_TENSOR.OUTPUT:        "output",
+        MODEL_TENSOR.ROPE_FREQS:    "rope_freqs",
+        MODEL_TENSOR.ATTN_NORM:     "blk.{bid}.attn_norm",
+        MODEL_TENSOR.ATTN_Q:        "blk.{bid}.attn_q",
+        MODEL_TENSOR.ATTN_K:        "blk.{bid}.attn_k",
+        MODEL_TENSOR.ATTN_V:        "blk.{bid}.attn_v",
+        MODEL_TENSOR.ATTN_OUT:      "blk.{bid}.attn_output",
+        MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
+        MODEL_TENSOR.FFN_NORM:      "blk.{bid}.ffn_norm",
+        MODEL_TENSOR.FFN_GATE:      "blk.{bid}.ffn_gate",
+        MODEL_TENSOR.FFN_DOWN:      "blk.{bid}.ffn_down",
+        MODEL_TENSOR.FFN_UP:        "blk.{bid}.ffn_up",
+    },
+    MODEL_ARCH.STARCODER: {
+        MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
+        MODEL_TENSOR.POS_EMBD:      "position_embd",
+        MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
+        MODEL_TENSOR.OUTPUT:        "output",
+        MODEL_TENSOR.ATTN_NORM:     "blk.{bid}.attn_norm",
+        MODEL_TENSOR.ATTN_QKV:      "blk.{bid}.attn_qkv",
+        MODEL_TENSOR.ATTN_OUT:      "blk.{bid}.attn_output",
+        MODEL_TENSOR.FFN_NORM:      "blk.{bid}.ffn_norm",
+        MODEL_TENSOR.FFN_DOWN:      "blk.{bid}.ffn_down",
+        MODEL_TENSOR.FFN_UP:        "blk.{bid}.ffn_up",
+    },
    MODEL_ARCH.GPT2: {
        # TODO
    },
@@ -165,6 +197,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_ROT_EMBD,
    ],
+    MODEL_ARCH.BAICHUAN: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }


@@ -187,7 +223,7 @@ class TensorNameMap:
        # Output
        MODEL_TENSOR.OUTPUT: (
            "embed_out", # gptneox
-            "lm_head",   # gpt2 mpt falcon llama-hf
+            "lm_head",   # gpt2 mpt falcon llama-hf baichuan
            "output",    # llama-pth
        ),

@@ -195,7 +231,7 @@ class TensorNameMap:
        MODEL_TENSOR.OUTPUT_NORM: (
            "gpt_neox.final_layer_norm", # gptneox
            "transformer.ln_f",          # gpt2 falcon
-            "model.norm",                # llama-hf
+            "model.norm",                # llama-hf baichuan
            "norm",                      # llama-pth
        ),

@@ -311,6 +347,7 @@ class TensorNameMap:
            tensor_name = tensor_names.get(tensor)
            if tensor_name is None:
                continue
+            mapping[tensor_name] = (tensor, tensor_name)
            for key in keys:
                mapping[key] = (tensor, tensor_name)
        for bid in range(n_blocks):
@@ -319,11 +356,12 @@ class TensorNameMap:
                if tensor_name is None:
                    continue
                tensor_name = tensor_name.format(bid = bid)
+                mapping[tensor_name] = (tensor, tensor_name)
                for key in keys:
                    key = key.format(bid = bid)
                    mapping[key] = (tensor, tensor_name)

-    def get_type_and_name(self, key: str, try_suffixes: Sequence[str]) -> tuple[MODEL_TENSOR, str] | None:
+    def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
        result = self.mapping.get(key)
        if result is not None:
            return result
@@ -334,13 +372,13 @@ class TensorNameMap:
                    return (result[0], result[1] + suffix)
        return None

-    def get_name(self, key: str, try_suffixes: Sequence[str]) -> str | None:
+    def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None:
        result = self.get_type_and_name(key, try_suffixes = try_suffixes)
        if result is None:
            return None
        return result[1]

-    def get_type(self, key: str, try_suffixes: Sequence[str]) -> MODEL_TENSOR | None:
+    def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None:
        result = self.get_type_and_name(key, try_suffixes = try_suffixes)
        if result is None:
            return None
--- a/gguf-py/pyproject.toml
+++ b/gguf-py/pyproject.toml
@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.3.2"
+version = "0.3.3"
 description = "Write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [
--- a/k_quants.c
+++ b/k_quants.c
@@ -2609,7 +2609,10 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri

        memcpy(utmp, x[i].scales, 12);

-        const uint32x2_t mins8 = {utmp[1] & kmask1, ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4)};
+        uint32x2_t mins8 = { 0 };
+        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
+        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
+
        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[0] &= kmask1;

--- a/llama.cpp
+++ b/llama.cpp
--- a/llama.h
+++ b/llama.h
@@ -37,6 +37,8 @@

 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF

+#define LLAMA_MAX_RNG_STATE (64*1024)
+
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'

 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
@@ -60,13 +62,9 @@ extern "C" {
    struct llama_model;
    struct llama_context;

-    typedef int llama_token;
-
-    enum llama_log_level {
-        LLAMA_LOG_LEVEL_ERROR = 2,
-        LLAMA_LOG_LEVEL_WARN  = 3,
-        LLAMA_LOG_LEVEL_INFO  = 4
-    };
+    typedef int32_t llama_pos;
+    typedef int32_t llama_token;
+    typedef int32_t llama_seq_id;

    enum llama_vocab_type {
        LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
@@ -86,24 +84,24 @@ extern "C" {
    // model file types
    enum llama_ftype {
        LLAMA_FTYPE_ALL_F32              = 0,
-        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_F16           = 1,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4,  // tok_embeddings.weight and output.weight are F16
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5,  // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6,  // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors

        LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
    };
@@ -122,6 +120,35 @@ extern "C" {

    typedef void (*llama_progress_callback)(float progress, void *ctx);

+    // Input data for llama_decode
+    // A llama_batch object can contain input about one or many sequences
+    // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
+    //
+    // - token  : the token ids of the input (used when embd is NULL)
+    // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
+    // - pos    : the positions of the respective token in the sequence
+    // - seq_id : the sequence to which the respective token belongs
+    // - logits : if zero, the logits for the respective token will not be output
+    //
+    typedef struct llama_batch {
+        int32_t n_tokens;
+
+        llama_token  * token;
+        float        * embd;
+        llama_pos    * pos;
+        llama_seq_id * seq_id;
+        int8_t       * logits;
+
+        // NOTE: helpers for smooth API transition - can be deprecated in the future
+        //       for future-proof code, use the above fields instead and ignore everything below
+        //
+        // pos[i] = all_pos_0 + i*all_pos_1
+        //
+        llama_pos    all_pos_0;  // used if pos == NULL
+        llama_pos    all_pos_1;  // used if pos == NULL
+        llama_seq_id all_seq_id; // used if seq_id == NULL
+    } llama_batch;
+
    struct llama_context_params {
        uint32_t seed;         // RNG seed, -1 for random
        int32_t  n_ctx;        // text context
@@ -151,13 +178,6 @@ extern "C" {
        bool embedding;  // embedding mode only
    };

-    // Signature for logging events
-    // Note that text includes the new line character at the end for most events.
-    // If your logging mechanism cannot handle that, check if the last character is '\n' and strip it
-    // if it exists.
-    // It might not exist for progress report where '.' is output repeatedly.
-    typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data);
-
    // model quantization parameters
    typedef struct llama_model_quantize_params {
        int nthread;                 // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
@@ -215,6 +235,7 @@ extern "C" {
        int32_t n_eval;
    };

+    // Helpers for getting default parameters
    LLAMA_API struct llama_context_params llama_context_default_params(void);
    LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void);

@@ -259,11 +280,16 @@ extern "C" {

    // Get a string describing the model type
    LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+
    // Returns the total size of all the tensors in the model in bytes
    LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
+
    // Returns the total number of parameters in the model
    LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);

+    // Get a llama model tensor
+    LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
+
    // Returns 0 on success
    LLAMA_API int llama_model_quantize(
            const char * fname_inp,
@@ -279,21 +305,65 @@ extern "C" {
    LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
            struct llama_context * ctx,
                      const char * path_lora,
+                           float   scale,
                      const char * path_base_model,
                             int   n_threads),
-            "please use llama_model_apply_lora_from_file instead");
+            "use llama_model_apply_lora_from_file instead");

    LLAMA_API int llama_model_apply_lora_from_file(
            const struct llama_model * model,
-                          const char * path_lora,
-                          const char * path_base_model,
-                                 int   n_threads);
+                      const char * path_lora,
+                           float   scale,
+                      const char * path_base_model,
+                             int   n_threads);
+
+    //
+    // KV cache
+    //

    // Returns the number of tokens in the KV cache
-    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
+    LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
+            "avoid using this, it will be removed in the future, instead - count the tokens in user code");

-    // Sets the current rng seed.
-    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+    // Remove all tokens data of cells in [c0, c1)
+    LLAMA_API void llama_kv_cache_tokens_rm(
+            struct llama_context * ctx,
+                         int32_t   c0,
+                         int32_t   c1);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    LLAMA_API void llama_kv_cache_seq_rm(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
+    LLAMA_API void llama_kv_cache_seq_cp(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id_src,
+                    llama_seq_id   seq_id_dst,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_kv_cache_seq_keep(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    LLAMA_API void llama_kv_cache_seq_shift(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                       llama_pos   delta);
+
+    //
+    // State / sessions
+    //

    // Returns the maximum size in bytes of the state (rng, logits, embedding
    // and kv_cache) - will often be smaller after compacting tokens
@@ -302,48 +372,100 @@ extern "C" {
    // Copies the state to the specified destination address.
    // Destination needs to have allocated enough memory.
    // Returns the number of bytes copied
-    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
+    LLAMA_API size_t llama_copy_state_data(
+            struct llama_context * ctx,
+                         uint8_t * dst);

    // Set the state reading from the specified address
    // Returns the number of bytes read
-    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
+    LLAMA_API size_t llama_set_state_data(
+            struct llama_context * ctx,
+                         uint8_t * src);

    // Save/load session file
-    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
-    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+    LLAMA_API bool llama_load_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+                     llama_token * tokens_out,
+                          size_t   n_token_capacity,
+                          size_t * n_token_count_out);

-    // Run the llama inference to obtain the logits and probabilities for the next token.
+    LLAMA_API bool llama_save_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+               const llama_token * tokens,
+                          size_t   n_token_count);
+
+    //
+    // Decoding
+    //
+
+    // Run the llama inference to obtain the logits and probabilities for the next token(s).
    // tokens + n_tokens is the provided batch of new tokens to process
    // n_past is the number of tokens to use from previous eval calls
    // Returns 0 on success
-    LLAMA_API int llama_eval(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval(
            struct llama_context * ctx,
-               const llama_token * tokens,
-                             int   n_tokens,
+                     llama_token * tokens,
+                         int32_t   n_tokens,
                             int   n_past,
-                             int   n_threads);
+                             int   n_threads),
+            "use llama_decode() instead");

    // Same as llama_eval, but use float matrix input directly.
-    LLAMA_API int llama_eval_embd(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval_embd(
            struct llama_context * ctx,
-                     const float * embd,
-                             int   n_tokens,
+                           float * embd,
+                         int32_t   n_tokens,
                             int   n_past,
-                             int   n_threads);
+                             int   n_threads),
+            "use llama_decode() instead");

-    // Export a static computation graph for context of 511 and batch size of 1
-    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
-    //       parameters here to keep things simple
-    // IMPORTANT: do not use for anything else other than debugging and testing!
-    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+    // Return batch for single sequence of tokens starting at pos_0
+    //
+    // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
+    //
+    LLAMA_API struct llama_batch llama_batch_get_one(
+                  llama_token * tokens,
+                      int32_t   n_tokens,
+                    llama_pos   pos_0,
+                 llama_seq_id   seq_id);
+
+    // Allocates a batch of tokens on the heap
+    // The batch has to be freed with llama_batch_free()
+    // If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float)
+    // Otherwise, llama_batch.token will be allocated to store n_tokens llama_token
+    // The rest of the llama_batch members are allocated with size n_tokens
+    // All members are left uninitialized
+    LLAMA_API struct llama_batch llama_batch_init(
+            int32_t n_tokens,
+            int32_t embd);
+
+    // Frees a batch of tokens allocated with llama_batch_init()
+    LLAMA_API void llama_batch_free(struct llama_batch batch);
+
+    // Positive return values does not mean a fatal error, but rather a warning.
+    //   0 - success
+    //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+    // < 0 - error
+    LLAMA_API int llama_decode(
+            struct llama_context * ctx,
+              struct llama_batch   batch,
+                             int   n_threads);

    // Token logits obtained from the last call to llama_eval()
    // The logits for the last token are stored in the last row
-    // Can be mutated in order to change the probabilities of the next token
-    // Rows: n_tokens
+    // Logits for which llama_batch.logits[i] == 0 are undefined
+    // Rows: n_tokens provided with llama_batch
    // Cols: n_vocab
    LLAMA_API float * llama_get_logits(struct llama_context * ctx);

+    // Logits for the ith token. Equivalent to:
+    // llama_get_logits(ctx) + i*n_vocab
+    LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
+
    // Get the embeddings for the input
    // shape: [n_embd] (1-dimensional)
    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
@@ -374,6 +496,7 @@ extern "C" {
    LLAMA_API int llama_tokenize(
            struct llama_context * ctx,
                      const char * text,
+                             int   text_len,
                     llama_token * tokens,
                             int   n_max_tokens,
                            bool   add_bos);
@@ -381,6 +504,7 @@ extern "C" {
    LLAMA_API int llama_tokenize_with_model(
        const struct llama_model * model,
                      const char * text,
+                             int   text_len,
                     llama_token * tokens,
                             int   n_max_tokens,
                            bool   add_bos);
@@ -418,11 +542,25 @@ extern "C" {
    // Sampling functions
    //

+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+
    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
-    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
+    LLAMA_API void llama_sample_repetition_penalty(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   last_tokens_size,
+                          float    penalty);

    /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
-    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+    LLAMA_API void llama_sample_frequency_and_presence_penalties(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   last_tokens_size,
+                           float   alpha_frequency,
+                           float   alpha_presence);

    /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
@@ -435,23 +573,54 @@ extern "C" {
                             float   scale);

    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API void llama_sample_softmax(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);

    /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
+    LLAMA_API void llama_sample_top_k(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                             int   k,
+                          size_t   min_keep);

    /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+    LLAMA_API void llama_sample_top_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);

    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
+    LLAMA_API void llama_sample_tail_free(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   z,
+                          size_t   min_keep);

    /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
-    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
-    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+    LLAMA_API void llama_sample_typical(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
+    LLAMA_API void llama_sample_temp(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   temp);
+
+    LLAMA_API DEPRECATED(void llama_sample_temperature(
+                struct llama_context * ctx,
+              llama_token_data_array * candidates,
+                               float   temp),
+            "use llama_sample_temp instead");

    /// @details Apply constraints from grammar
-    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+    LLAMA_API void llama_sample_grammar(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+      const struct llama_grammar * grammar);

    /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
@@ -459,23 +628,41 @@ extern "C" {
    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
    /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                             int   m,
+                           float * mu);

    /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                           float * mu);

    /// @details Selects the token with the highest probability.
-    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API llama_token llama_sample_token_greedy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);

    /// @details Randomly selects a token from the candidates based on their probabilities.
-    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API llama_token llama_sample_token(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);

    /// @details Accepts the sampled token into the grammar
-    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+    LLAMA_API void llama_grammar_accept_token(
+            struct llama_context * ctx,
+            struct llama_grammar * grammar,
+                     llama_token   token);

    //
    // Beam search
@@ -483,9 +670,10 @@ extern "C" {

    struct llama_beam_view {
        const llama_token * tokens;
+
        size_t n_tokens;
-        float p;   // Cumulative beam probability (renormalized relative to all beams)
-        bool eob;  // Callback should set this to true when a beam is at end-of-beam.
+        float  p;        // Cumulative beam probability (renormalized relative to all beams)
+        bool   eob;      // Callback should set this to true when a beam is at end-of-beam.
    };

    // Passed to beam_search_callback function.
@@ -494,9 +682,10 @@ extern "C" {
    // These pointers are valid only during the synchronous callback, so should not be saved.
    struct llama_beams_state {
        struct llama_beam_view * beam_views;
+
        size_t n_beams;               // Number of elements in beam_views[].
        size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
-        bool last_call;               // True iff this is the last callback invocation.
+        bool   last_call;             // True iff this is the last callback invocation.
    };

    // Type of pointer to the beam_search_callback function.
@@ -512,10 +701,18 @@ extern "C" {
    /// @param n_past Number of tokens already evaluated.
    /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
    /// @param n_threads Number of threads as passed to llama_eval().
-    LLAMA_API void llama_beam_search(struct llama_context * ctx, llama_beam_search_callback_fn_t callback, void * callback_data, size_t n_beams, int n_past, int n_predict, int n_threads);
+    LLAMA_API void llama_beam_search(
+                   struct llama_context * ctx,
+        llama_beam_search_callback_fn_t   callback,
+                                   void * callback_data,
+                                 size_t   n_beams,
+                                    int   n_past,
+                                    int   n_predict,
+                                    int   n_threads);

    // Performance information
    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
+
    LLAMA_API void llama_print_timings(struct llama_context * ctx);
    LLAMA_API void llama_reset_timings(struct llama_context * ctx);

@@ -524,7 +721,7 @@ extern "C" {

    // Set callback for all future logging events.
    // If this is not called, or NULL is supplied, everything is output on stderr.
-    LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data);
+    LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data);

    LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);

@@ -540,7 +737,9 @@ extern "C" {

 struct ggml_tensor;

-const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+);

 #endif // LLAMA_API_INTERNAL

--- a/pocs/vdot/vdot.cpp
+++ b/pocs/vdot/vdot.cpp
@@ -16,7 +16,7 @@

 constexpr int kVecSize = 1 << 18;

-float drawFromGaussianPdf(std::mt19937& rndm) {
+static float drawFromGaussianPdf(std::mt19937& rndm) {
    constexpr double kScale = 1./(1. + std::mt19937::max());
    constexpr double kTwoPiTimesScale = 6.28318530717958647692*kScale;
    static float lastX;
@@ -28,7 +28,8 @@ float drawFromGaussianPdf(std::mt19937& rndm) {
    haveX = true;
    return r*cos(phi);
 }
-void fillRandomGaussianFloats(std::vector<float>& values, std::mt19937& rndm, float mean = 0) {
+
+static void fillRandomGaussianFloats(std::vector<float>& values, std::mt19937& rndm, float mean = 0) {
    for (auto& v : values) v = mean + drawFromGaussianPdf(rndm);
 }

--- a/prompts/chat-with-baichuan.txt
+++ b/prompts/chat-with-baichuan.txt
@@ -0,0 +1,4 @@
+以下内容为人类用户与与一位智能助手的对话。
+
+用户:你好！
+助手:
--- a/run_with_preset.py
+++ b/run_with_preset.py
@@ -13,7 +13,7 @@ CLI_ARGS_MAIN_PERPLEXITY = [
    "hellaswag-tasks", "ignore-eos", "in-prefix", "in-prefix-bos", "in-suffix", "instruct",
    "interactive", "interactive-first", "keep", "logdir", "logit-bias", "lora", "lora-base",
    "low-vram", "main-gpu", "memory-f32", "mirostat", "mirostat-ent", "mirostat-lr", "mlock",
-    "model", "mtest", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q",
+    "model", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q",
    "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt",
    "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "random-prompt", "repeat-last-n",
    "repeat-penalty", "reverse-prompt", "rope-freq-base", "rope-freq-scale", "rope-scale", "seed",
--- a/scripts/LlamaConfig.cmake.in
+++ b/scripts/LlamaConfig.cmake.in
@@ -0,0 +1,69 @@
+set(LLAMA_VERSION @LLAMA_INSTALL_VERSION@)
+set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
+set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
+set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
+set(LLAMA_BLAS @LLAMA_BLAS@)
+set(LLAMA_CUBLAS @LLAMA_CUBLAS@)
+set(LLAMA_METAL @LLAMA_METAL@)
+set(LLAMA_MPI @LLAMA_MPI@)
+set(LLAMA_CLBLAST @LLAMA_CLBLAST@)
+set(LLAMA_HIPBLAS @LLAMA_HIPBLAS@)
+set(LLAMA_ACCELERATE @LLAMA_ACCELERATE@)
+
+@PACKAGE_INIT@
+
+set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
+set_and_check(LLAMA_LIB_DIR "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
+set_and_check(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
+
+# Ensure transient dependencies satisfied
+
+find_package(Threads REQUIRED)
+if (APPLE AND LLAMA_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+endif()
+
+if (LLAMA_BLAS)
+    find_package(BLAS REQUIRED)
+endif()
+
+if (LLAMA_CUBLAS)
+    find_package(CUDAToolkit REQUIRED)
+endif()
+
+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK Metal REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+endif()
+
+if (LLAMA_MPI)
+    find_package(MPI REQUIRED)
+endif()
+
+if (LLAMA_CLBLAST)
+    find_package(CLBlast REQUIRED)
+endif()
+
+if (LLAMA_HIPBLAS)
+    find_package(hip REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+endif()
+
+find_library(llama_LIBRARY llama
+    REQUIRED
+    HINTS ${LLAMA_LIB_DIR})
+
+set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
+add_library(llama UNKNOWN IMPORTED)
+set_target_properties(llama
+    PROPERTIES
+        INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
+        INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
+        IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+        IMPORTED_LOCATION "${llama_LIBRARY}"
+        INTERFACE_COMPILE_FEATURES cxx_std_11
+        POSITION_INDEPENDENT_CODE ON )
+
+check_required_components(Llama)
--- a/scripts/build-info.cmake
+++ b/scripts/build-info.cmake
@@ -2,20 +2,18 @@ set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.h.in")
 set(HEADER_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
 set(BUILD_NUMBER 0)
 set(BUILD_COMMIT "unknown")
+set(BUILD_COMPILER "unknown")
+set(BUILD_TARGET "unknown")

 # Look for git
 find_package(Git)
 if(NOT Git_FOUND)
-    execute_process(
-        COMMAND which git
-        OUTPUT_VARIABLE GIT_EXECUTABLE
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-    if(NOT GIT_EXECUTABLE STREQUAL "")
+    find_program(GIT_EXECUTABLE NAMES git git.exe)
+    if(GIT_EXECUTABLE)
        set(Git_FOUND TRUE)
-        message(STATUS "Found Git using 'which': ${GIT_EXECUTABLE}")
+        message(STATUS "Found Git: ${GIT_EXECUTABLE}")
    else()
-        message(WARNING "Git not found using 'find_package' or 'which'. Build info will not be accurate. Consider installing Git or ensuring it is in the PATH.")
+        message(WARNING "Git not found. Build info will not be accurate.")
    endif()
 endif()

@@ -26,26 +24,49 @@ if(Git_FOUND)
        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
        OUTPUT_VARIABLE HEAD
        OUTPUT_STRIP_TRAILING_WHITESPACE
-        RESULT_VARIABLE GIT_HEAD_RESULT
    )
    execute_process(
        COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD
        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
        OUTPUT_VARIABLE COUNT
        OUTPUT_STRIP_TRAILING_WHITESPACE
-        RESULT_VARIABLE GIT_COUNT_RESULT
    )
-    if(GIT_HEAD_RESULT EQUAL 0 AND GIT_COUNT_RESULT EQUAL 0)
-        set(BUILD_COMMIT ${HEAD})
-        set(BUILD_NUMBER ${COUNT})
-    endif()
+    set(BUILD_COMMIT ${HEAD})
+    set(BUILD_NUMBER ${COUNT})
+endif()
+
+if(MSVC)
+    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+else()
+    execute_process(
+        COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_COMPILER ${OUT})
+    execute_process(
+        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_TARGET ${OUT})
 endif()

 # Only write the header if it's changed to prevent unnecessary recompilation
 if(EXISTS ${HEADER_FILE})
-    file(STRINGS ${HEADER_FILE} CONTENTS REGEX "BUILD_COMMIT \"([^\"]*)\"")
-    list(GET CONTENTS 0 EXISTING)
-    if(NOT EXISTING STREQUAL "#define BUILD_COMMIT \"${BUILD_COMMIT}\"")
+    file(READ ${HEADER_FILE} CONTENTS)
+    string(REGEX MATCH "BUILD_COMMIT \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_COMMIT ${CMAKE_MATCH_1})
+    string(REGEX MATCH "BUILD_COMPILER \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_COMPILER ${CMAKE_MATCH_1})
+    string(REGEX MATCH "BUILD_TARGET \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_TARGET ${CMAKE_MATCH_1})
+    if (
+        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
+        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
+        NOT OLD_TARGET   STREQUAL BUILD_TARGET
+    )
        configure_file(${TEMPLATE_FILE} ${HEADER_FILE})
    endif()
 else()
--- a/scripts/build-info.h.in
+++ b/scripts/build-info.h.in
@@ -3,5 +3,7 @@

 #define BUILD_NUMBER @BUILD_NUMBER@
 #define BUILD_COMMIT "@BUILD_COMMIT@"
+#define BUILD_COMPILER "@BUILD_COMPILER@"
+#define BUILD_TARGET "@BUILD_TARGET@"

 #endif // BUILD_INFO_H
--- a/scripts/build-info.sh
+++ b/scripts/build-info.sh
@@ -1,23 +1,35 @@
 #!/bin/sh

-BUILD_NUMBER="0"
-BUILD_COMMIT="unknown"
+CC=$1

-REV_LIST=$(git rev-list --count HEAD)
-if [ $? -eq 0 ]; then
-  BUILD_NUMBER=$REV_LIST
+build_number="0"
+build_commit="unknown"
+build_compiler="unknown"
+build_target="unknown"
+
+if out=$(git rev-list --count HEAD); then
+  # git is broken on WSL so we need to strip extra newlines
+  build_number=$(printf '%s' "$out" | tr -d '\n')
 fi

-REV_PARSE=$(git rev-parse --short HEAD)
-if [ $? -eq 0 ]; then
-  BUILD_COMMIT=$REV_PARSE
+if out=$(git rev-parse --short HEAD); then
+  build_commit=$(printf '%s' "$out" | tr -d '\n')
+fi
+
+if out=$($CC --version | head -1); then
+  build_compiler=$out
+fi
+
+if out=$($CC -dumpmachine); then
+  build_target=$out
 fi

 echo "#ifndef BUILD_INFO_H"
 echo "#define BUILD_INFO_H"
-echo ""
-echo "#define BUILD_NUMBER $BUILD_NUMBER" | tr -d '\n'
-echo ""
-echo "#define BUILD_COMMIT \"$BUILD_COMMIT\"" | tr -d '\n'
-echo ""
+echo
+echo "#define BUILD_NUMBER $build_number"
+echo "#define BUILD_COMMIT \"$build_commit\""
+echo "#define BUILD_COMPILER \"$build_compiler\""
+echo "#define BUILD_TARGET \"$build_target\""
+echo
 echo "#endif // BUILD_INFO_H"
--- a/scripts/verify-checksum-models.py
+++ b/scripts/verify-checksum-models.py
@@ -1,4 +1,4 @@
-#!/bin/env python3
+#!/usr/bin/env python3

 import os
 import hashlib
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -29,15 +29,16 @@ llama_build_executable(test-tokenizer-0-llama.cpp)
 llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
 llama_build_executable(test-tokenizer-0-falcon.cpp)
 #llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-llama_build_executable(test-tokenizer-1.cpp)
-# test-tokenizer-1 requires a BPE vocab. re-enable when we have one.
-#llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_build_executable(test-tokenizer-1-llama.cpp)
+llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
 #llama_test_executable(test-tokenizer-1.aquila test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp) # SLOW
 # llama_build_and_test_executable(test-opt.cpp) # SLOW

+llama_build_and_test_executable(test-rope.cpp)
+
 # dummy executable - not installed
 get_filename_component(TEST_TARGET test-c.c NAME_WE)
 add_executable(${TEST_TARGET} test-c.c)
--- a/tests/test-grad0.cpp
+++ b/tests/test-grad0.cpp
@@ -251,18 +251,20 @@ static bool check_gradient(
        printf("GGML_N_THREADS = %d\n", n_threads);
    }

-    struct ggml_cgraph gf = ggml_build_forward (f);
-    struct ggml_cgraph gb = ggml_build_backward(ctx0, &gf, false);
+    struct ggml_cgraph * gf = ggml_build_forward_ctx(ctx0, f);
+    struct ggml_cgraph * gb = ggml_new_graph(ctx0);
+    *gb = *gf;
+    ggml_build_backward_expand(ctx0, gf, gb, false);

-    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+    ggml_graph_compute_with_ctx(ctx0, gf, n_threads);

-    ggml_graph_reset  (&gf);
+    ggml_graph_reset  (gf);
    ggml_set_f32      (f->grad, 1.0f);

-    ggml_graph_compute_with_ctx(ctx0, &gb, n_threads);
+    ggml_graph_compute_with_ctx(ctx0, gb, n_threads);

-    // ggml_graph_dump_dot(&gf, NULL, "test-grad0-forward.dot");
-    // ggml_graph_dump_dot(&gb, &gf,  "test-grad0-backward.dot");
+    // ggml_graph_dump_dot(gf, NULL, "test-grad0-forward.dot");
+    // ggml_graph_dump_dot(gb, gf,  "test-grad0-backward.dot");

    for (int i = 0; i < nargs; ++i) {
        const int nelements = ggml_nelements(x[i]);
@@ -273,13 +275,13 @@ static bool check_gradient(
            const float xp = x0 + eps;
            ggml_set_f32_1d(x[i], k, xp);

-            ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+            ggml_graph_compute_with_ctx(ctx0, gf, n_threads);

            const double f0 = ggml_get_f32_1d(f, 0);

            ggml_set_f32_1d(x[i], k, xm);

-            ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+            ggml_graph_compute_with_ctx(ctx0, gf, n_threads);

            const double f1 = ggml_get_f32_1d(f, 0);
            const double g0 = (f0 - f1)/(2.0*(double) eps);
@@ -287,10 +289,10 @@ static bool check_gradient(
            ggml_set_f32_1d(x[i], k, x0);

            // compute gradient using backward graph
-            ggml_graph_reset  (&gf);
+            ggml_graph_reset  (gf);
            ggml_set_f32      (f->grad, 1.0f);

-            ggml_graph_compute_with_ctx(ctx0, &gb, n_threads);
+            ggml_graph_compute_with_ctx(ctx0, gb, n_threads);

            const double g1 = ggml_get_f32_1d(x[i]->grad, k);

@@ -373,7 +375,7 @@ static bool check_mat_mul(

 int main(int argc, const char ** argv) {
    struct ggml_init_params params = {
-        /* .mem_size   = */ 128*1024*1024,
+        /* .mem_size   = */ 256*1024*1024,
        /* .mem_buffer = */ NULL,
        /* .no_alloc   = */ false,
    };
@@ -405,6 +407,7 @@ int main(int argc, const char ** argv) {
        }
    }

+    unsigned seed_iter = 1;

    // original loop: 1000
    int niter = 4;
@@ -416,6 +419,10 @@ int main(int argc, const char ** argv) {
        niter = atoi(argv[1]);
    }
    for (int iter = 0; iter < niter; ++iter) {
+        srand(seed_iter);
+        seed_iter = rand();
+        unsigned seed = rand();
+
        printf("test-grad0: iter:%d/%d\n", iter, niter);
        struct ggml_context * ctx0 = ggml_init(params);

@@ -425,6 +432,7 @@ int main(int argc, const char ** argv) {

        // add f32
        {
+            srand(seed);
            const int nargs = 2;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -441,6 +449,7 @@ int main(int argc, const char ** argv) {

        // add f16
        {
+            srand(seed);
            const int nargs = 2;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -457,6 +466,7 @@ int main(int argc, const char ** argv) {

        // sub
        {
+            srand(seed);
            const int nargs = 2;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -473,6 +483,7 @@ int main(int argc, const char ** argv) {

        // mul
        {
+            srand(seed);
            const int nargs = 2;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -489,6 +500,7 @@ int main(int argc, const char ** argv) {

        // div
        {
+            srand(seed);
            const int nargs = 2;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -505,6 +517,7 @@ int main(int argc, const char ** argv) {

        // sqr
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -521,6 +534,7 @@ int main(int argc, const char ** argv) {

        // sqrt
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -537,6 +551,7 @@ int main(int argc, const char ** argv) {

        // log
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -553,6 +568,7 @@ int main(int argc, const char ** argv) {

        // sum
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -570,6 +586,7 @@ int main(int argc, const char ** argv) {

        // sum_rows
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -587,6 +604,7 @@ int main(int argc, const char ** argv) {
        // mean, not yet fully implemented
        if(0)
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -604,6 +622,7 @@ int main(int argc, const char ** argv) {
        // argmax
        if (0)
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -620,6 +639,7 @@ int main(int argc, const char ** argv) {

        // repeat
        {
+            srand(seed);
            int64_t ne2[4];
            get_random_dims(ne2, 4);

@@ -642,6 +662,7 @@ int main(int argc, const char ** argv) {

        // repeat back
        {
+            srand(seed);
            int64_t ne2[4];
            get_random_dims(ne2, 4);

@@ -680,6 +701,7 @@ int main(int argc, const char ** argv) {

        // sgn
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -696,6 +718,7 @@ int main(int argc, const char ** argv) {

        // neg
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -712,6 +735,7 @@ int main(int argc, const char ** argv) {

        // step
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -729,6 +753,7 @@ int main(int argc, const char ** argv) {
        // tanh, not yet fully implemented
        if(0)
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -745,33 +770,45 @@ int main(int argc, const char ** argv) {

        // mul_mat
        {
+            srand(seed);
            const int nargs = 2;

-            for (int ndims = 2; ndims <= 2; ++ndims) {
+            for (int ndims = 2; ndims <= 4; ++ndims) {
+                int max_nrep = (ndims >= 3) ? 2 : 1;
                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                {
-                    int64_t ne2[4];
-                    get_random_dims(ne2, 4);
-                    ne2[0] = ne[0];
-                    x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+                for (int nrep2 = 1; nrep2 < max_nrep; ++nrep2) {
+                    for (int nrep3 = 1; nrep3 < max_nrep; ++nrep3) {
+                        {
+                            int64_t ne2[4];
+                            get_random_dims(ne2, 4);
+                            ne2[0] = ne[0];
+                            ne2[2] = nrep2 * ne[2];
+                            ne2[3] = nrep3 * ne[3];
+                            x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+                        }
+
+                        ggml_set_param(ctx0, x[0]);
+                        ggml_set_param(ctx0, x[1]);
+
+                        struct ggml_tensor * m = ggml_mul_mat(ctx0, x[1], x[0]);
+                        struct ggml_tensor * f = ggml_sum(ctx0, m);
+
+                        GGML_PRINT_DEBUG("testing: mul_mat, [%lld, %lld] (%d) * [%lld, %lld] (%d)\n", x[1]->ne[0], x[1]->ne[1], x[1]->n_dims, x[0]->ne[0], x[0]->ne[1], x[0]->n_dims);
+
+                        check_gradient("mul_mat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                        if (ndims == 2) {
+                            // check_mat_mul does not support ndims > 2
+                            check_mat_mul(m, x[1], x[0]);
+                        }
+                    }
                }
-
-                ggml_set_param(ctx0, x[0]);
-                ggml_set_param(ctx0, x[1]);
-
-                struct ggml_tensor * m = ggml_mul_mat(ctx0, x[1], x[0]);
-                struct ggml_tensor * f = ggml_sum(ctx0, m);
-
-                GGML_PRINT_DEBUG("testing: mul_mat, [%lld, %lld] (%d) * [%lld, %lld] (%d)\n", x[1]->ne[0], x[1]->ne[1], x[1]->n_dims, x[0]->ne[0], x[0]->ne[1], x[0]->n_dims);
-
-                check_gradient("mul_mat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
-                check_mat_mul(m, x[1], x[0]);
            }
        }

        // elu, not yet fully implemented
        if(0)
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -788,6 +825,7 @@ int main(int argc, const char ** argv) {

        // relu
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -805,6 +843,7 @@ int main(int argc, const char ** argv) {
        // gelu, not yet fully implemented
        if(0)
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -821,6 +860,7 @@ int main(int argc, const char ** argv) {

        // silu
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -842,6 +882,7 @@ int main(int argc, const char ** argv) {

        // rms_norm
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -858,6 +899,7 @@ int main(int argc, const char ** argv) {

        // scale
        {
+            srand(seed);
            const int nargs = 2;

            int64_t ne2[4];
@@ -878,6 +920,7 @@ int main(int argc, const char ** argv) {

        // cpy f32
        {
+            srand(seed);
            const int nargs = 2;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -895,6 +938,7 @@ int main(int argc, const char ** argv) {

        // cpy f16
        {
+            srand(seed);
            const int nargs = 2;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -912,6 +956,7 @@ int main(int argc, const char ** argv) {

        // reshape (1d->nd)
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -935,6 +980,7 @@ int main(int argc, const char ** argv) {

        // reshape (nd->1d)
        {
+            srand(seed);
            const int nargs = 1;

            for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -958,6 +1004,7 @@ int main(int argc, const char ** argv) {

        // acc 1d
        {
+            srand(seed);
            int64_t ne2[4] = { 1, 1, 1, 1 };

            const int nargs = 2;
@@ -985,6 +1032,7 @@ int main(int argc, const char ** argv) {

        // acc 2d
        {
+            srand(seed);
            int64_t ne2[4]         = { 1, 1, 1, 1 };
            int64_t max_offsets[4] = { 0, 0, 0, 0 };
            int64_t offsets[4]     = { 0, 0, 0, 0 };
@@ -1017,6 +1065,7 @@ int main(int argc, const char ** argv) {

        // acc 3d
        {
+            srand(seed);
            int64_t ne2[4]         = { 1, 1, 1, 1 };
            int64_t max_offsets[4] = { 0, 0, 0, 0 };
            int64_t offsets[4]     = { 0, 0, 0, 0 };
@@ -1051,6 +1100,7 @@ int main(int argc, const char ** argv) {

        // acc 4d
        {
+            srand(seed);
            int64_t ne2[4]         = { 1, 1, 1, 1 };
            int64_t max_offsets[4] = { 0, 0, 0, 0 };
            int64_t offsets[4]     = { 0, 0, 0, 0 };
@@ -1087,6 +1137,7 @@ int main(int argc, const char ** argv) {

        // set_1d
        {
+            srand(seed);
            int64_t ne2[4];

            const int nargs = 2;
@@ -1114,6 +1165,7 @@ int main(int argc, const char ** argv) {

        // set_2d
        {
+            srand(seed);
            int64_t ne2[4];
            int64_t max_offsets[4] = { 0, 0, 0, 0 };
            int64_t offsets[4]     = { 0, 0, 0, 0 };
@@ -1146,6 +1198,7 @@ int main(int argc, const char ** argv) {

        // view_1d
        {
+            srand(seed);
            const int nargs = 1;
            for (int ndims = 1; ndims <= 4; ++ndims) {

@@ -1169,6 +1222,7 @@ int main(int argc, const char ** argv) {

        // view_2d
        {
+            srand(seed);
            int64_t ne2[4];
            int64_t nb2[4];

@@ -1199,6 +1253,7 @@ int main(int argc, const char ** argv) {

        // view_3d
        {
+            srand(seed);
            int64_t ne2[4] = {1,1,1,1};
            int64_t nb2[4] = {0,0,0,0};

@@ -1230,6 +1285,7 @@ int main(int argc, const char ** argv) {

        // permute
        {
+            srand(seed);
            int64_t ne2[4];

            const int nargs = 1;
@@ -1263,6 +1319,7 @@ int main(int argc, const char ** argv) {

        // transpose
        {
+            srand(seed);
            int64_t ne2[4];

            const int nargs = 1;
@@ -1290,6 +1347,7 @@ int main(int argc, const char ** argv) {

        // get_rows
        {
+            srand(seed);
            int64_t ne2[4] = {ne[0], ne[1], 1, 1};
            int64_t ne3[4] = {1+irand(ne[1]), 1, 1, 1};
            const int nargs = 1;
@@ -1306,6 +1364,7 @@ int main(int argc, const char ** argv) {

        // diag_mask_inf
        {
+            srand(seed);
            const int nargs = 1;
            const int ndims = 2;

@@ -1321,6 +1380,7 @@ int main(int argc, const char ** argv) {

        // diag_mask_zero
        {
+            srand(seed);
            const int nargs = 1;
            const int ndims = 2;

@@ -1336,6 +1396,7 @@ int main(int argc, const char ** argv) {

        // softmax
        {
+            srand(seed);
            const int nargs = 1;

            int64_t ne2[4];
@@ -1357,11 +1418,16 @@ int main(int argc, const char ** argv) {
                                                    ggml_new_f32(ctx0, eps))));

                check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY);
+                // NOTE: softmax forward is computed using f16 table lookup instead of using actual expf, but backward assumes actual expf.
+                // this may result in different gradients too finite differences.
+                // when this test reports errors, first try to replace the table lookup with actual expf and test again to see if just that was the cause.
+                // if only the table lookup causes gradients to differ this is acceptable.
            }
        }

        // cross_entropy_loss
        {
+            srand(seed);
            const int nargs = 1;

            int64_t ne2[4];
@@ -1392,6 +1458,7 @@ int main(int argc, const char ** argv) {

        // rope f32
        {
+            srand(seed);
            const int nargs = 1;

            int64_t ne2[4];
@@ -1404,6 +1471,11 @@ int main(int argc, const char ** argv) {
                    for (int n_past = 1; n_past < ne2[2]; ++n_past) {
                        x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);

+                        struct ggml_tensor * p = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne2[2]);
+                        for (int i = 0; i < ne2[2]; ++i) {
+                            ((int32_t *) p->data)[i] = n_past + i;
+                        }
+
                        ggml_set_param(ctx0, x[0]);

                        const bool skip_past = (mode & 1);
@@ -1415,7 +1487,7 @@ int main(int argc, const char ** argv) {
                            continue;
                        }

-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode, 0));

                        GGML_PRINT_DEBUG("rope f32: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
                        check_gradient("rope f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);
@@ -1426,6 +1498,7 @@ int main(int argc, const char ** argv) {

        // rope f16
        {
+            srand(seed);
            const int nargs = 1;

            int64_t ne2[4];
@@ -1438,6 +1511,11 @@ int main(int argc, const char ** argv) {
                    for (int n_past = 1; n_past < ne2[2]; ++n_past) {
                        x[0] = get_random_tensor_f16(ctx0, ndims, ne2, -1.0f, 1.0f);

+                        struct ggml_tensor * p = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne2[2]);
+                        for (int i = 0; i < ne2[2]; ++i) {
+                            ((int32_t *) p->data)[i] = n_past + i;
+                        }
+
                        ggml_set_param(ctx0, x[0]);

                        const bool skip_past = (mode & 1);
@@ -1449,7 +1527,7 @@ int main(int argc, const char ** argv) {
                            continue;
                        }

-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode, 0));

                        GGML_PRINT_DEBUG("rope f16: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
                        check_gradient("rope f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);
@@ -1460,6 +1538,7 @@ int main(int argc, const char ** argv) {

        // flash_attn f32
        {
+            srand(seed);
            const int nargs = 3;

            int64_t ne2[4];
@@ -1472,28 +1551,31 @@ int main(int argc, const char ** argv) {

            for (int masked = 0; masked <= 1; ++masked) {
                for (int ndims = 2; ndims <= 4; ++ndims) {
-                    int64_t neq[4] = { D, N, B, ne[3] };
-                    int64_t nek[4] = { D, M, B, ne[3] };
-                    int64_t nev[4] = { M, D, B, ne[3] };
-                    if (ndims == 2) {
-                        neq[2] = 1; neq[3] = 1;
-                        nek[2] = 1; nek[3] = 1;
-                        nev[2] = 1; nev[3] = 1;
-                    } else if (ndims == 3) {
-                        neq[3] = 1;
-                        nek[3] = 1;
-                        nev[3] = 1;
+                    int max_nrep = (ndims >= 3) ? 2 : 1;
+                    for (int nrep = 1; nrep < max_nrep; ++nrep) {
+                        int64_t neq[4] = { D, N, B*nrep, ne[3] };
+                        int64_t nek[4] = { D, M, B, ne[3] };
+                        int64_t nev[4] = { M, D, B, ne[3] };
+                        if (ndims == 2) {
+                            neq[2] = 1; neq[3] = 1;
+                            nek[2] = 1; nek[3] = 1;
+                            nev[2] = 1; nev[3] = 1;
+                        } else if (ndims == 3) {
+                            neq[3] = 1;
+                            nek[3] = 1;
+                            nev[3] = 1;
+                        }
+                        x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                        x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                        x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                        ggml_set_param(ctx0, x[0]);
+                        ggml_set_param(ctx0, x[1]);
+                        ggml_set_param(ctx0, x[2]);
+
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                        check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
                    }
-                    x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
-                    x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
-                    x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
-                    ggml_set_param(ctx0, x[0]);
-                    ggml_set_param(ctx0, x[1]);
-                    ggml_set_param(ctx0, x[2]);
-
-                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
-
-                    check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
                }
            }
        }
@@ -1501,6 +1583,7 @@ int main(int argc, const char ** argv) {
        // flash_attn f16, not yet fully implemented
        if(0)
        {
+            srand(seed);
            const int nargs = 3;

            int64_t ne2[4];
--- a/Show More
+++ b/Show More