wip

err handling + fixing scope issue
wip with json stuff and cleanup
2025-12-31 19:50:04 -05:00 · 2025-03-24 11:02:02 -07:00 · 2025-02-11 19:29:48 -08:00 · 2025-02-11 16:40:40 -08:00 · 2025-02-04 22:09:22 -08:00 · 2025-02-03 16:05:15 -08:00
181 changed files with 7043 additions and 478394 deletions
--- a/.dockerignore
+++ b/.dockerignore
@@ -3,9 +3,7 @@ ollama
 app
 macapp
 dist
-build
 .env
 .cache
 test_data
-.git
-
+llama/build
--- a/.gitattributes
+++ b/.gitattributes
@@ -15,10 +15,6 @@ ml/backend/**/*.cu linguist-vendored
 ml/backend/**/*.cuh linguist-vendored
 ml/backend/**/*.m linguist-vendored
 ml/backend/**/*.metal linguist-vendored
-ml/backend/**/CMakeLists.txt linguist-vendored
-
-llama/build-info.cpp linguist-generated
-ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.s linguist-generated

 * text=auto
 *.go text eol=lf
--- a/.github/ISSUE_TEMPLATE/10_bug_report.yml
+++ b/.github/ISSUE_TEMPLATE/10_bug_report.yml
@@ -9,14 +9,6 @@ body:
      description: What happened? What did you expect to happen?
    validations:
      required: true
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) for details.
-      render: shell
-    validations:
-      required: false
  - type: dropdown
    id: os
    attributes:
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -40,106 +40,28 @@ jobs:

  linux:
    needs: [changes]
-    if: needs.changes.outputs.changed == 'True'
+    if: ${{ needs.changes.outputs.changed == 'True' }}
    strategy:
      matrix:
        include:
-          - preset: CPU
-          - preset: CUDA
-            container: nvidia/cuda:11.8.0-devel-ubuntu22.04
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
-          - preset: ROCm
-            container: rocm/dev-ubuntu-22.04:6.1.2
+          - container: nvidia/cuda:11.8.0-devel-ubuntu22.04
+            preset: CUDA
+          - container: rocm/dev-ubuntu-22.04:6.1.2
+            preset: ROCm
            extra-packages: rocm-libs
-            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_PREFIX_PATH=/opt/rocm'
-    runs-on: linux
+    runs-on: ubuntu-latest
    container: ${{ matrix.container }}
    steps:
      - uses: actions/checkout@v4
      - run: |
-          [ -n "${{ matrix.container }}" ] || sudo=sudo
-          $sudo apt-get update
-          $sudo apt-get install -y cmake ccache ${{ matrix.extra-packages }}
+          apt-get update
+          apt-get install -y cmake pkg-config ${{ matrix.extra-packages }}
        env:
          DEBIAN_FRONTEND: noninteractive
-      - uses: actions/cache@v4
-        with:
-          path: /github/home/.cache/ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
      - run: |
-          cmake --preset ${{ matrix.preset }} ${{ matrix.flags }}
+          cmake --preset ${{ matrix.preset }}
          cmake --build --preset ${{ matrix.preset }} --parallel

-  windows:
-    needs: [changes]
-    if: needs.changes.outputs.changed == 'True'
-    strategy:
-      matrix:
-        include:
-          - preset: CPU
-          - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_522.06_windows.exe
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
-          - preset: ROCm
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe
-            flags: '-DAMDGPU_TARGETS=gfx1010'
-    runs-on: windows
-    steps:
-      - run: |
-          choco install -y --no-progress ccache ninja
-          ccache -o cache_dir=${{ github.workspace }}\.ccache
-      - if: matrix.preset == 'CUDA' || matrix.preset == 'ROCm'
-        id: cache-install
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
-            C:\Program Files\AMD\ROCm
-          key: ${{ matrix.install }}
-      - if: matrix.preset == 'CUDA'
-        name: Install CUDA ${{ matrix.cuda-version }}
-        run: |
-          $ErrorActionPreference = "Stop"
-          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
-            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_11.8", "nvcc_11.8", "cublas_11.8", "cublas_dev_11.8")) -NoNewWindow -Wait
-          }
-
-          $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
-          echo "$cudaPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-      - if: matrix.preset == 'ROCm'
-        name: Install ROCm ${{ matrix.rocm-version }}
-        run: |
-          $ErrorActionPreference = "Stop"
-          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
-            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList '-install' -NoNewWindow -Wait
-          }
-
-          $hipPath = (Resolve-Path "C:\Program Files\AMD\ROCm\*").path
-          echo "$hipPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CC=$hipPath\bin\clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "CXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-      - if: ${{ !cancelled() && steps.cache-install.outputs.cache-hit != 'true' }}
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
-            C:\Program Files\AMD\ROCm
-          key: ${{ matrix.install }}
-      - uses: actions/checkout@v4
-      - uses: actions/cache@v4
-        with:
-          path: ${{ github.workspace }}\.ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
-      - run: |
-          Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
-          cmake --build --parallel --preset "${{ matrix.preset }}"
-        env:
-          CMAKE_GENERATOR: Ninja
-
  test:
    strategy:
      matrix:
@@ -163,5 +85,5 @@ jobs:
      - uses: actions/checkout@v4
      - name: Verify patches apply cleanly and do not change files
        run: |
-          make -f Makefile.sync clean sync
+          make -f Makefile2 clean checkout sync
          git diff --compact-summary --exit-code
--- a/.gitignore
+++ b/.gitignore
@@ -4,13 +4,12 @@
 .venv
 .swp
 dist
-build
 ollama
 .cache
 *.exe
 .idea
 test_data
 *.crt
-__debug_bin*
 llama/build
-llama/vendor
+__debug_bin*
+llama/vendor
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -19,30 +19,11 @@ set(GGML_CCACHE ON)
 set(GGML_BACKEND_DL ON)
 set(GGML_BACKEND_SHARED ON)
 set(GGML_SCHED_MAX_COPIES 4)
-
-set(GGML_LLAMAFILE ON)
+set(GGML_CPU_ALL_VARIANTS ON)
 set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
-set(GGML_CUDA_GRAPHS ON)
+set(GGML_LLAMAFILE ON)

-if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
-    OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
-    set(GGML_CPU_ALL_VARIANTS ON)
-endif()
-
-if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
-    set(CMAKE_BUILD_RPATH "@loader_path")
-    set(CMAKE_INSTALL_RPATH "@loader_path")
-endif()
-
-set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
-set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama)
-
-set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
-set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
-set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)

 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include)
@@ -53,77 +34,12 @@ set(GGML_CPU ON)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)

-get_target_property(CPU_VARIANTS ggml-cpu MANUALLY_ADDED_DEPENDENCIES)
-if(NOT CPU_VARIANTS)
-    set(CPU_VARIANTS "ggml-cpu")
-endif()
-
-install(TARGETS ggml-base ${CPU_VARIANTS}
-    RUNTIME_DEPENDENCIES
-        PRE_EXCLUDE_REGEXES ".*"
-    RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-    LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-    FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-)
-
 check_language(CUDA)
 if(CMAKE_CUDA_COMPILER)
-    if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24" AND NOT CMAKE_CUDA_ARCHITECTURES)
-        set(CMAKE_CUDA_ARCHITECTURES "native")
-    endif()
-
-    find_package(CUDAToolkit)
    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
-    set(OLLAMA_CUDA_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/cuda_v${CUDAToolkit_VERSION_MAJOR})
-    install(TARGETS ggml-cuda
-        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_LIBRARY_DIR}
-            PRE_INCLUDE_REGEXES cublas cublasLt cudart
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
-        LIBRARY DESTINATION ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
-    )
 endif()

-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a):xnack[+-]$"
-    CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a):xnack[+-]$\"."
-)
-
 check_language(HIP)
 if(CMAKE_HIP_COMPILER)
-    set(HIP_PLATFORM "amd")
-
-    find_package(hip REQUIRED)
-    if(NOT AMDGPU_TARGETS)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012])$")
-    elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
-        list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
-    endif()
-
-    if(AMDGPU_TARGETS)
-        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
-
-        if (WIN32)
-            target_compile_definitions(ggml-hip PRIVATE GGML_CUDA_NO_PEER_COPY=1)
-        endif()
-
-        set(OLLAMA_HIP_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/rocm)
-        install(TARGETS ggml-hip
-            RUNTIME_DEPENDENCIES
-                DIRECTORIES ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR}
-                PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register drm drm_amdgpu numa elf
-                PRE_EXCLUDE_REGEXES ".*"
-                POST_EXCLUDE_REGEXES "system32"
-            RUNTIME DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP
-        )
-
-        foreach(HIP_LIB_BIN_INSTALL_DIR IN ITEMS ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR})
-            if(EXISTS ${HIP_LIB_BIN_INSTALL_DIR}/rocblas)
-                install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP)
-                break()
-            endif()
-        endforeach()
-    endif()
+    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
 endif()
--- a/CMakePresets.json
+++ b/CMakePresets.json
@@ -4,15 +4,10 @@
    {
      "name": "Default",
      "binaryDir": "${sourceDir}/build",
-      "installDir": "${sourceDir}/dist",
      "cacheVariables": {
        "CMAKE_BUILD_TYPE": "Release"
      }
    },
-    {
-      "name": "CPU",
-      "inherits": [ "Default" ]
-    },
    {
      "name": "CUDA",
      "inherits": [ "Default" ]
@@ -47,29 +42,20 @@
    },
    {
      "name": "ROCm",
-      "inherits": [ "Default" ],
-      "cacheVariables": {
-        "CMAKE_HIP_PLATFORM": "amd"
-      }
+      "inherits": [ "Default" ]
    },
    {
      "name": "ROCm 6",
      "inherits": [ "ROCm" ],
      "cacheVariables": {
-        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
+        "CMAKE_HIP_ARCHITECTURES": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
      }
    }
  ],
  "buildPresets": [
    {
      "name": "Default",
-      "configurePreset": "Default",
-      "configuration": "Release"
-    },
-    {
-      "name": "CPU",
-      "configurePreset": "Default",
-      "targets": [ "ggml-cpu" ]
+      "configurePreset": "Default"
    },
    {
      "name": "CUDA",
--- a/299
+++ b/299
@@ -1,128 +1,201 @@
-# vim: filetype=dockerfile
+ARG GOLANG_VERSION=1.22.8
+ARG CUDA_VERSION_11=11.3.1
+ARG CUDA_VERSION_12=12.4.0
+ARG ROCM_VERSION=6.1.2
+ARG JETPACK_6=r36.2.0
+ARG JETPACK_5=r35.4.1

-ARG FLAVOR=${TARGETARCH}
+### To create a local image for building linux binaries on mac or windows with efficient incremental builds
+#
+# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile --target unified-builder-amd64 .
+# docker run --platform linux/amd64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-amd64
+#
+### Then incremental builds will be much faster in this container
+#
+# make -j 10 dist
+#
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
+ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
+COPY ./scripts/rh_linux_deps.sh /
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:/usr/local/cuda/bin:$PATH
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+RUN GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+# TODO intel oneapi goes here...
+ENV GOARCH amd64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]

-ARG ROCMVERSION=6.1.2
-ARG JETPACK5VERSION=r35.4.1
-ARG JETPACK6VERSION=r36.2.0
-ARG CMAKEVERSION=3.31.2
+### To create a local image for building linux binaries on mac or linux/arm64 with efficient incremental builds
+# Note: this does not contain jetson variants
+#
+# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile --target unified-builder-arm64 .
+# docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
+#
+FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
+ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
+COPY ./scripts/rh_linux_deps.sh /
+RUN GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
+    dnf config-manager --set-enabled appstream && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH:/usr/local/cuda/bin
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
+ENV GOARCH arm64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]

-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCMVERSION}-complete AS base-amd64
-RUN sed -i -e 's/mirror.centos.org/vault.centos.org/g' -e 's/^#.*baseurl=http/baseurl=http/g' -e 's/^mirrorlist=http/#mirrorlist=http/g' /etc/yum.repos.d/*.repo \
-    && yum install -y yum-utils devtoolset-10-gcc devtoolset-10-gcc-c++ \
-    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo \
-    && curl -s -L https://github.com/ccache/ccache/releases/download/v4.10.2/ccache-4.10.2-linux-x86_64.tar.xz | tar -Jx -C /usr/local/bin --strip-components 1
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:/opt/rh/devtoolset-11/root/usr/bin:$PATH
-
-FROM --platform=linux/arm64 rockylinux:8 AS base-arm64
-# install epel-release for ccache
-RUN yum install -y yum-utils epel-release \
-    && yum install -y clang ccache \
-    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
-ENV CC=clang CXX=clang++
-
-FROM base-${TARGETARCH} AS base
-ARG CMAKEVERSION
-RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-ENV LDFLAGS=-s
-
-FROM base AS cpu
-# amd64 uses gcc which requires devtoolset-11 for AVX extensions while arm64 uses clang
-RUN if [ "$(uname -m)" = "x86_64" ]; then yum install -y devtoolset-11-gcc devtoolset-11-gcc-c++; fi
-ENV PATH=/opt/rh/devtoolset-11/root/usr/bin:$PATH
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CPU' \
-        && cmake --build --parallel --preset 'CPU' \
-        && cmake --install build --component CPU --strip --parallel 8
-
-FROM base AS cuda-11
-ARG CUDA11VERSION=11.3
-RUN yum install -y cuda-toolkit-${CUDA11VERSION//./-}
-ENV PATH=/usr/local/cuda-11/bin:$PATH
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 11' \
-        && cmake --build --parallel --preset 'CUDA 11' \
-        && cmake --install build --component CUDA --strip --parallel 8
-
-FROM base AS cuda-12
-ARG CUDA12VERSION=12.4
-RUN yum install -y cuda-toolkit-${CUDA12VERSION//./-}
-ENV PATH=/usr/local/cuda-12/bin:$PATH
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 12' \
-        && cmake --build --parallel --preset 'CUDA 12' \
-        && cmake --install build --component CUDA --strip --parallel 8
-
-FROM base AS rocm-6
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'ROCm 6' \
-        && cmake --build --parallel --preset 'ROCm 6' \
-        && cmake --install build --component HIP --strip --parallel 8
-
-FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
-ARG CMAKEVERSION
-RUN apt-get update && apt-get install -y curl ccache \
-    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 5' \
-        && cmake --build --parallel --preset 'JetPack 5' \
-        && cmake --install build --component CUDA --strip --parallel 8
-
-FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK6VERSION} AS jetpack-6
-ARG CMAKEVERSION
-RUN apt-get update && apt-get install -y curl ccache \
-    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 6' \
-        && cmake --build --parallel --preset 'JetPack 6' \
-        && cmake --install build --component CUDA --strip --parallel 8
-
-FROM base AS build
-ARG GOVERSION=1.23.4
-RUN curl -fsSL https://golang.org/dl/go${GOVERSION}.linux-$(case $(uname -m) in x86_64) echo amd64 ;; aarch64) echo arm64 ;; esac).tar.gz | tar xz -C /usr/local
-ENV PATH=/usr/local/go/bin:$PATH
-WORKDIR /go/src/github.com/ollama/ollama
+FROM --platform=linux/amd64 unified-builder-amd64 AS build-amd64
 COPY . .
-ARG GOFLAGS="'-ldflags=-w -s'"
-ENV CGO_ENABLED=1
-RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -trimpath -buildmode=pie -o /bin/ollama .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_ROCM_GENERATE
+ARG OLLAMA_FAST_BUILD
+ARG VERSION
+ARG CUSTOM_CPU_FLAGS
+RUN --mount=type=cache,target=/root/.ccache \
+    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
+        make -j $(nproc) dist ; \
+    else \
+        make -j 5 dist ; \
+    fi
+RUN cd dist/linux-$GOARCH && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
+    cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
+    fi

-FROM --platform=linux/amd64 scratch AS amd64
-COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
-COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
+# Jetsons need to be built in discrete stages
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_5} AS runners-jetpack5-arm64
+ARG GOLANG_VERSION
+RUN apt-get update && apt-get install -y git curl ccache && \
+    curl -s -L https://dl.google.com/go/go${GOLANG_VERSION}.linux-arm64.tar.gz | tar xz -C /usr/local && \
+    ln -s /usr/local/go/bin/go /usr/local/bin/go && \
+    ln -s /usr/local/go/bin/gofmt /usr/local/bin/gofmt && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+WORKDIR /go/src/github.com/ollama/ollama/
+COPY . .
+ARG CGO_CFLAGS
+ENV GOARCH arm64
+ARG VERSION
+RUN --mount=type=cache,target=/root/.ccache \
+    make -j 5 dist_cuda_v11 \
+        CUDA_ARCHITECTURES="72;87" \
+        GPU_RUNNER_VARIANT=_jetpack5 \
+        DIST_LIB_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ollama \
+        DIST_GPU_RUNNER_DEPS_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ollama/cuda_jetpack5

-FROM --platform=linux/arm64 scratch AS arm64
-COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
-COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
-COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 lib/ollama/cuda_jetpack5
-COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 lib/ollama/cuda_jetpack6
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_6} AS runners-jetpack6-arm64
+ARG GOLANG_VERSION
+RUN apt-get update && apt-get install -y git curl ccache && \
+    curl -s -L https://dl.google.com/go/go${GOLANG_VERSION}.linux-arm64.tar.gz | tar xz -C /usr/local && \
+    ln -s /usr/local/go/bin/go /usr/local/bin/go && \
+    ln -s /usr/local/go/bin/gofmt /usr/local/bin/gofmt && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+WORKDIR /go/src/github.com/ollama/ollama/
+COPY . .
+ARG CGO_CFLAGS
+ENV GOARCH arm64
+ARG VERSION
+RUN --mount=type=cache,target=/root/.ccache \
+    make -j 5 dist_cuda_v12 \
+        CUDA_ARCHITECTURES="87" \
+        GPU_RUNNER_VARIANT=_jetpack6 \
+        DIST_LIB_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ollama \
+        DIST_GPU_RUNNER_DEPS_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ollama/cuda_jetpack6

-FROM --platform=linux/arm64 scratch AS rocm
-COPY --from=rocm-6 dist/lib/ollama/rocm /lib/ollama/rocm
+FROM --platform=linux/arm64 unified-builder-arm64 AS build-arm64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_FAST_BUILD
+ARG VERSION
+RUN --mount=type=cache,target=/root/.ccache \
+    make -j 5 dist
+COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+RUN cd dist/linux-$GOARCH && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN cd dist/linux-$GOARCH-jetpack5 && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack5.tgz
+RUN cd dist/linux-$GOARCH-jetpack6 && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack6.tgz

-FROM ${FLAVOR} AS archive
-COPY --from=cpu dist/lib/ollama /lib/ollama
-COPY --from=build /bin/ollama /bin/ollama
+FROM --platform=linux/amd64 scratch AS dist-amd64
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM --platform=linux/arm64 scratch AS dist-arm64
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM dist-$TARGETARCH AS dist

-FROM ubuntu:20.04
-RUN apt-get update \
-    && apt-get install -y ca-certificates \
-    && apt-get clean \
-    && rm -rf /var/lib/apt/lists/*
-COPY --from=archive /bin /usr/bin
+
+# For amd64 container images, filter out cuda/rocm to minimize size
+FROM build-amd64 AS runners-cuda-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_hipblas.so \
+    ./dist/linux-amd64/lib/ollama/runners/rocm*
+
+FROM build-amd64 AS runners-rocm-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_cuda*.so \
+    ./dist/linux-amd64/lib/ollama/libcu*.so* \
+    ./dist/linux-amd64/lib/ollama/runners/cuda*
+
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
+FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ /lib/
+COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ /lib/
+
+
+# ROCm libraries larger so we keep it distinct from the CPU/CUDA image
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
+# Frontload the rocm libraries which are large, and rarely change to increase chance of a common layer
+# across releases
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
+EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
+
+ENTRYPOINT ["/bin/ollama"]
+CMD ["serve"]
+
+FROM runtime-$TARGETARCH
+EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
 ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
-COPY --from=archive /lib/ollama /usr/lib/ollama
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
 ENV NVIDIA_VISIBLE_DEVICES=all
-ENV OLLAMA_HOST=0.0.0.0:11434
-EXPOSE 11434
+
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]
--- a/66
+++ b/66
@@ -0,0 +1,66 @@
+ARG CUDA_11_VERSION=11.3
+ARG CUDA_12_VERSION=12.4
+ARG ROCM_VERSION=6.1.2
+ARG JETPACK_5_VERSION=r35.4.1
+ARG JETPACK_6_VERSION=r36.2.0
+ARG CMAKE_VERSION=3.31.2
+
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS base
+ARG CMAKE_VERSION
+RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-x86_64.tar.gz | tar xz -C /usr --strip-components 1
+RUN sed -i -e 's/mirror.centos.org/vault.centos.org/g' -e 's/^#.*baseurl=http/baseurl=http/g' -e 's/^mirrorlist=http/#mirrorlist=http/g' /etc/yum.repos.d/*.repo \
+    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
+
+# FROM --platform=linux/arm64 rockylinux:8 AS base
+# ARG CMAKE_VERSION
+# RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-aarch64.tar.gz | tar xz -C /usr --strip-components 1
+# RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
+
+FROM base AS amd64
+ARG CUDA_11_VERSION
+ARG CUDA_12_VERSION
+RUN yum install -y cuda-toolkit-${CUDA_11_VERSION//./-} \
+    && yum install -y cuda-toolkit-${CUDA_12_VERSION//./-}
+COPY CMakeLists.txt CMakeLists.txt
+COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+
+FROM --platform=linux/amd64 amd64 AS cuda_11
+ENV PATH=/usr/local/cuda-${CUDA_11_VERSION}/bin:$PATH
+RUN cmake -S . -B build -DCMAKE_CUDA_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
+RUN cmake --build build --target ggml-cuda -j
+
+FROM --platform=linux/amd64 amd64 AS cuda_12
+ENV PATH=/usr/local/cuda-${CUDA_12_VERSION}/bin:$PATH
+RUN cmake -S . -B build -DCMAKE_CUDA_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
+RUN cmake --build build --target ggml-cuda -j
+
+FROM --platform=linux/amd64 amd64 AS rocm
+RUN cmake -S . -B build -DCMAKE_HIP_ARCHITECTURES="gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
+RUN cmake --build build --target ggml-hip -j
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_5_VERSION} AS jetpack_5
+ARG CMAKE_VERSION
+RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-aarch64.tar.gz | tar xz -C /usr --strip-components 1
+COPY CMakeLists.txt .
+COPY ml/backend/ggml/ggml .
+RUN cmake -S . -B build \
+    -DCMAKE_CUDA_ARCHITECTURES="72;87"
+RUN cmake --build build --target ggml-cuda
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_6_VERSION} AS jetpack_6
+ARG CMAKE_VERSION
+RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-aarch64.tar.gz | tar xz -C /usr --strip-components 1
+COPY CMakeLists.txt .
+COPY ml/backend/ggml/ggml .
+RUN cmake -S . -B build \
+    -DCMAKE_CUDA_ARCHITECTURES="87"
+RUN cmake --build build --target ggml-cuda
+
+FROM --platform=linux/amd64 golang:1.23
+COPY --from=cuda_11 build/ml/backend/ggml/ggml/src/ggml-cuda/libggml-cuda.so libggml-cuda-11.so
+COPY --from=cuda_12 build/ml/backend/ggml/ggml/src/ggml-cuda/libggml-cuda.so libggml-cuda-12.so
+COPY --from=rocm build/ml/backend/ggml/ggml/src/ggml-hip/libggml-hip.so libggml-hip.so
+
+# FROM --platform=linux/arm64 golang:1.23
+# COPY --from=jetpack_5 build/ml/backend/ggml/ggml/src/ggml-cuda/libggml-cuda.so libggml-cuda-jetpack-5.so
+# COPY --from=jetpack_6 build/ml/backend/ggml/ggml/src/ggml-cuda/libggml-cuda.so libggml-cuda-jetpack-6.so
--- a/Makefile.sync
+++ b/Makefile.sync
@@ -1,60 +0,0 @@
-UPSTREAM=https://github.com/ggerganov/llama.cpp.git
-WORKDIR=llama/vendor
-FETCH_HEAD=46e3556e01b824e52395fb050b29804b6cff2a7c
-
-.PHONY: help
-help:
-	@echo "Available targets:"
-	@echo "    sync                 Sync with upstream repositories"
-	@echo "    checkout             Checkout upstream repository"
-	@echo "    apply-patches        Apply patches to local repository"
-	@echo "    format-patches       Format patches from local repository"
-	@echo "    clean                Clean local repository"
-	@echo
-	@echo "Example:"
-	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean sync"
-
-.PHONY: sync
-sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml apply-patches
-
-.PHONY: llama/build-info.cpp
-llama/build-info.cpp: llama/build-info.cpp.in
-	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' $< > $@
-
-.PHONY: llama/llama.cpp
-llama/llama.cpp: llama/vendor/ apply-patches
-	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
-
-.PHONY: ml/backend/ggml/ggml apply-patches
-ml/backend/ggml/ggml: llama/vendor/ggml/ apply-patches
-	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
-
-PATCHES=$(wildcard llama/patches/*.patch)
-
-.PHONY: apply-patches
-.NOTPARALLEL:
-apply-patches: $(addsuffix ed, $(PATCHES))
-
-%.patched: %.patch
-	@if git -c user.name=nobody -c 'user.email=<>' -C $(WORKDIR) am -3 $(realpath $<); then touch $@; else git -C $(WORKDIR) am --abort; exit 1; fi
-
-.PHONY: checkout
-checkout: $(WORKDIR)
-	git -C $(WORKDIR) fetch
-	git -C $(WORKDIR) checkout -f $(FETCH_HEAD)
-
-$(WORKDIR):
-	git clone $(UPSTREAM) $(WORKDIR)
-
-.PHONE: format-patches
-format-patches: llama/patches
-	git -C $(WORKDIR) format-patch \
-		--no-signature \
-		--no-numbered \
-		--zero-commit \
-		-o $(realpath $<) \
-		$(FETCH_HEAD)
-
-.PHONE: clean
-clean: checkout
-	$(RM) $(addsuffix ed, $(PATCHES))
--- a/46
+++ b/46
@@ -0,0 +1,46 @@
+UPSTREAM=https://github.com/ggerganov/llama.cpp.git
+WORKDIR=llama/vendor
+FETCH_HEAD=46e3556e01b824e52395fb050b29804b6cff2a7c
+
+all: sync
+
+.PHONY: sync
+sync: llama/llama.cpp ml/backend/ggml/ggml
+
+.PHONY: llama/llama.cpp
+llama/llama.cpp: llama/vendor/ apply_patches
+	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
+
+.PHONY: ml/backend/ggml/ggml apply_patches
+ml/backend/ggml/ggml: llama/vendor/ggml/ apply_patches
+	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
+
+PATCHES=$(wildcard llama/patches/*.patch)
+
+.PHONY: apply_patches
+.NOTPARALLEL:
+apply_patches: $(addsuffix ed, $(PATCHES))
+
+%.patched: %.patch
+	@if git -c user.name=nobody -c 'user.email=<>' -C $(WORKDIR) am -3 $(realpath $<); then touch $@; else git -C $(WORKDIR) am --abort; exit 1; fi
+
+.PHONY: checkout
+checkout: $(WORKDIR)
+	git -C $(WORKDIR) fetch
+	git -C $(WORKDIR) checkout -f $(FETCH_HEAD)
+
+$(WORKDIR):
+	git clone $(UPSTREAM) $(WORKDIR)
+
+.PHONE: format_patches
+format_patches: llama/patches
+	git -C $(WORKDIR) format-patch \
+		--no-signature \
+		--no-numbered \
+		--zero-commit \
+		-o $(realpath $<) \
+		$(FETCH_HEAD)
+
+.PHONE: clean
+clean: checkout
+	$(RM) $(addsuffix ed, $(PATCHES))
--- a/README.md
+++ b/README.md
@@ -18,7 +18,7 @@ Get up and running with large language models.

 ### Linux

-```shell
+```
 curl -fsSL https://ollama.com/install.sh | sh
 ```

@@ -42,7 +42,7 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla

 To run and chat with [Llama 3.2](https://ollama.com/library/llama3.2):

-```shell
+```
 ollama run llama3.2
 ```

@@ -54,8 +54,6 @@ Here are some example models that can be downloaded:

 | Model              | Parameters | Size  | Download                         |
 | ------------------ | ---------- | ----- | -------------------------------- |
-| DeepSeek-R1        | 7B         | 4.7GB | `ollama run deepseek-r1`         |
-| DeepSeek-R1        | 671B       | 404GB | `ollama run deepseek-r1:671b`    |
 | Llama 3.3          | 70B        | 43GB  | `ollama run llama3.3`            |
 | Llama 3.2          | 3B         | 2.0GB | `ollama run llama3.2`            |
 | Llama 3.2          | 1B         | 1.3GB | `ollama run llama3.2:1b`         |
@@ -94,13 +92,13 @@ Ollama supports importing GGUF models in the Modelfile:

 2. Create the model in Ollama

-   ```shell
+   ```
   ollama create example -f Modelfile
   ```

 3. Run the model

-   ```shell
+   ```
   ollama run example
   ```

@@ -112,7 +110,7 @@ See the [guide](docs/import.md) on importing models for more information.

 Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.2` model:

-```shell
+```
 ollama pull llama3.2
 ```

@@ -147,13 +145,13 @@ For more information on working with a Modelfile, see the [Modelfile](docs/model

 `ollama create` is used to create a model from a Modelfile.

-```shell
+```
 ollama create mymodel -f ./Modelfile
 ```

 ### Pull a model

-```shell
+```
 ollama pull llama3.2
 ```

@@ -161,13 +159,13 @@ ollama pull llama3.2

 ### Remove a model

-```shell
+```
 ollama rm llama3.2
 ```

 ### Copy a model

-```shell
+```
 ollama cp llama3.2 my-model
 ```

@@ -186,39 +184,37 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol

 ```
 ollama run llava "What's in this image? /Users/jmorgan/Desktop/smile.png"
+The image features a yellow smiley face, which is likely the central focus of the picture.
 ```

-> **Output**: The image features a yellow smiley face, which is likely the central focus of the picture.
-
 ### Pass the prompt as an argument

-```shell
-ollama run llama3.2 "Summarize this file: $(cat README.md)"
 ```
-
-> **Output**: Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
+$ ollama run llama3.2 "Summarize this file: $(cat README.md)"
+ Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
+```

 ### Show model information

-```shell
+```
 ollama show llama3.2
 ```

 ### List models on your computer

-```shell
+```
 ollama list
 ```

 ### List which models are currently loaded

-```shell
+```
 ollama ps
 ```

 ### Stop a model which is currently running

-```shell
+```
 ollama stop llama3.2
 ```

@@ -234,13 +230,13 @@ See the [developer guide](https://github.com/ollama/ollama/blob/main/docs/develo

 Next, start the server:

-```shell
+```
 ./ollama serve
 ```

 Finally, in a separate shell, run a model:

-```shell
+```
 ./ollama run llama3.2
 ```

@@ -250,7 +246,7 @@ Ollama has a REST API for running and managing models.

 ### Generate a response

-```shell
+```
 curl http://localhost:11434/api/generate -d '{
  "model": "llama3.2",
  "prompt":"Why is the sky blue?"
@@ -259,7 +255,7 @@ curl http://localhost:11434/api/generate -d '{

 ### Chat with a model

-```shell
+```
 curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2",
  "messages": [
@@ -357,7 +353,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Web management](https://github.com/lemonit-eric-mao/ollama-web-management) (Web management page)
 - [Promptery](https://github.com/promptery/promptery) (desktop client for Ollama.)
 - [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
- [chat-ollama](https://github.com/annilq/chat-ollama) (a React Native client for Ollama)
 - [SpaceLlama](https://github.com/tcsenpai/spacellama) (Firefox and Chrome extension to quickly summarize web pages with ollama in a sidebar)
 - [YouLama](https://github.com/tcsenpai/youlama) (Webapp to quickly summarize any YouTube video, supporting Invidious as well)
 - [DualMind](https://github.com/tcsenpai/dualmind) (Experimental app allowing two models to talk to each other in the terminal or in a web interface)
@@ -374,14 +369,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Minima](https://github.com/dmayboroda/minima) (RAG with on-premises or fully local workflow)
 - [aidful-ollama-model-delete](https://github.com/AidfulAI/aidful-ollama-model-delete) (User interface for simplified model cleanup)
 - [Perplexica](https://github.com/ItzCrazyKns/Perplexica) (An AI-powered search engine & an open-source alternative to Perplexity AI)
- [Ollama Chat WebUI for Docker ](https://github.com/oslook/ollama-webui) (Support for local docker deployment, lightweight ollama webui)
- [AI Toolkit for Visual Studio Code](https://aka.ms/ai-tooklit/ollama-docs) (Microsoft-official VSCode extension to chat, test, evaluate models with Ollama support, and use them in your AI applications.)
- [MinimalNextOllamaChat](https://github.com/anilkay/MinimalNextOllamaChat) (Minimal Web UI for Chat and Model Control)
- [Chipper](https://github.com/TilmanGriesel/chipper) AI interface for tinkerers (Ollama, Haystack RAG, Python)
- [ChibiChat](https://github.com/CosmicEventHorizon/ChibiChat) (Kotlin-based Android app to chat with Ollama and Koboldcpp API endpoints)
- [LocalLLM](https://github.com/qusaismael/localllm) (Minimal Web-App to run ollama models on it with a GUI)
- [Ollamazing](https://github.com/buiducnhat/ollamazing) (Web extension to run Ollama models)
- [OpenDeepResearcher-via-searxng](https://github.com/benhaotang/OpenDeepResearcher-via-searxng) (A Deep Research equivent endpoint with Ollama support for running locally)

 ### Cloud

@@ -439,10 +426,9 @@ See the [API documentation](./docs/api.md) for all endpoints.

 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
 - [Gentoo](https://github.com/gentoo/guru/tree/master/app-misc/ollama)
- [Homebrew](https://formulae.brew.sh/formula/ollama)
 - [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
 - [Guix channel](https://codeberg.org/tusharhero/ollama-guix)
- [Nix package](https://search.nixos.org/packages?show=ollama&from=0&size=50&sort=relevance&type=packages&query=ollama)
+- [Nix package](https://search.nixos.org/packages?channel=24.05&show=ollama&from=0&size=50&sort=relevance&type=packages&query=ollama)
 - [Flox](https://flox.dev/blog/ollama-part-one)

 ### Libraries
@@ -495,9 +481,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [GoLamify](https://github.com/prasad89/golamify)
 - [Ollama for Haskell](https://github.com/tusharad/ollama-haskell)
 - [multi-llm-ts](https://github.com/nbonamy/multi-llm-ts) (A Typescript/JavaScript library allowing access to different LLM in unified API)
- [LlmTornado](https://github.com/lofcz/llmtornado) (C# library providing a unified interface for major FOSS & Commercial inference APIs)
- [Ollama for Zig](https://github.com/dravenk/ollama-zig)
- [Abso](https://github.com/lunary-ai/abso) (OpenAI-compatible TypeScript SDK for any LLM provider)

 ### Mobile

@@ -548,16 +531,12 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [TextCraft](https://github.com/suncloudsmoon/TextCraft) (Copilot in Word alternative using Ollama)
 - [Alfred Ollama](https://github.com/zeitlings/alfred-ollama) (Alfred Workflow)
 - [TextLLaMA](https://github.com/adarshM84/TextLLaMA) A Chrome Extension that helps you write emails, correct grammar, and translate into any language
- [Simple-Discord-AI](https://github.com/zyphixor/simple-discord-ai)
- [LLM Telegram Bot](https://github.com/innightwolfsleep/llm_telegram_bot) (telegram bot, primary for RP. Oobabooga-like buttons, [A1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) API integration e.t.c)

 ### Supported backends

 - [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.

 ### Observability
- [Lunary](https://lunary.ai/docs/integrations/ollama) is the leading open-source LLM observability platform. It provides a variety of enterprise-grade features such as real-time analytics, prompt templates management, PII masking, and comprehensive agent tracing.
+
 - [OpenLIT](https://github.com/openlit/openlit) is an OpenTelemetry-native tool for monitoring Ollama Applications & GPUs using traces and metrics.
- [HoneyHive](https://docs.honeyhive.ai/integrations/ollama) is an AI observability and evaluation platform for AI agents. Use HoneyHive to evaluate agent performance, interrogate failures, and monitor quality in production.
- [Langfuse](https://langfuse.com/docs/integrations/ollama) is an open source LLM observability platform that enables teams to collaboratively monitor, evaluate and debug AI applications.
- [MLflow Tracing](https://mlflow.org/docs/latest/llms/tracing/index.html#automatic-tracing) is an open source LLM observability tool with a convenient API to log and visualize traces, making it easy to debug and evaluate GenAI applications.
+- [HoneyHive](https://docs.honeyhive.ai/integrations/ollama) is an AI observability and evaluation platform for AI agents. Use HoneyHive to evaluate agent performance, interrogate failures, and monitor quality in production. 
--- a/api/client.go
+++ b/api/client.go
@@ -126,8 +126,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 			return err
 		}
 	}
-
-	return ctx.Err()
+	return nil
 }

 const maxBufferSize = 512 * format.KiloByte
@@ -190,7 +189,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		}
 	}

-	return ctx.Err()
+	return nil
 }

 // GenerateResponseFunc is a function that [Client.Generate] invokes every time
--- a/api/examples/README.md
+++ b/api/examples/README.md
@@ -2,10 +2,9 @@

 Run the examples in this directory with:

-```shell
+```
 go run example_name/main.go
 ```
-
 ## Chat - Chat with a model
 - [chat/main.go](chat/main.go)

--- a/app/README.md
+++ b/app/README.md
@@ -17,6 +17,6 @@ If you want to build the installer, youll need to install
 In the top directory of this repo, run the following powershell script
 to build the ollama CLI, ollama app, and ollama installer.

-```powershell
+```
 powershell -ExecutionPolicy Bypass -File .\scripts\build_windows.ps1
 ```
--- a/cache/cache.go
+++ b/cache/cache.go
@@ -0,0 +1,63 @@
+package cache
+
+import (
+	"github.com/ollama/ollama/ml"
+)
+
+type Options struct {
+	Position int
+}
+
+type Cache interface {
+	Sub(i int) Cache
+	Put(ctx ml.Context, key, value ml.Tensor, opts Options) (ml.Tensor, ml.Tensor)
+}
+
+type Simple struct {
+	DType    ml.DType
+	Capacity int
+
+	keys, values []ml.Tensor
+}
+
+func (c *Simple) Sub(i int) Cache {
+	if i >= len(c.keys) {
+		c.keys = append(c.keys, make([]ml.Tensor, i-len(c.keys)+1)...)
+		c.values = append(c.values, make([]ml.Tensor, i-len(c.values)+1)...)
+	}
+
+	return &Simple{
+		keys:     c.keys[i : i+1],
+		values:   c.values[i : i+1],
+		Capacity: c.Capacity,
+		DType:    c.DType,
+	}
+}
+
+func (c *Simple) Put(ctx ml.Context, key, value ml.Tensor, opts Options) (ml.Tensor, ml.Tensor) {
+	if c.keys[0] == nil || c.values[0] == nil {
+		c.keys[0] = ctx.Zeros(c.DType, int(key.Dim(0)*key.Dim(1))*c.Capacity)
+		c.values[0] = ctx.Zeros(c.DType, int(value.Dim(0)*value.Dim(1))*c.Capacity)
+	}
+
+	ctx.Forward(key.Copy(ctx, c.keys[0].View(ctx, int(key.Stride(2))*opts.Position, int(key.Dim(0)*key.Dim(1)*key.Dim(2)))))
+	ctx.Forward(value.Copy(ctx, c.values[0].View(ctx, int(value.Stride(2))*opts.Position, int(value.Dim(0)*value.Dim(1)*value.Dim(2)))))
+
+	n := min(c.Capacity, int(key.Dim(2))+opts.Position)
+
+	key = c.keys[0].View(ctx, 0,
+		int(key.Dim(0)), int(key.Stride(1)),
+		int(key.Dim(1)), int(key.Stride(2)),
+		n,
+	)
+
+	value = c.values[0].View(ctx, 0,
+		int(value.Dim(0)), int(value.Stride(1)),
+		int(value.Dim(1)), int(value.Stride(2)),
+		n,
+	)
+
+	// TODO shift context if necessary
+
+	return key, value
+}
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -15,11 +15,13 @@ import (
 	"net"
 	"net/http"
 	"os"
+	"os/signal"
 	"path/filepath"
 	"runtime"
 	"strconv"
 	"strings"
 	"sync/atomic"
+	"syscall"
 	"time"

 	"github.com/containerd/console"
@@ -33,9 +35,9 @@ import (
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/llama"
+	"github.com/ollama/ollama/llama/runner"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
-	"github.com/ollama/ollama/runner"
 	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -57,7 +59,7 @@ func getModelfileName(cmd *cobra.Command) (string, error) {

 	_, err = os.Stat(absName)
 	if err != nil {
-		return "", err
+		return filename, err
 	}

 	return absName, nil
@@ -328,7 +330,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 			if err := PullHandler(cmd, []string{name}); err != nil {
 				return nil, err
 			}
-
 			return client.Show(cmd.Context(), &api.ShowRequest{Name: name})
 		}
 		return info, err
@@ -337,10 +338,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}

-	// TODO(jessegross): We should either find another way to know if this is
-	// a vision model or remove the logic. Also consider that other modalities will
-	// need different behavior anyways.
-	opts.MultiModal = len(info.ProjectorInfo) != 0 || envconfig.NewEngine()
+	opts.MultiModal = len(info.ProjectorInfo) != 0
 	opts.ParentModel = info.Details.ParentModel

 	if interactive {
@@ -857,6 +855,17 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	spinner := progress.NewSpinner("")
 	p.Add("", spinner)

+	cancelCtx, cancel := context.WithCancel(cmd.Context())
+	defer cancel()
+
+	sigChan := make(chan os.Signal, 1)
+	signal.Notify(sigChan, syscall.SIGINT)
+
+	go func() {
+		<-sigChan
+		cancel()
+	}()
+
 	var state *displayResponseState = &displayResponseState{}
 	var latest api.ChatResponse
 	var fullResponse strings.Builder
@@ -891,7 +900,10 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		req.KeepAlive = opts.KeepAlive
 	}

-	if err := client.Chat(cmd.Context(), req, fn); err != nil {
+	if err := client.Chat(cancelCtx, req, fn); err != nil {
+		if errors.Is(err, context.Canceled) {
+			return nil, nil
+		}
 		return nil, err
 	}

@@ -931,6 +943,17 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		generateContext = []int{}
 	}

+	ctx, cancel := context.WithCancel(cmd.Context())
+	defer cancel()
+
+	sigChan := make(chan os.Signal, 1)
+	signal.Notify(sigChan, syscall.SIGINT)
+
+	go func() {
+		<-sigChan
+		cancel()
+	}()
+
 	var state *displayResponseState = &displayResponseState{}

 	fn := func(response api.GenerateResponse) error {
@@ -966,7 +989,10 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		KeepAlive: opts.KeepAlive,
 	}

-	if err := client.Generate(cmd.Context(), &request, fn); err != nil {
+	if err := client.Generate(ctx, &request, fn); err != nil {
+		if errors.Is(err, context.Canceled) {
+			return nil
+		}
 		return err
 	}

@@ -988,7 +1014,8 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		latest.Summary()
 	}

-	cmd.SetContext(context.WithValue(cmd.Context(), generateContextKey("context"), latest.Context))
+	ctx = context.WithValue(cmd.Context(), generateContextKey("context"), latest.Context)
+	cmd.SetContext(ctx)

 	return nil
 }
--- a/cmd/cmd_test.go
+++ b/cmd/cmd_test.go
@@ -279,7 +279,7 @@ func TestGetModelfileName(t *testing.T) {
 			name:          "no modelfile specified, no modelfile exists",
 			modelfileName: "",
 			fileExists:    false,
-			expectedName:  "",
+			expectedName:  "Modelfile",
 			expectedErr:   os.ErrNotExist,
 		},
 		{
@@ -293,7 +293,7 @@ func TestGetModelfileName(t *testing.T) {
 			name:          "modelfile specified, no modelfile exists",
 			modelfileName: "crazyfile",
 			fileExists:    false,
-			expectedName:  "",
+			expectedName:  "crazyfile",
 			expectedErr:   os.ErrNotExist,
 		},
 		{
--- a/cmd/runner/main.go
+++ b/cmd/runner/main.go
@@ -4,7 +4,7 @@ import (
 	"fmt"
 	"os"

-	"github.com/ollama/ollama/runner"
+	"github.com/ollama/ollama/llama/runner"
 )

 func main() {
--- a/convert/convert.go
+++ b/convert/convert.go
@@ -191,8 +191,6 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		conv = &qwen2Model{}
 	case "BertModel":
 		conv = &bertModel{}
-	case "CohereForCausalLM":
-		conv = &commandrModel{}
 	default:
 		return errors.New("unsupported architecture")
 	}
--- a/convert/convert_commandr.go
+++ b/convert/convert_commandr.go
@@ -1,76 +0,0 @@
-package convert
-
-import (
-	"cmp"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type commandrModel struct {
-	ModelParameters
-	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-	HiddenSize            uint32  `json:"hidden_size"`
-	HiddenLayers          uint32  `json:"num_hidden_layers"`
-	IntermediateSize      uint32  `json:"intermediate_size"`
-	NumAttentionHeads     uint32  `json:"num_attention_heads"`
-	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
-	LayerNormEPS          float32 `json:"layer_norm_eps"`
-	RopeTheta             float32 `json:"rope_theta"`
-	UseQKNorm             bool    `json:"use_qk_norm"`
-	MaxLength             uint32  `json:"model_max_length"`
-	LogitScale            float32 `json:"logit_scale"`
-	NCtx                  uint32  `json:"n_ctx"`
-}
-
-var _ ModelConverter = (*commandrModel)(nil)
-
-func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
-	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "command-r"
-	kv["general.name"] = "command-r"
-	kv["command-r.context_length"] = cmp.Or(p.MaxLength, p.MaxPositionEmbeddings, p.NCtx)
-	kv["command-r.embedding_length"] = p.HiddenSize
-	kv["command-r.block_count"] = p.HiddenLayers
-	kv["command-r.feed_forward_length"] = p.IntermediateSize
-	kv["command-r.attention.head_count"] = p.NumAttentionHeads
-	kv["command-r.attention.head_count_kv"] = p.NumKeyValueHeads
-	kv["command-r.attention.layer_norm_epsilon"] = p.LayerNormEPS
-	kv["command-r.rope.freq_base"] = p.RopeTheta
-	kv["command-r.max_position_embeddings"] = cmp.Or(p.MaxLength, p.MaxPositionEmbeddings)
-	kv["command-r.logit_scale"] = p.LogitScale
-	kv["command-r.rope.scaling.type"] = "none"
-
-	return kv
-}
-
-func (p *commandrModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
-	for _, t := range ts {
-		out = append(out, ggml.Tensor{
-			Name:     t.Name(),
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (p *commandrModel) Replacements() []string {
-	return []string{
-		"self_attn.q_norm", "attn_q_norm",
-		"self_attn.k_norm", "attn_k_norm",
-		"model.layers", "blk",
-		"input_layernorm", "attn_norm",
-		"mlp.down_proj", "ffn_down",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.up_proj", "ffn_up",
-		"self_attn.k_proj", "attn_k",
-		"self_attn.o_proj", "attn_output",
-		"self_attn.q_proj", "attn_q",
-		"self_attn.v_proj", "attn_v",
-		"model.norm", "output_norm",
-		"model.embed_tokens", "token_embd",
-	}
-}
--- a/convert/convert_qwen2.go
+++ b/convert/convert_qwen2.go
@@ -2,6 +2,7 @@ package convert

 import "github.com/ollama/ollama/fs/ggml"

+
 type qwen2Model struct {
 	ModelParameters
 	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
--- a/convert/convert_test.go
+++ b/convert/convert_test.go
@@ -109,7 +109,6 @@ func TestConvertModel(t *testing.T) {
 		"all-MiniLM-L6-v2",
 		"gemma-2-9b-it",
 		"Qwen2.5-0.5B-Instruct",
-		"c4ai-command-r-v01",
 	}

 	for i := range cases {
--- a/convert/testdata/c4ai-command-r-v01.json
+++ b/convert/testdata/c4ai-command-r-v01.json
@@ -1,344 +0,0 @@
-{
-    "general.architecture": "command-r",
-    "general.name": "command-r",
-    "command-r.attention.head_count": "64",
-    "command-r.attention.head_count_kv": "64",
-    "command-r.attention.layer_norm_epsilon": "1e-05",
-    "command-r.block_count": "40",
-    "command-r.context_length": "131072",
-    "command-r.embedding_length": "8192",
-    "command-r.feed_forward_length": "22528",
-    "command-r.logit_scale": "0.0625",
-    "command-r.rope.freq_base": "8e+06",
-    "command-r.rope.scaling.type": "none",
-    "tokenizer.ggml.add_bos_token": "true",
-    "tokenizer.ggml.add_eos_token": "false",
-    "tokenizer.ggml.bos_token_id": "5",
-    "tokenizer.ggml.eos_token_id": "255001",
-    "tokenizer.ggml.merges": "902a060cac8884a5793d2a857dd2e53a259de46c8d08c4deb243c239671e1350",
-    "tokenizer.ggml.model": "gpt2",
-    "tokenizer.ggml.padding_token_id": "0",
-    "tokenizer.ggml.token_type": "b7a352ccd1c99d4413bcf452c2db707b0526d0e1216616b865560fab80296462",
-    "tokenizer.ggml.tokens": "815ac90ff23565081522d7258f46648c8a0619eb847a9c7c31b238a9b984e4ae",
-    "blk.0.attn_k.weight": "6fcfdb466f9ceb1229404ce4ec4e480751b8d00da12707a11783dad7256cb864",
-    "blk.0.attn_norm.weight": "6063317f731371864049c7704a70772f1eb632194201ebdc2ed0f8e483507c72",
-    "blk.0.attn_output.weight": "920f49716a1e2fc73b6794ec777947f1c122701e63ed302422ac89e90f06e9da",
-    "blk.0.attn_q.weight": "ddbcd7cde197e632564ac58e4f25d9e3a8ca52917329eeb6081eb41a797932ab",
-    "blk.0.attn_v.weight": "318fc02a189d87420f0cbf57f47f11e00c21ec1ed472ce0a2a895b44f7fa0fca",
-    "blk.0.ffn_down.weight": "aa71975b6eb1f4c77b03d2ac4a194cf8d95718efac741bb12f0f3ff79a27f9bc",
-    "blk.0.ffn_gate.weight": "42967702fa0bc738b88dc50007ace26dbe74a5a9e0978124dd093f818241a9e1",
-    "blk.0.ffn_up.weight": "5282c8788b086bd30f46525e7995a17464882a72703fd27165491afdd8bfd4af",
-    "blk.1.attn_k.weight": "cd248882e64fd2c3402c44790ebe12440133dc671b6893fdad0564c461973adc",
-    "blk.1.attn_norm.weight": "ba84e1c8fd30af6ec94208db4078befac8c921aad3acb887812887f3282ea2be",
-    "blk.1.attn_output.weight": "2efa3ef7c5666ccceb05e339b83ad680cc0d2c3ec78203f5da5959f23a80e14f",
-    "blk.1.attn_q.weight": "5106f2e255358a1303c22e8b5f0ec044852bb30a866c52cabefd30017a7a6b7d",
-    "blk.1.attn_v.weight": "a211a634a1a5df1d5f973645438be0461dd922210f9747c6b04e386c7f1ebe95",
-    "blk.1.ffn_down.weight": "37093afe48d32c578ec956c9ed85242cd000d6aa979e60526aafa10c822dbb10",
-    "blk.1.ffn_gate.weight": "469860819e9159caefb1aad0bc66db790f3393f05fd87b08e52256a7ed256543",
-    "blk.1.ffn_up.weight": "736742c97d35d1a011f9cafd3c0ce947ad559bb2fba6da73c816f6bfd0fa9aeb",
-    "blk.2.attn_k.weight": "92c219d92804d832ab404bd6dc7339c90877bb7cf405dd030c121f8b27757739",
-    "blk.2.attn_norm.weight": "61e4466069474b76b6d1e702566420eb669faf3556b00ff7b824784aca13a2d6",
-    "blk.2.attn_output.weight": "d2fb38a2b2171fd91caf037faa585a62225819aa232d86fd4f7f9d2c3c8a45e9",
-    "blk.2.attn_q.weight": "f6faf5cc6844e3daa4f9f68d90f5458c64879de68a7728860e38374e30c3429d",
-    "blk.2.attn_v.weight": "f340ef8f7341d987a6f37c0e9afe0aef5be67be00c0ce5f57612daf73319cce1",
-    "blk.2.ffn_down.weight": "c7be61a701d779860b621b143fb6365b607bf99ec7c0f153b07908ac8120885a",
-    "blk.2.ffn_gate.weight": "b64f0878187bd3392abfa4c3e8ad2f8b4c133903e54246747ff8f3b4639ad83e",
-    "blk.2.ffn_up.weight": "50b11c712652e90ee7428dbb45cffebb80662ac982bc72bd9eafff361b5eb5a8",
-    "blk.3.attn_k.weight": "2b7bcbe9ee5c9c630c8c8d7483887e78b73581016f4cbb6933db2a147a25f431",
-    "blk.3.attn_norm.weight": "0181dac7f4eee7252980323e8032cf339bef2046ce0a16c0fd72af7c98a8a37b",
-    "blk.3.attn_output.weight": "aef8843b636ce231da9e7c9acbee197883cc15df0e2887709324c6a50f16da7b",
-    "blk.3.attn_q.weight": "55404130fa10e81322d33eb378aa0de31a92990ce7730f1338c0ace0406bb1b1",
-    "blk.3.attn_v.weight": "76f7fb8040d82b957d689ce34fea2302a6640ad5bbaa0052ad2b7ebce270c33d",
-    "blk.3.ffn_down.weight": "648628933eff3b357c3729c33c5b1ae51c28e59b9c19acd1601a2ff7c5d5d9a5",
-    "blk.3.ffn_gate.weight": "6a588885d16e98d5f50ebed05af089154f680085ca9c97691e5b489088630a4a",
-    "blk.3.ffn_up.weight": "e12455a1d702f4986e1a663493e3d5102b367af74d45557522002a35d63ecac2",
-    "blk.4.attn_k.weight": "40d943380a8a85e4eab147934bf6e16f23cc8ab753f6636526382c074d182288",
-    "blk.4.attn_norm.weight": "4ab2c098983d4599fe540eef624c4df954adb7473faebda7471ef0ba4134814c",
-    "blk.4.attn_output.weight": "d14b91e40f58bf4a3c8c2eca0b12bb541de406574af39027d56f6c588a147082",
-    "blk.4.attn_q.weight": "e1224960a3562107488589f883fa32414bae41712fa8dbd47c5f3e3a7801452f",
-    "blk.4.attn_v.weight": "063f297bc4aa6e709fc32c4c32e35af7d07d80e83cb939b76adbba858006c03d",
-    "blk.4.ffn_down.weight": "f88a18020c5e1caaa29596895eb348e76ee5bfad27ed57651a86cd8cd1f9b5aa",
-    "blk.4.ffn_gate.weight": "48e7e1eed3fb52e92e61d3557dd0ec002418327090e034ce4322fd68542266f8",
-    "blk.4.ffn_up.weight": "1ca8a7aa17355b6ce0d9ad5539fdad3899fa47fd359c285fbfb31f19f47bf073",
-    "blk.5.attn_k.weight": "2bdf15f8e73d068d972380f25d207004cf0bf3b5bfa46946803ba6fba07d9175",
-    "blk.5.attn_norm.weight": "60448d7cde6e1b6467aa31bdea012e39cdb08c88081cee7d102dca4f93f766ef",
-    "blk.5.attn_output.weight": "f9f687d7c457537f9fca8a4087a59f1c3bebfaf5537b94e42c831a13224f7799",
-    "blk.5.attn_q.weight": "987db7a2ad68657a92625e1980effbb1f79697c2183f2b9f3b3a0570c51b0ab9",
-    "blk.5.attn_v.weight": "cf696891148f3e4783ad1d20f93462ae091eb8651c656bba9b662253b6263e02",
-    "blk.5.ffn_down.weight": "c0662b0bd0929136005fb9d691fdd9b2c33867d9ce9622339a6a456b720b059a",
-    "blk.5.ffn_gate.weight": "200bbdfab615d7a3a84719b6ced7751e3ce52757ef212d96f87798bc1de5e987",
-    "blk.5.ffn_up.weight": "df5d23e7e035fb1b9d163da7ddfdfe38da6a37e86e96534dc02ad20f011b55b3",
-    "blk.6.attn_k.weight": "c0dae2d272a7c5a2fa004bbb8475dbab362fc1f6d008e73d5a4434a9382ac6ba",
-    "blk.6.attn_norm.weight": "51c57ac8b55e04354d5dca6bb9c0cf4177639d3b038e80209e33036209688f64",
-    "blk.6.attn_output.weight": "229d97892c62f85bcdf431675250e01c976ad69ffa450b01fb543bf88f14a2fb",
-    "blk.6.attn_q.weight": "c20e49621821bd46ed156e6823864a5bda4f317750e71ab8dc54e44eb48cf7c2",
-    "blk.6.attn_v.weight": "53ceb1a2ee43fce3c7b5b33c58a9fc5ee7f44dc1c6f29bc9dbefc37582102dc9",
-    "blk.6.ffn_down.weight": "7923c943b7629d560a032d1efa210d1d75c6692140f1be94464ee7ed24f44ed0",
-    "blk.6.ffn_gate.weight": "57593d350361af753a6a39f53b066282634c0fb44f396f6f2966a574b01d8f8c",
-    "blk.6.ffn_up.weight": "327b6a7a387098b8899d3ded04a4d4e7c658ca61b80d4e7b17594be232721602",
-    "blk.7.attn_k.weight": "9ca48b87a10116fd8868e62b76f211d4bb91f166096be9061439ee2e1c3a5c20",
-    "blk.7.attn_norm.weight": "cd56cfcc4e2ad6b96e23ea7b0d32b4caf236107d99a0b22c56760b62e63c8cfd",
-    "blk.7.attn_output.weight": "7352b509a03cae2491ffc060e577d189341a0f861233f18c96f9d275dc4234bf",
-    "blk.7.attn_q.weight": "2b3791c8c008c33ddbe12bedba8191322ceea2dcce5cf0eb7a93d40ad254e672",
-    "blk.7.attn_v.weight": "3ae721d52466487a3d48150581e57f6d64ea1e83ab929f23b28c3d777422eeb6",
-    "blk.7.ffn_down.weight": "3b6fa8ececdb3c34af3a5363863d6f94289c1c95bf47fce3a3ddcf184c5f0848",
-    "blk.7.ffn_gate.weight": "dbd7df6c5ae5eb4adb859f0d36453813a4e289a359a1ba8f72d67fcbf21c3e22",
-    "blk.7.ffn_up.weight": "de68380a334b4c5cfd4c318b0e9854aec59bd79aa0f0c30af3f56414f83482b0",
-    "blk.8.attn_k.weight": "7303c4e4480abc72a7ee271811311199245fb5c2ea27a2bd3b8cad3a53a03c27",
-    "blk.8.attn_norm.weight": "2e3d1921898d1b943ce1a1b6818546c8b471d6d542da24f51a8b514b8c3dd4ef",
-    "blk.8.attn_output.weight": "30421520887b66bf97a18dbcdc283bc8d0b60590b612fd638a319a6eae923227",
-    "blk.8.attn_q.weight": "73e064d5433c9b500068a1c31744dbd53f4ade298fb450a0e8c97f62cf1f8a8d",
-    "blk.8.attn_v.weight": "27e21f8b9a9a8533e8178ca34a72aa1d786393d57302b7806dcdf3e51de511a8",
-    "blk.8.ffn_down.weight": "bf694bd8e00047982108000e7b3dee7b225db8b19abc595e5697b6bbefd92e7c",
-    "blk.8.ffn_gate.weight": "d55fdbf8606d9141b774b0500c58944fd1253b9e69d1f765eaa9a680b9f2ca40",
-    "blk.8.ffn_up.weight": "1ae3f580655e7c8e8dd6c34fa4ac574fdfc5e3f1a8536da0c5442d3a2976f0e7",
-    "blk.9.attn_k.weight": "b18080626012d8aabcf78542d6c7bf31c712bf55a70172fbfe173fcf34481036",
-    "blk.9.attn_norm.weight": "2e3620620dc09998c6d3063a7d5de5433fbbae8c11e5b00d13f145d39140e162",
-    "blk.9.attn_output.weight": "69c3c0e27ef1c0fc933eeb7b612b70909f18cde238873c0d576a2ba9714ef174",
-    "blk.9.attn_q.weight": "68330e5aa28a28873c9a6e67f032186ef651df2df5844e0f27094ba349fbe4ab",
-    "blk.9.attn_v.weight": "3df8d45a102be082d0793a51cb82aa62a43cd0e9d047ba4115ca0f2414b39325",
-    "blk.9.ffn_down.weight": "1d6cc162b73745b135b4f040a0aac3c06d5135a3dc5b2421e7ee2af48662fd7f",
-    "blk.9.ffn_gate.weight": "034a9d40fb1e32b534b45f4bccd65cbe43c4a6a3f5d01132bd245ca0005de5fc",
-    "blk.9.ffn_up.weight": "c838c38d0e1a0ac0da17eb2a66023ed31929f07d8fcfe1cc546df26096c91f0c",
-    "blk.10.attn_k.weight": "a78507cb72f744b86ceaa032596e74e5571c822d0226d334881169addb32cbd5",
-    "blk.10.attn_norm.weight": "35f48d0b28ee0e6b4cad4e983925737562d64824be5b168b3e26df3d6b260cf1",
-    "blk.10.attn_output.weight": "53712db06796de39b131323e7abf9a58551b6d52da6db66a471580386d396252",
-    "blk.10.attn_q.weight": "efe08429ba196026b81cd1c471e1c7418afd9e966659feb3936b674aa0803b58",
-    "blk.10.attn_v.weight": "7ec6055e134f89da0cbe79ec9f13ef2e442ac584b1f03c3e13e7d0cdad0078bd",
-    "blk.10.ffn_down.weight": "37e66af4bcd1f3079e841e892255b8255070655901864ea3a8c602a7f681a640",
-    "blk.10.ffn_gate.weight": "1825282bc34830d371c6edcc3c1e73e6ecc1e10f4aea0122dbb7acc1d6f7b1bc",
-    "blk.10.ffn_up.weight": "819b3b276a4d4c14a35ed6682d5ef18a5e8ed468e5ce3f12e8c75ec18ac20ec4",
-    "blk.11.attn_k.weight": "5327e6a2af82dfff0619a14971f5864a15553c36fead84e1af42c7630f2729c6",
-    "blk.11.attn_norm.weight": "fec363b3c4a43036d2c635fb8aa9e122dd87ee79811839f2f6cd955be3373e7b",
-    "blk.11.attn_output.weight": "ccf7b38f18ee8798b8a6a35018e2df3eb3e007de62876befb68025dd66c79763",
-    "blk.11.attn_q.weight": "da8c4a1c824ffe174e39f126cd72f7ef83c56aff1259d452a1212de80f98f5e9",
-    "blk.11.attn_v.weight": "d17ae6bb77f03982b55d341eb67acb5969e9ad3da5994b96eafc09793dcfe3a0",
-    "blk.11.ffn_down.weight": "a6bac521e2791345f22c57205fa1c2f2f687794dfd24d0e98d50ae0d0eb6088a",
-    "blk.11.ffn_gate.weight": "5ed902c488cb51ba5635f3df08258c5f84f31a679a00211ea5f9d8b824ef6d9d",
-    "blk.11.ffn_up.weight": "ee9f1437eb890d2cf9df2574afa1cecf20aafdd847cd75b152d7eb74419afd34",
-    "blk.12.attn_k.weight": "5a069c06e1019b0f889088e67458f7a11ec77fa190ada6069e46211f62219947",
-    "blk.12.attn_norm.weight": "194d7e5fcc8c49aea62daf1940532419cf3c505afdce6be377286b677db5db8f",
-    "blk.12.attn_output.weight": "6534995fd4d6fecb55e317add4b1723aba4d825e1e9471d0b08813dfdc247176",
-    "blk.12.attn_q.weight": "4ab51ca519b5995581fa34f846276feca3b907ef2b51f192f6cc0b3263c3f5a2",
-    "blk.12.attn_v.weight": "5652ca3fa81ef9a1ac1543d71fc6813f8517f8ec54b25c701f6f98061614830f",
-    "blk.12.ffn_down.weight": "4b2c263f54c88516b8eb273bb8d9615b01c5c8b484dc70358adb91b50b300edd",
-    "blk.12.ffn_gate.weight": "8f50c3c3e3e8568991d6c1b0e74b500cf4f208e7700bbb8e87c3f6a6d359b6b5",
-    "blk.12.ffn_up.weight": "1c1a581fec1fbe959e1427fa513f400100b5e1ee9d83932630be9905fb49c231",
-    "blk.13.attn_k.weight": "efd7a38c46f08d8376d82974f33c644e3a02220e142d63b1704718699a8a884c",
-    "blk.13.attn_norm.weight": "d28fa4f1bd75abbd063b0e622e08f579c89cd0c0c5ce63c1952ec9f944f8ee13",
-    "blk.13.attn_output.weight": "71e0068a639288718bdb70a6cfdefd50bc8b3ec3993347a65129e70001ca5827",
-    "blk.13.attn_q.weight": "b97077adc92cff07a2e07d80ee38f214ad8713571c69cd5c70ebd43dc501ac87",
-    "blk.13.attn_v.weight": "79b3e2749ab4b459c81e96e322b215f1e8af645eb346e176c326bd00cf6ed2fd",
-    "blk.13.ffn_down.weight": "9f8687d11effa1db7cfecf7bec5631734bcf2962aad74a9f519144491e08ec85",
-    "blk.13.ffn_gate.weight": "7d14dfa0543852e7777fe8fff29ca533744cbcf1ebcf10067e5adfc4eb345e65",
-    "blk.13.ffn_up.weight": "852b9527b97fdab211ff3f832a660ee1d93ccb56906144c50f01319a6e8ee615",
-    "blk.14.attn_k.weight": "79e926b20f36f66d58226cb358881f2f68ae7b468787d33cafae5110287a14a0",
-    "blk.14.attn_norm.weight": "97d481b63deb0df6142c2c6cd23043720c62eb609e390f47a7113751c79974ec",
-    "blk.14.attn_output.weight": "aa6e94d7176d5c79fbb89b96e5f13ce75702ce3dd23ee52986446da436a6c3d6",
-    "blk.14.attn_q.weight": "214becb6d1bb460da9fb8ace0f99b9a5afa9edf7aa7acc19606c7401b11d6305",
-    "blk.14.attn_v.weight": "488b0e6d7f1a7a2ed0972aaa6d10ef9c775ee5373460324efcf5b3e3da9311df",
-    "blk.14.ffn_down.weight": "29c7ad16cf9542e30996a1a01ab95b844533b28051f04cc7949c371afb796471",
-    "blk.14.ffn_gate.weight": "b7ef208f2b054803665b377f5a5980c122c026841809cf855c6ba06d1c3a885a",
-    "blk.14.ffn_up.weight": "76a5cc28100748d79c4398ce7b9176aab4d661548b6293a82f99144812e5b70e",
-    "blk.15.attn_k.weight": "a6b8f9e98ab878fa7ebc5d080978ebf2d050acc2ab2fa8ea9188eb10e27702c8",
-    "blk.15.attn_norm.weight": "a26d07a9752d6dccb68e3a8a2a49fd0752cdd0a415e05547819bc37d9ba63d5e",
-    "blk.15.attn_output.weight": "c63616c69048ccbee801e05be4f56d21fda21aa0cc470f41d57c31b4d9283a4d",
-    "blk.15.attn_q.weight": "fd595a67bf96c6ba16eb148a9d02fa52fa3c1d33ed10be28a08f851409fd6e64",
-    "blk.15.attn_v.weight": "1c5c9d33fa07c05d5f4ed0032c6c4aa83d863f0d31c94a66109d239dcd03cea3",
-    "blk.15.ffn_down.weight": "585ea62ab8aff7d7d212ea5c1a03226fda6b68370c890b776834af70c948dcbc",
-    "blk.15.ffn_gate.weight": "a13c63f86f879b03a573d5dd2a25cfd1f4dc73e8132e6454ecc23e538b4cdf6f",
-    "blk.15.ffn_up.weight": "f7112450f57c12fcd511f049e0dc0b541625a107a7901c3261ed9e984299f65c",
-    "blk.16.attn_k.weight": "2d2c8b11dd71fba6d1c106aa1673c113a5448653cca7eab897c8739212ed5003",
-    "blk.16.attn_norm.weight": "95c2ec7be9469690e18a9a1779684acb3e9da44b13e263a0da840305646fbf8a",
-    "blk.16.attn_output.weight": "31a65046e677f54dae654ded4e733479fcc0f7283d83076b7dc7cbcae8528230",
-    "blk.16.attn_q.weight": "bfc6292b9c6d49b7118d08060242a138182eb182d136ba5dfaf469437c16081d",
-    "blk.16.attn_v.weight": "68f81d037340217d87c7853ff4d6edfbc46d9e827ee6d5bff7c3f6238e3a95ad",
-    "blk.16.ffn_down.weight": "bbd6629691950cef4d5113e1c6670e91b216a9b872cb92cee02dfda4d6c4f7b8",
-    "blk.16.ffn_gate.weight": "63cb56f282b7401ed6c76e5bb6fdf1bf68a64f9af0c82c014209b55bcb5191d0",
-    "blk.16.ffn_up.weight": "b54f39a2541063cbfb6f713aa81c3b69a04100e999aa2ebbeec195dc382eceec",
-    "blk.17.attn_k.weight": "3d9ba49799cc56664ec30a002bcad61eb651294212a68c3ddb573eb042aef5a4",
-    "blk.17.attn_norm.weight": "42ee0db4b9d63257bca0012a30b12737ead1caafeb5ed3d93c8f48ffec4b46de",
-    "blk.17.attn_output.weight": "a38fd100f05c9041c592bc739e287de0b10d08ef2bda41a879225bdca9002f71",
-    "blk.17.attn_q.weight": "8a3bee285b0180a9eb35662e449ee4cbe16d992bdd48fb3a94bc4a347728cfa2",
-    "blk.17.attn_v.weight": "d7f8f1b8b863494ed4392a1656775912e9b264ad36016547b12e832a1d6757d6",
-    "blk.17.ffn_down.weight": "bb7ee58f61da8630972e25b621996fbe8ec06f4dc9ab1e268ab5b120c526ca28",
-    "blk.17.ffn_gate.weight": "6b652dbf167fee09a45ebfd78d500ff6548fb2756dbe5343ffec3f7e6207179f",
-    "blk.17.ffn_up.weight": "3b67f727e55e742715de978fab80457781e7a3762bc48f79d13b45dcb8de664c",
-    "blk.18.attn_k.weight": "ff7fe57c57b90c6fcc0aefc39ec24593c3a7d1ea1c23770480075a015450e0f5",
-    "blk.18.attn_norm.weight": "1d40faca082d2633ef0ccf19e121870dd6c7c3e2154607c7f3543fa96e99cb2d",
-    "blk.18.attn_output.weight": "9adfecaaa397a92db4687efd5fcabfa0daef9e6b0493763b7ff5ebc185c43a6c",
-    "blk.18.attn_q.weight": "ad1803eb9b291948639277afe981e666b07167eb3fcae903ba5b73bf86d8f50b",
-    "blk.18.attn_v.weight": "308cf23399adccf27401a4ab60d74dac6fb9d4cd4b9c5940d9145118d1881b34",
-    "blk.18.ffn_down.weight": "7de4ac9a561fb580619b745687dfd7ca8a69ef70471dee978741b80e9ff7bead",
-    "blk.18.ffn_gate.weight": "0c66970f696b33bd5ee8f1f2fbcb41fd78fa5ccabdc927e11a4d5a4089f19c69",
-    "blk.18.ffn_up.weight": "66a42e988e8a1f468fabf976c48e9e4bb045eaac6916ef16555ac101cd674abc",
-    "blk.19.attn_k.weight": "a928ab50390bacbcebe2e4b66922498134ce22d7b93beaa87d6cf4ab52eb7174",
-    "blk.19.attn_norm.weight": "b4a02c55b46c2a96aec9c64a254087cf48e6c1d4b6f31782c77a46fc4daebad1",
-    "blk.19.attn_output.weight": "b768319c641dff1eac5d1f8ceb960c9899c795bf2b24c1d6bf70aa24fda45f77",
-    "blk.19.attn_q.weight": "79ef3f57d187d3954a26362096e1b6c222d76f537dff73e034d6e9999935b8bc",
-    "blk.19.attn_v.weight": "ce13d6b13e24fcb2d5bc6a2662e5bd295b31b12db10a6d0307f86cf29b8d5001",
-    "blk.19.ffn_down.weight": "cf90d7e2137482cfd50934a8223ad774621d08554969da80a9712df5e6227eb0",
-    "blk.19.ffn_gate.weight": "71ce30150f003b6eeb3bf7464e05b6ae615f135110d8e47f0a47fd973e537c0f",
-    "blk.19.ffn_up.weight": "7f92aca0cc29866633feec701ec01a85a8ee2fd4e2b9630173a6cffb1d9d50ee",
-    "blk.20.attn_k.weight": "a2df23159d6fb74ef28e14b61028fe8b00a693a2fc9234a980be74f20b958682",
-    "blk.20.attn_norm.weight": "c6cd5f1b096fc5efa4eb59ca1c8c4bd28730f3dcedd59a63601663eccc6724ed",
-    "blk.20.attn_output.weight": "896a8a166d0f006d4b09867ae4345426303cbc3fb13a18d3d4e1bde00f16dbdf",
-    "blk.20.attn_q.weight": "01eb79588fe61baea0da43e99f4dc5939590e1bafd01e12dadb8326f102bfea2",
-    "blk.20.attn_v.weight": "bd39630fdd5a7c859ac1addaf53e63faf524c3f32f5f4896d86b6e746b1d5c06",
-    "blk.20.ffn_down.weight": "0304a5d39957a0e3f031c4bcc4549a135d396c8d97c8d276fd1c823ce86560c2",
-    "blk.20.ffn_gate.weight": "117b79d595b1dca0c8b37586beaecc4d84411507276212dc286cde7fc36c9bef",
-    "blk.20.ffn_up.weight": "6e799346db145c125f01783539749d3828fcc451cd4f10c5352f047a47e28714",
-    "blk.21.attn_k.weight": "1c37e4c0664147e775bb006b226b9553e3421140cd96288ea755f81731ab80ba",
-    "blk.21.attn_norm.weight": "00ae783a29000ccda5e4bdbff03df0752fb82805dc3f9b987500ebd80714476e",
-    "blk.21.attn_output.weight": "7588b84f9fb19f15095b5265c60b4a4e7ae74bcc47d4607dfa5d0bfab6f136cb",
-    "blk.21.attn_q.weight": "a65f1c0dd06d45bb97532d3e932689c1eecfe7359089b39174a96a149335cbc1",
-    "blk.21.attn_v.weight": "4220b77e7d5e8709b4eef33a679b5dad11f297085ef44c9977f9e54ef08f7a2d",
-    "blk.21.ffn_down.weight": "b8c082a0530d4b5328e67db0df84c5498f2af956de23c639fa0198ffea853950",
-    "blk.21.ffn_gate.weight": "cd1b656ee72d00e9835ef667c19ef89a88de261eb8eb7c0e936e0f9ddf83ef9f",
-    "blk.21.ffn_up.weight": "dc445f73e36ec7a3bd86884186b728f8e0187f32848c3b8b69d4d41f8571bf31",
-    "blk.22.attn_k.weight": "e37cf0b893ec8b9ee8c78dd139b8d9c45cb997a3bc0c3d93a70ca1c3f6af8859",
-    "blk.22.attn_norm.weight": "248a27838d3c46cc03a5c312facc84e2e0e2c990ef8401e93da25918497f88d1",
-    "blk.22.attn_output.weight": "fc191a18f6d18332c66761f7ab28008bfe295dd1f5c8741a2488442f9e00d0f5",
-    "blk.22.attn_q.weight": "4b193a2ab8bc2b085db18f2bf3eeba26e02b537b2cdd738160c8f14b165d0f5a",
-    "blk.22.attn_v.weight": "7a60ce5ccac7e045e55ba1e1e85bd2a0f93f8c781daee96c5223665e22f0c666",
-    "blk.22.ffn_down.weight": "e0a34fb4244e2c7168f3dbaa1904c15d339ec39999cdf27128bbaf619ee0a237",
-    "blk.22.ffn_gate.weight": "8bac872d4b8549c8812f927efa309f1792b524f33601095fff61b826de5a5615",
-    "blk.22.ffn_up.weight": "b67fa2b94dd901b6ec64c0853ce8ca2d86fe9cb1cc6d2f15fbbbe0e691c0c648",
-    "blk.23.attn_k.weight": "2c32e66ad01942b819ac09a197c71579fe66f02226a264fdd72ad1e02c67a27e",
-    "blk.23.attn_norm.weight": "825fdc94deb439cb93c713eeb077c1052b90ed658d6d464fc4ad3d611e911d48",
-    "blk.23.attn_output.weight": "95ca6707a95b8750b0c7c5d379d368f0f2e7ebef631954e7d4d8ec0f41f13a3a",
-    "blk.23.attn_q.weight": "6eccc84faca5fac015d1b26e2854501edcfd292a302228fe14cf99f5eb59a34b",
-    "blk.23.attn_v.weight": "b343ac3d226040f1033ee049668aa1d89b1774bc18431965682e5dbdce78ccdc",
-    "blk.23.ffn_down.weight": "9fc599befea8d3b1e342d564a110074f66d2542df406c4b90b6bdc5828fbb2b2",
-    "blk.23.ffn_gate.weight": "488556c1b0c9f0b20b0c99b4bac2e0f4046b81edb601d7b91e7e5b3bab47d667",
-    "blk.23.ffn_up.weight": "1088e291d7008dd9c7c2dd6830af686a8a84b724d123a016209bd5156d6898f1",
-    "blk.24.attn_k.weight": "a923fbe35e61e009a53927d7828818e0592bb737d6a1106c4b0b5a1efc367e07",
-    "blk.24.attn_norm.weight": "9b51aaaa939cefafdd9b13a7e5b74ac7fa2d603427e55a16a909d6f3f353750a",
-    "blk.24.attn_output.weight": "1beb2baba56f8409466434b037771248c2f620ec5f53e15f44c271d5a2d9ecf4",
-    "blk.24.attn_q.weight": "4b0194fe5bfae0c6bf6131dcf8cb6e2b994f6ea10b27cb03574f0f4f8cc0c950",
-    "blk.24.attn_v.weight": "6ac34b1ab0f66226d85bca1194a7c212cd93d384ecbc8b8395de48aec0970a61",
-    "blk.24.ffn_down.weight": "5508f74cb732a662c2936b32ac5e90742d172b9f961a747b0e5cba0e5906a89d",
-    "blk.24.ffn_gate.weight": "095e39b8584403835f9bb1ac33e0e81f54175575e4800273d281b845bff381e7",
-    "blk.24.ffn_up.weight": "2d43ec21637dda12973de367b0113ee9840b0d815bf6fce042f7c3f270b0b530",
-    "blk.25.attn_k.weight": "9e2aee029f3d2c7f67dfc7926e72c8228fb978382c8e5a4701bbf82c93801419",
-    "blk.25.attn_norm.weight": "220cd7164fb4cdbe22d26058e4153b26c27c7b5ce2bec8e95bf2c0ea08d23103",
-    "blk.25.attn_output.weight": "a17f4a5dc6aa51f03dbd75602d98e9491767c205cdc2c3a5f8667fc54bbf7c64",
-    "blk.25.attn_q.weight": "f60827496835c440c794bf57ce9780704d10a59d8229886bf75ebb18900ba4ef",
-    "blk.25.attn_v.weight": "9cac217e9e9f4f4c85f14ee51165a77c580165bd4a34b202389169bbe61a1ced",
-    "blk.25.ffn_down.weight": "a0f36949b663e80849581dfb71e7babcc73580793bbcb0c80ab26d5a6e000359",
-    "blk.25.ffn_gate.weight": "df4d1be4d50d6afe5ad3ef0d0e0fac76a33e85c963dea769641d612dd53e7d13",
-    "blk.25.ffn_up.weight": "992da76be762632e25ebc5ef4d03728eece1b43f7c4e31827df19ca724aea694",
-    "blk.26.attn_k.weight": "34199ff856ac32a500c754539d070258574192a34ecba87a182897cb59fdff52",
-    "blk.26.attn_norm.weight": "a8e9dfb2dae5d22b5c0aec5f3675991c0e3c3e6a44153db2579136b73f456e00",
-    "blk.26.attn_output.weight": "1c4f257ffb0d7db0f11cfb275e38b4af736917b43ad82de1badce3f1d227da4d",
-    "blk.26.attn_q.weight": "33d55786274c2e718cf61e8fbecf3dfa5ee0c208f0b716d42b061f55459acb3c",
-    "blk.26.attn_v.weight": "684b636939cd4ffcfec5a6238a0790ffa43d853c95783af9b9e8275e74071a7a",
-    "blk.26.ffn_down.weight": "89d0bf066db154e6d312b5433aed1714f6a28b40f4c52e3e1530ee07703303c8",
-    "blk.26.ffn_gate.weight": "393d649bebe5e2940e1b043649f6c860b4b8b9f380f30e9da1744a830f358156",
-    "blk.26.ffn_up.weight": "179edc85ababd9d8440cc6093eecd1004290aa1cb96434b26ecf7585b6cca17b",
-    "blk.27.attn_k.weight": "334841445a7f1e14731b08f56eb0b1f0938c63823d28bc6d078c4c5f05b36f19",
-    "blk.27.attn_norm.weight": "57344471bbda2e9deffdfdb2dd05a07aa47f8761e24de53525588639145bf551",
-    "blk.27.attn_output.weight": "506126af9ee54b535d49f97e36f630e74834f480329f098d6d62e96246d8d65a",
-    "blk.27.attn_q.weight": "dd984df1acb4783849e25ba7ae378bfd385cd9efc540fb798cd5bdd873f0118f",
-    "blk.27.attn_v.weight": "b4b3fe9a4455d34c297ff20a2f537b647cef424741d840a747b265f23d320ac0",
-    "blk.27.ffn_down.weight": "621fdb185ba0d35ba5476dae73d2c81ec1482a0e878d5bfd5c3b29fe837af013",
-    "blk.27.ffn_gate.weight": "e4fbab45f2ec506fa374103251a0bdb7baa6f576080bdd796f3e9db92098e08f",
-    "blk.27.ffn_up.weight": "a0c57e463e988002bbd6a6c6792baa21a65e6f89ae303a2c301951b0ae6e4bbe",
-    "blk.28.attn_k.weight": "bac36cbd52ec5056841663865e1291ddab4b47ef9a2544dd285d4503bfb0e4a0",
-    "blk.28.attn_norm.weight": "5774a9df2bbb2e86d1f70179c7b92d81e1f401160148b3328fb64db6646a5425",
-    "blk.28.attn_output.weight": "e8712622d1569557000c75f26c3f55fad267fd300463c2c2cfe3afbfa1c8f908",
-    "blk.28.attn_q.weight": "11677751fddee52cc739699c02836f7be54d96038be4240be5d4f53d00161608",
-    "blk.28.attn_v.weight": "e5ee459b8958d65e1445997b9aa1e90e2f5d17761ebcf5357313119a45322507",
-    "blk.28.ffn_down.weight": "3934518f9f85292da8475fe38a8edcbfc4e24ac56c351b472d6351f98750871e",
-    "blk.28.ffn_gate.weight": "6ba735d57e98d0847e487f25ffaa25256deaa8abec76f428cb70bd9774279d83",
-    "blk.28.ffn_up.weight": "977fae6e1e5353114fc645dd98429464749758765cbc6e6457593d596e57850c",
-    "blk.29.attn_k.weight": "8122a457307d580ad6f1e0acea09a2f593d97f595ba0d6737f5fea16d2433642",
-    "blk.29.attn_norm.weight": "d626f721e05aa1202439b01027031d4caf1adace61ed37870a277cb6297c77cc",
-    "blk.29.attn_output.weight": "7fb7122ab1b6b1e6615ca746897da27bc52c92cb70d3147183cdde61795b72b3",
-    "blk.29.attn_q.weight": "be43e94ff6b6e391024dc824101efa0ddf4005d5b002ac26cb03765c0c73c2fa",
-    "blk.29.attn_v.weight": "af93c85ebff908f74f9935b81bde0516ca487c84139868a1ce079c3ae20036b1",
-    "blk.29.ffn_down.weight": "39dae12340ed3120bd19c495fe0872b559613641e41fde69d02d8631900b84c0",
-    "blk.29.ffn_gate.weight": "36fd482439840ef197c9f3b8905d86acfcea49bcf018544106ca465d4bf8d5c7",
-    "blk.29.ffn_up.weight": "5243fbdfdc1e2a1dd84b6210a9869d18a014db9088897e345240cdc99990bd5d",
-    "blk.30.attn_k.weight": "948f263616bd3788b2b968baafd69b9c5bd1b77578665f096c4b7e247b4cea42",
-    "blk.30.attn_norm.weight": "e168df981e744874ff303faf2eb470e5f6868c2040ba5f383f6c5148669975e7",
-    "blk.30.attn_output.weight": "4cf0ccca04b792573b756655a24fc89cfb1f272da8305633f0bc66ef14990b93",
-    "blk.30.attn_q.weight": "21e07d6cba6c50d65350289258209717174a13c42be57e8141d69712cbaf32c1",
-    "blk.30.attn_v.weight": "65a8ca29c7237b3182ccf03e2fc94e84f9a53d0e160fb679ab401c853170dd9c",
-    "blk.30.ffn_down.weight": "8b00500a6d00d84058f6658ee1d6f06fb4fcae2f90d4341792259362923b3c13",
-    "blk.30.ffn_gate.weight": "5bc0e19ab7a31b50ac2118ad1b36e31055271a322cd8ff661d47c3ac0210703c",
-    "blk.30.ffn_up.weight": "f37a0561955725bd59ee2d064fa9f4e00a12a1b620b624db3bc3add5330bc321",
-    "blk.31.attn_k.weight": "9a5663edda227f5d87533897146764f8e8a7481b9e71fae197c39204f8463221",
-    "blk.31.attn_norm.weight": "060a4f438a1ee5e220b5b5278ad2f5c085a428bf38c515766781815597c87529",
-    "blk.31.attn_output.weight": "6ada5d3cad9dea4780ffbb43302bb6ccc2f24eddd0fc4f5f84c9ce0fc0c6e5dd",
-    "blk.31.attn_q.weight": "bb5d08c08603907981ad388d5d8b70fcc9b98034ba264b8474c8890cc0297af0",
-    "blk.31.attn_v.weight": "e01b4252ea9c6a889c32b21144b441a347464d04536ef4f6572425be55759796",
-    "blk.31.ffn_down.weight": "8ba4d679c36e93ba65ba03180385ef35ea86b3b7cdf2fded9df59369f1c09630",
-    "blk.31.ffn_gate.weight": "e5b41dc93645f8b5e8eebae3ada3ea43a18f97ce2654228655170b07b463ccb0",
-    "blk.31.ffn_up.weight": "25b88cdddc8b547af294ed107d3d1312e90b983cae87936fa6062ecd8ea02539",
-    "blk.32.attn_k.weight": "4bcf86dc0858c8ca2fbdf6aa76674d43eb698f78979fdc1a38f556a7af1facc4",
-    "blk.32.attn_norm.weight": "cdcc12f3b8b9773c6722736bfb748a2729230b21478cbcc4104859d3148df815",
-    "blk.32.attn_output.weight": "d43f1196822995ed89a9365c97054753a8b30ce20b6e273c8edcc42673a1e141",
-    "blk.32.attn_q.weight": "ebf2972bb3865cbc5be4840113a322089752038344beab2a0122c7cb4fb399b6",
-    "blk.32.attn_v.weight": "714db81704ff34fa137512903c1013acee7877467473e46600728b9240582eb7",
-    "blk.32.ffn_down.weight": "2cde3da1258bb170a79d5d3cdfe10c86a71eb34b77da46b74c5ed71e7f4fe274",
-    "blk.32.ffn_gate.weight": "c7e1ed792532613ff9d4e5834b6536e2e0f47df2303bc0fdaa90aac0c1f4e8db",
-    "blk.32.ffn_up.weight": "d8d6f13fe66a716e28f79101a29817f0c0d6f99969a6f017d51bafd1a16c600c",
-    "blk.33.attn_k.weight": "a0a28f6cbca88da00cab2ca37094d9b0503bf9defdae77b91895b911c408cbb6",
-    "blk.33.attn_norm.weight": "0251200c24cc8445607ace6dc8c5aa0566567997262b7cca53a11ac23cc564b2",
-    "blk.33.attn_output.weight": "b2423205bdf6a1096d43c44d8d12f1a84fcd4e1bb70fcf6dc8542b8b8a71a13c",
-    "blk.33.attn_q.weight": "00b425c3ef71065ce5e0234e702bf38143b4952da78a85f52ab2c2e3073d97ab",
-    "blk.33.attn_v.weight": "035edd2335df816c42c765a5e66b9d9b9e15a822a8dc1863508145499c942c14",
-    "blk.33.ffn_down.weight": "4894a923a3db75bae4496ba3ce5f28796ad31fe33996a066271fb8654964310e",
-    "blk.33.ffn_gate.weight": "8f6c819b8bbfbe3357fae89e1ac5a3d58be85b3b04be3bacf7b62775869046ff",
-    "blk.33.ffn_up.weight": "257c3544b5b544fd5d839665bf5caf107a329b59dbc3751efcaa24ae63c56179",
-    "blk.34.attn_k.weight": "b6cd8bba892e38dac4a2ebc3ba1bce49e71b967fc436fde30c6d76f54a18935f",
-    "blk.34.attn_norm.weight": "2b3c8e60a064cba9955752bbbbdd92c71ba5c2f1bd721097bdbe88b5abc68787",
-    "blk.34.attn_output.weight": "8cc272551c9aaca9db5a660c6927bab94a0243d74a30b2bc165f06bd577714ea",
-    "blk.34.attn_q.weight": "74b561eb4792484e6a94b58fe2583848c3ae28ff2f1bf3d02939a0cfdfa49990",
-    "blk.34.attn_v.weight": "dba19e24ff05154dc5a1f55c023729303a583d13d68732ce22ea74d4410dc8f0",
-    "blk.34.ffn_down.weight": "76eca5dfeb274c35774e0bf9f22ee420ed9085c8e99aa2cd5a236e4918b44c61",
-    "blk.34.ffn_gate.weight": "9af0862d5fcbc24732846488e653db8242a467765c0cdbc00332b3a40256b4a6",
-    "blk.34.ffn_up.weight": "2a03126bf73587eaba99ece2066103d12e47bcd4ce30ff6c17b2f383b81d40df",
-    "blk.35.attn_k.weight": "52513fc0cd4e997a842729af7d21dd09399bce0a339558374738be266d0fa2f0",
-    "blk.35.attn_norm.weight": "e5281fa911964263ccf1630b14762edbd41d0b9472d6ec695fc600fed4892c35",
-    "blk.35.attn_output.weight": "b391d6705d5dc6f48326b5fd16573f679edf64109d86fb729a498819676590ca",
-    "blk.35.attn_q.weight": "d16446921966db9b0e0539626ad22a2511ace780e59379d6a4162d8c5441440b",
-    "blk.35.attn_v.weight": "9d8cdf23ffdb0c5c74106843390b94b24c9f33ef0eb9998d39f78c73390101ea",
-    "blk.35.ffn_down.weight": "938eb6301f7bbf162d7dd965682a5ed11d0a4a530c6fedd7e5469ce80012fc17",
-    "blk.35.ffn_gate.weight": "5ad84f5a0c8edcfea1ecf1a3e3d21d85ceda0c4ad9e3c6ca68885eeff8ed3c2f",
-    "blk.35.ffn_up.weight": "1c4330d9dc71bf4c98812c34356c51f520f47610a534152aa6d29284b758090d",
-    "blk.36.attn_k.weight": "ef720655e5ca2465f13db2dfc4732fb4ef2c9d53acde52f514fd4f301e974081",
-    "blk.36.attn_norm.weight": "88f4b9310b3c8c2644e3029160cd35678c79dfa59280430e03f5c29a6fe84a58",
-    "blk.36.attn_output.weight": "aec6f915fffd7bb72cd783273e871b4f09605950089d45e72059d1316b6c4b01",
-    "blk.36.attn_q.weight": "72f9408a2405d42f8db6ce5fcf1d26a3660b6f225fc60e77d0277109cfcb82ed",
-    "blk.36.attn_v.weight": "0f3b3d851dc44b3893ef53f6cca5b4acc9658bacfe1cc2d13c3d704ddd409b67",
-    "blk.36.ffn_down.weight": "470aec48ce8c5129a6654d9fd26fcae72776f9fc1429a8bb05818072a876475d",
-    "blk.36.ffn_gate.weight": "7f5f296d09cf55679767b5d15de3eff489c456782119f25204be4b1647f18dcf",
-    "blk.36.ffn_up.weight": "b7ef74a1f7ffb4982711d93f1787be3a70edc3d2358d5203c41d8900508037d4",
-    "blk.37.attn_k.weight": "c4ffa5412e4ff2dcfe1aed991c1f54169fd171a4c7638e4b9f21a1ca64c5e1d6",
-    "blk.37.attn_norm.weight": "4eb6c888d841cccfacf5b963f8611120f6ff24b84af0b5714fd9ab36dcda422f",
-    "blk.37.attn_output.weight": "db2a7bbf9682f9f6eea672dae8e150738f1bf74dbc80edc7022017a3f040c8ac",
-    "blk.37.attn_q.weight": "e38c0462aff139afcbab289189823527e453abc9e541154adde5e7af88cacf0b",
-    "blk.37.attn_v.weight": "952eb2492ed452a72f96bcc12d4b2affad9dfdf46ee39ce4a5d7b57a5dc301e5",
-    "blk.37.ffn_down.weight": "25f23a8fbc44febf6dc4848fd7fe03a580e2822bd3b3b5a51f4990826bfe3e4e",
-    "blk.37.ffn_gate.weight": "707da5eb40118b035305d3262444382351f170a20a537386a70e90c5a83a7817",
-    "blk.37.ffn_up.weight": "d2d2ba5cfc4ef47338dd7384219e22bf030a5a2209e0354d88f5bbaaafd20e87",
-    "blk.38.attn_k.weight": "abc4bb189dedf7ce661e79028427623a4f91ac091c2cd60e31b58bc62b1cda71",
-    "blk.38.attn_norm.weight": "9f4803a7d03fd40fcb83d85f84eb1d5682ea4e5bb084f210c02850675d804c3d",
-    "blk.38.attn_output.weight": "77cb66007f1a41df7135d0e7f900ceb499c2f667dfc3f1a6ac01a3203bbd3ccf",
-    "blk.38.attn_q.weight": "d94a8b26cd375bf2bcaa76597e314aa8268ee50a479d00931e5e0e021feadb5d",
-    "blk.38.attn_v.weight": "660c907888bc5016dc69b7d35fe6f55c7ded697c93be0e2d332a2f17aff88758",
-    "blk.38.ffn_down.weight": "6f06173bae5b00ffaf88ef383619a8b9c6a8d0d5c6494695d17f6c1de1a68a13",
-    "blk.38.ffn_gate.weight": "89f99be149d03f116527bfcabe073c50001c874de40fb6e817f6619027f3cd05",
-    "blk.38.ffn_up.weight": "8d57557c8d5e2d2688b73f01dddf1ce8d5194990cda6358153320aea88aac7f8",
-    "blk.39.attn_k.weight": "21be09c988b46c8393e6c2ec9230f3b5136eb7607dd1953ba92d0811c2f0dd75",
-    "blk.39.attn_norm.weight": "ba7c1912dd1c4e2d16917201f62396fd0600e4a451137eaddff255548c209abd",
-    "blk.39.attn_output.weight": "acfaf4abb3fd27fd899b5563c3877f176b597d8f6cdb2f2fd3f3a0bd4da15ed6",
-    "blk.39.attn_q.weight": "e8adbc140d4c8f0db2a27ca584c5531d5b1e080555fe627e34d80d0814a92bed",
-    "blk.39.attn_v.weight": "92f96b0e1f724e73a0f90a76c145654418844c04a6d4b14c05eb5af8a62bf8dc",
-    "blk.39.ffn_down.weight": "4d9ee7c65fc16fe95d10c47b79ac6a525741947600a64b5fcea5d300a82c50de",
-    "blk.39.ffn_gate.weight": "7e18507989f39b32191133d2657c2ee3b74f42f070579204d727eb72215793d1",
-    "blk.39.ffn_up.weight": "22cda752269c9757ba918abede1df95bb0f83a5c772dea13c8deea3d5f2723d9",
-    "output_norm.weight": "2858cf0e39d32caf52b7861378ace076000241e147f10b9eb21d8a5cd149e3cb"
-}
--- a/discover/amd_common.go
+++ b/discover/amd_common.go
@@ -9,6 +9,8 @@ import (
 	"path/filepath"
 	"runtime"
 	"strings"
+
+	"github.com/ollama/ollama/envconfig"
 )

 // Determine if the given ROCm lib directory is usable by checking for existence of some glob patterns
@@ -39,10 +41,13 @@ func commonAMDValidateLibDir() (string, error) {
 	// Favor our bundled version

 	// Installer payload location if we're running the installed binary
-	rocmTargetDir := filepath.Join(LibOllamaPath, "rocm")
-	if rocmLibUsable(rocmTargetDir) {
-		slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-		return rocmTargetDir, nil
+	exe, err := os.Executable()
+	if err == nil {
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), envconfig.LibRelativeToExe(), "lib", "ollama")
+		if rocmLibUsable(rocmTargetDir) {
+			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
+			return rocmTargetDir, nil
+		}
 	}

 	// Prefer explicit HIP env var
--- a/discover/amd_linux.go
+++ b/discover/amd_linux.go
@@ -77,7 +77,8 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {

 	gfxOverride := envconfig.HsaOverrideGfxVersion()
 	var supported []string
-	var libDir string
+	depPaths := LibraryDirs()
+	libDir := ""

 	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
 	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
@@ -352,8 +353,9 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				})
 				return nil, err
 			}
+			depPaths = append(depPaths, libDir)
 		}
-		gpuInfo.DependencyPath = []string{libDir}
+		gpuInfo.DependencyPath = depPaths

 		if gfxOverride == "" {
 			// Only load supported list once
--- a/discover/amd_windows.go
+++ b/discover/amd_windows.go
@@ -5,6 +5,7 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"os"
 	"path/filepath"
 	"slices"
 	"strconv"
@@ -49,13 +50,14 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		slog.Info(err.Error())
 		return nil, err
 	}
-
+	depPaths := LibraryDirs()
 	libDir, err := AMDValidateLibDir()
 	if err != nil {
 		err = fmt.Errorf("unable to verify rocm library: %w", err)
 		slog.Warn(err.Error())
 		return nil, err
 	}
+	depPaths = append(depPaths, libDir)

 	var supported []string
 	gfxOverride := envconfig.HsaOverrideGfxVersion()
@@ -111,7 +113,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				UnreliableFreeMemory: true,

 				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
-				DependencyPath: []string{libDir},
+				DependencyPath: depPaths,
 				MinimumMemory:  rocmMinimumMemory,
 				Name:           name,
 				Compute:        gfx,
@@ -162,7 +164,9 @@ func AMDValidateLibDir() (string, error) {
 	}

 	// Installer payload (if we're running from some other location)
-	rocmTargetDir := filepath.Join(LibOllamaPath, "rocm")
+	localAppData := os.Getenv("LOCALAPPDATA")
+	appDir := filepath.Join(localAppData, "Programs", "Ollama")
+	rocmTargetDir := filepath.Join(appDir, envconfig.LibRelativeToExe(), "lib", "ollama")
 	if rocmLibUsable(rocmTargetDir) {
 		slog.Debug("detected ollama installed ROCm at " + rocmTargetDir)
 		return rocmTargetDir, nil
--- a/discover/gpu.go
+++ b/discover/gpu.go
@@ -23,6 +23,7 @@ import (

 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )

 type cudaHandles struct {
@@ -100,7 +101,15 @@ func initCudaHandles() *cudaHandles {

 	// Aligned with driver, we can't carry as payloads
 	nvcudaMgmtPatterns := NvcudaGlobs
-	cudartMgmtPatterns = append(cudartMgmtPatterns, filepath.Join(LibOllamaPath, "cuda_v*", CudartMgmtName))
+
+	if runtime.GOOS == "windows" {
+		localAppData := os.Getenv("LOCALAPPDATA")
+		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", CudartMgmtName)}
+	}
+	libDirs := LibraryDirs()
+	for _, d := range libDirs {
+		cudartMgmtPatterns = append(cudartMgmtPatterns, filepath.Join(d, CudartMgmtName))
+	}
 	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)

 	if len(NvmlGlobs) > 0 {
@@ -231,7 +240,7 @@ func GetGPUInfo() GpuInfoList {
 		if err != nil {
 			slog.Warn("error looking up system memory", "error", err)
 		}
-
+		depPaths := LibraryDirs()
 		details, err := GetCPUDetails()
 		if err != nil {
 			slog.Warn("failed to lookup CPU details", "error", err)
@@ -239,9 +248,11 @@ func GetGPUInfo() GpuInfoList {
 		cpus = []CPUInfo{
 			{
 				GpuInfo: GpuInfo{
-					memInfo: mem,
-					Library: "cpu",
-					ID:      "0",
+					memInfo:        mem,
+					Library:        "cpu",
+					Variant:        runners.GetCPUCapability().String(),
+					ID:             "0",
+					DependencyPath: depPaths,
 				},
 				CPUs: details,
 			},
@@ -283,13 +294,17 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
 				variant := cudaVariant(gpuInfo)
-
-				// Start with our bundled libraries
-				if variant != "" {
-					variantPath := filepath.Join(LibOllamaPath, "cuda_"+variant)
-					if _, err := os.Stat(variantPath); err == nil {
-						// Put the variant directory first in the search path to avoid runtime linking to the wrong library
-						gpuInfo.DependencyPath = append([]string{variantPath}, gpuInfo.DependencyPath...)
+				if depPaths != nil {
+					gpuInfo.DependencyPath = depPaths
+					// Check for variant specific directory
+					if variant != "" {
+						for _, d := range depPaths {
+							if _, err := os.Stat(filepath.Join(d, "cuda_"+variant)); err == nil {
+								// Put the variant directory first in the search path to avoid runtime linking to the wrong library
+								gpuInfo.DependencyPath = append([]string{filepath.Join(d, "cuda_"+variant)}, gpuInfo.DependencyPath...)
+								break
+							}
+						}
 					}
 				}
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
@@ -361,7 +376,7 @@ func GetGPUInfo() GpuInfoList {
 						gpuInfo.FreeMemory = uint64(memInfo.free)
 						gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 						gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-						gpuInfo.DependencyPath = []string{LibOllamaPath}
+						gpuInfo.DependencyPath = depPaths
 						oneapiGPUs = append(oneapiGPUs, gpuInfo)
 					}
 				}
@@ -497,30 +512,33 @@ func GetGPUInfo() GpuInfoList {

 func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
+	var ldPaths []string
 	gpuLibPaths := []string{}
 	slog.Debug("Searching for GPU library", "name", baseLibName)

-	// search our bundled libraries first
-	patterns := []string{filepath.Join(LibOllamaPath, baseLibName)}
-
-	var ldPaths []string
-	switch runtime.GOOS {
-	case "windows":
-		ldPaths = strings.Split(os.Getenv("PATH"), string(os.PathListSeparator))
-	case "linux":
-		ldPaths = strings.Split(os.Getenv("LD_LIBRARY_PATH"), string(os.PathListSeparator))
+	// Start with our bundled libraries
+	patterns := []string{}
+	for _, d := range LibraryDirs() {
+		patterns = append(patterns, filepath.Join(d, baseLibName))
 	}

-	// then search the system's LD_LIBRARY_PATH
-	for _, p := range ldPaths {
-		p, err := filepath.Abs(p)
+	switch runtime.GOOS {
+	case "windows":
+		ldPaths = strings.Split(os.Getenv("PATH"), ";")
+	case "linux":
+		ldPaths = strings.Split(os.Getenv("LD_LIBRARY_PATH"), ":")
+	default:
+		return gpuLibPaths
+	}
+
+	// Then with whatever we find in the PATH/LD_LIBRARY_PATH
+	for _, ldPath := range ldPaths {
+		d, err := filepath.Abs(ldPath)
 		if err != nil {
 			continue
 		}
-		patterns = append(patterns, filepath.Join(p, baseLibName))
+		patterns = append(patterns, filepath.Join(d, baseLibName))
 	}
-
-	// finally, search the default patterns provided by the caller
 	patterns = append(patterns, defaultPatterns...)
 	slog.Debug("gpu library search", "globs", patterns)
 	for _, pattern := range patterns {
@@ -697,6 +715,23 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 	}
 }

+func LibraryDirs() []string {
+	// dependencies can exist wherever we found the runners (e.g. build tree for developers) and relative to the executable
+	// This can be simplified once we no longer carry runners as payloads
+	exe, err := os.Executable()
+	if err != nil {
+		slog.Warn("failed to lookup executable path", "error", err)
+		return nil
+	}
+
+	lib := filepath.Join(filepath.Dir(exe), envconfig.LibRelativeToExe(), "lib", "ollama")
+	if _, err := os.Stat(lib); err != nil {
+		return nil
+	}
+
+	return []string{lib}
+}
+
 func GetSystemInfo() SystemInfo {
 	gpus := GetGPUInfo()
 	gpuMutex.Lock()
--- a/discover/gpu_darwin.go
+++ b/discover/gpu_darwin.go
@@ -15,6 +15,7 @@ import (
 	"syscall"

 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )

 const (
@@ -27,6 +28,7 @@ func GetGPUInfo() GpuInfoList {
 		return []GpuInfo{
 			{
 				Library: "cpu",
+				Variant: runners.GetCPUCapability().String(),
 				memInfo: mem,
 			},
 		}
@@ -49,6 +51,7 @@ func GetCPUInfo() GpuInfoList {
 	return []GpuInfo{
 		{
 			Library: "cpu",
+			Variant: runners.GetCPUCapability().String(),
 			memInfo: mem,
 		},
 	}
--- a/discover/path.go
+++ b/discover/path.go
@@ -1,56 +0,0 @@
-package discover
-
-import (
-	"os"
-	"path/filepath"
-	"runtime"
-)
-
-// LibPath is a path to lookup dynamic libraries
-// in development it's usually 'build/lib/ollama'
-// in distribution builds it's 'lib/ollama' on Windows
-// '../lib/ollama' on Linux and the executable's directory on macOS
-// note: distribution builds, additional GPU-specific libraries are
-// found in subdirectories of the returned path, such as
-// 'cuda_v11', 'cuda_v12', 'rocm', etc.
-var LibOllamaPath string = func() string {
-	exe, err := os.Executable()
-	if err != nil {
-		return ""
-	}
-
-	if eval, err := filepath.EvalSymlinks(exe); err == nil {
-		exe = eval
-	}
-
-	var libPath string
-	switch runtime.GOOS {
-	case "windows":
-		libPath = filepath.Join(filepath.Dir(exe), "lib", "ollama")
-	case "linux":
-		libPath = filepath.Join(filepath.Dir(exe), "..", "lib", "ollama")
-	case "darwin":
-		libPath = filepath.Dir(exe)
-	}
-
-	cwd, err := os.Getwd()
-	if err != nil {
-		return ""
-	}
-
-	paths := []string{
-		libPath,
-
-		// build paths for development
-		filepath.Join(filepath.Dir(exe), "build", "lib", "ollama"),
-		filepath.Join(cwd, "build", "lib", "ollama"),
-	}
-
-	for _, p := range paths {
-		if _, err := os.Stat(p); err == nil {
-			return p
-		}
-	}
-
-	return filepath.Dir(exe)
-}()
--- a/discover/types.go
+++ b/discover/types.go
@@ -5,6 +5,7 @@ import (
 	"log/slog"

 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )

 type memInfo struct {
@@ -106,7 +107,7 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	for _, info := range l {
 		found := false
 		requested := info.Library
-		if info.Variant != "" {
+		if info.Variant != runners.CPUCapabilityNone.String() {
 			requested += "_" + info.Variant
 		}
 		for i, lib := range libs {
--- a/docs/README.md
+++ b/docs/README.md
@@ -2,7 +2,7 @@

 ### Getting Started
 * [Quickstart](../README.md#quickstart)
-* [Examples](./examples.md)
+* [Examples](../examples)
 * [Importing models](./import.md)
 * [Linux Documentation](./linux.md)
 * [Windows Documentation](./windows.md)
--- a/docs/api.md
+++ b/docs/api.md
@@ -31,7 +31,7 @@ Certain endpoints stream responses as JSON objects. Streaming can be disabled by

 ## Generate a completion

-```
+```shell
 POST /api/generate
 ```

@@ -306,7 +306,7 @@ curl http://localhost:11434/api/generate -d '{

 #### Response

-```json
+```
 {
  "model": "llava",
  "created_at": "2023-11-03T15:36:02.583064Z",
@@ -485,7 +485,7 @@ A single JSON object is returned:

 ## Generate a chat completion

-```
+```shell
 POST /api/chat
 ```

@@ -495,14 +495,14 @@ Generate the next message in a chat with a provided model. This is a streaming e

 - `model`: (required) the [model name](#model-names)
 - `messages`: the messages of the chat, this can be used to keep a chat memory
- `tools`: list of tools in JSON for the model to use if supported
+- `tools`: tools for the model to use if supported. Requires `stream` to be set to `false`

 The `message` object has the following fields:

 - `role`: the role of the message, either `system`, `user`, `assistant`, or `tool`
 - `content`: the content of the message
 - `images` (optional): a list of images to include in the message (for multimodal models such as `llava`)
- `tool_calls` (optional): a list of tools in JSON that the model wants to use
+- `tool_calls` (optional): a list of tools the model wants to use

 Advanced parameters (optional):

@@ -795,7 +795,7 @@ curl http://localhost:11434/api/chat -d '{

 ##### Request

-```shell
+```
 curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2",
  "messages": [
@@ -870,7 +870,7 @@ If the messages array is empty, the model will be loaded into memory.

 ##### Request

-```shell
+```
 curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2",
  "messages": []
@@ -878,7 +878,6 @@ curl http://localhost:11434/api/chat -d '{
 ```

 ##### Response
-
 ```json
 {
  "model": "llama3.2",
@@ -898,7 +897,7 @@ If the messages array is empty and the `keep_alive` parameter is set to `0`, a m

 ##### Request

-```shell
+```
 curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2",
  "messages": [],
@@ -925,7 +924,7 @@ A single JSON object is returned:

 ## Create a Model

-```
+```shell
 POST /api/create
 ```

@@ -1021,7 +1020,7 @@ curl http://localhost:11434/api/create -d '{

 A stream of JSON objects is returned:

-```json
+```
 {"status":"quantizing F16 model to Q4_K_M"}
 {"status":"creating new layer sha256:667b0c1932bc6ffc593ed1d03f895bf2dc8dc6df21db3042284a6f4416b06a29"}
 {"status":"using existing layer sha256:11ce4ee3e170f6adebac9a991c22e22ab3f8530e154ee669954c4bc73061c258"}
@@ -1052,7 +1051,7 @@ curl http://localhost:11434/api/create -d '{

 A stream of JSON objects is returned:

-```json
+```
 {"status":"parsing GGUF"}
 {"status":"using existing layer sha256:432f310a77f4650a88d0fd59ecdd7cebed8d684bafea53cbff0473542964f0c3"}
 {"status":"writing manifest"}
@@ -1119,7 +1118,7 @@ Return 200 OK if the blob exists, 404 Not Found if it does not.

 ## Push a Blob

-```
+```shell
 POST /api/blobs/:digest
 ```

@@ -1143,7 +1142,7 @@ Return 201 Created if the blob was successfully created, 400 Bad Request if the

 ## List Local Models

-```
+```shell
 GET /api/tags
 ```

@@ -1196,7 +1195,7 @@ A single JSON object will be returned.

 ## Show Model Information

-```
+```shell
 POST /api/show
 ```

@@ -1262,7 +1261,7 @@ curl http://localhost:11434/api/show -d '{

 ## Copy a Model

-```
+```shell
 POST /api/copy
 ```

@@ -1285,7 +1284,7 @@ Returns a 200 OK if successful, or a 404 Not Found if the source model doesn't e

 ## Delete a Model

-```
+```shell
 DELETE /api/delete
 ```

@@ -1311,7 +1310,7 @@ Returns a 200 OK if successful, 404 Not Found if the model to be deleted doesn't

 ## Pull a Model

-```
+```shell
 POST /api/pull
 ```

@@ -1383,7 +1382,7 @@ if `stream` is set to false, then the response is a single JSON object:

 ## Push a Model

-```
+```shell
 POST /api/push
 ```

@@ -1448,7 +1447,7 @@ If `stream` is set to `false`, then the response is a single JSON object:

 ## Generate Embeddings

-```
+```shell
 POST /api/embed
 ```

@@ -1516,7 +1515,7 @@ curl http://localhost:11434/api/embed -d '{
 ```

 ## List Running Models
-```
+```shell
 GET /api/ps
 ```

@@ -1563,7 +1562,7 @@ A single JSON object will be returned.

 > Note: this endpoint has been superseded by `/api/embed`

-```
+```shell
 POST /api/embeddings
 ```

@@ -1603,7 +1602,7 @@ curl http://localhost:11434/api/embeddings -d '{

 ## Version

-```
+```shell
 GET /api/version
 ```

--- a/docs/development.md
+++ b/docs/development.md
@@ -1,131 +1,165 @@
 # Development

-Install prerequisites:
+Install required tools:

- [Go](https://go.dev/doc/install)
- C/C++ Compiler e.g. Clang on macOS, [TDM-GCC](https://github.com/jmeubank/tdm-gcc/releases/latest) (Windows amd64) or [llvm-mingw](https://github.com/mstorsjo/llvm-mingw) (Windows arm64), GCC/Clang on Linux.
+- go version 1.22 or higher
+- OS specific C/C++ compiler (see below)
+- GNU Make

-Then build and run Ollama from the root directory of the repository:

-```shell
-go run . serve
+## Overview
+
+Ollama uses a mix of Go and C/C++ code to interface with GPUs.  The C/C++ code is compiled with both CGO and GPU library specific compilers.  A set of GNU Makefiles are used to compile the project.  GPU Libraries are auto-detected based on the typical environment variables used by the respective libraries, but can be overridden if necessary.  The default make target will build the runners and primary Go Ollama application that will run within the repo directory.  Throughout the examples below `-j 5` is suggested for 5 parallel jobs to speed up the build.  You can adjust the job count based on your CPU Core count to reduce build times.  If you want to relocate the built binaries, use the `dist` target and recursively copy the files in `./dist/$OS-$ARCH/` to your desired location. To learn more about the other make targets use `make help`
+
+Once you have built the GPU/CPU runners, you can compile the main application with `go build .` 
+
+### MacOS
+
+[Download Go](https://go.dev/dl/)
+
+```bash
+make -j 5
 ```

-## macOS (Apple Silicon)
+Now you can run `ollama`:

-macOS Apple Silicon supports Metal which is built-in to the Ollama binary. No additional steps are required.
-
-## macOS (Intel)
-
-Install prerequisites:
-
- [CMake](https://cmake.org/download/) or `brew install cmake`
-
-Then, configure and build the project:
-
-```shell
-cmake -B build
-cmake --build build
+```bash
+./ollama
 ```

-Lastly, run Ollama:
+#### Xcode 15 warnings

-```shell
-go run . serve
+If you are using Xcode newer than version 14, you may see a warning during `go build` about `ld: warning: ignoring duplicate libraries: '-lobjc'` due to Golang issue https://github.com/golang/go/issues/67799 which can be safely ignored.  You can suppress the warning with `export CGO_LDFLAGS="-Wl,-no_warn_duplicate_libraries"`
+
+### Linux
+
+#### Linux CUDA (NVIDIA)
+
+_Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
+
+Install `make`, `gcc` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
+development and runtime packages.
+
+Typically the makefile will auto-detect CUDA, however, if your Linux distro
+or installation approach uses alternative paths, you can specify the location by
+overriding `CUDA_PATH` to the location of the CUDA toolkit. You can customize
+a set of target CUDA architectures by setting `CUDA_ARCHITECTURES` (e.g. `CUDA_ARCHITECTURES=50;60;70`)
+
+```
+make -j 5
 ```

-## Windows
+If both v11 and v12 tookkits are detected, runners for both major versions will be built by default.  You can build just v12 with `make cuda_v12`

-Install prerequisites:
+#### Older Linux CUDA (NVIDIA)

- [CMake](https://cmake.org/download/)
- [Visual Studio 2022](https://visualstudio.microsoft.com/downloads/) including the Native Desktop Workload
- (Optional) AMD GPU support
-    - [ROCm](https://rocm.github.io/install.html)
-    - [Ninja](https://github.com/ninja-build/ninja/releases)
- (Optional) NVIDIA GPU support
-    - [CUDA SDK](https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_network)
+To support older GPUs with Compute Capability 3.5 or 3.7, you will need to use an older version of the Driver from [Unix Driver Archive](https://www.nvidia.com/en-us/drivers/unix/) (tested with 470) and [CUDA Toolkit Archive](https://developer.nvidia.com/cuda-toolkit-archive) (tested with cuda V11).  When you build Ollama, you will need to set two make variable to adjust the minimum compute capability Ollama supports via `make -j 5 CUDA_ARCHITECTURES="35;37;50;52" EXTRA_GOLDFLAGS="\"-X=github.com/ollama/ollama/discover.CudaComputeMajorMin=3\" \"-X=github.com/ollama/ollama/discover.CudaComputeMinorMin=5\""`.  To find the Compute Capability of your older GPU, refer to [GPU Compute Capability](https://developer.nvidia.com/cuda-gpus).

-> [!IMPORTANT]
-> Ensure prerequisites are in `PATH` before running CMake.
+#### Linux ROCm (AMD)

-> [!IMPORTANT]
-> ROCm is not compatible with Visual Studio CMake generators. Use `-GNinja` when configuring the project.
+_Your operating system distribution may already have packages for AMD ROCm. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_

-> [!IMPORTANT]
-> CUDA is only compatible with Visual Studio CMake generators.
+Install [ROCm](https://rocm.docs.amd.com/en/latest/) development packages first, as well as `make`, `gcc`, and `golang`.

-Then, configure and build the project:
+Typically the build scripts will auto-detect ROCm, however, if your Linux distro
+or installation approach uses unusual paths, you can specify the location by
+specifying an environment variable `HIP_PATH` to the location of the ROCm
+install (typically `/opt/rocm`). You can also customize
+the AMD GPU targets by setting HIP_ARCHS (e.g. `HIP_ARCHS=gfx1101;gfx1102`)

-```shell
-cmake -B build
-cmake --build build --config Release
+```
+make -j 5
 ```

-Lastly, run Ollama:
+ROCm requires elevated privileges to access the GPU at runtime. On most distros you can add your user account to the `render` group, or run as root.

-```shell
-go run . serve
+#### Containerized Linux Build
+
+If you have Docker and buildx available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting artifacts are placed in `./dist`  and by default the script builds both arm64 and amd64 binaries.  If you want to build only amd64, you can build with `PLATFORM=linux/amd64 ./scripts/build_linux.sh`
+
+### Windows
+
+The following tools are required as a minimal development environment to build CPU inference support.
+
+- Go version 1.22 or higher
+  - https://go.dev/dl/
+- Git
+  - https://git-scm.com/download/win
+- clang with gcc compat and Make.  There are multiple options on how to go about installing these tools on Windows.  We have verified the following, but others may work as well:  
+  - [MSYS2](https://www.msys2.org/)
+    - After installing, from an MSYS2 terminal, run `pacman -S mingw-w64-clang-x86_64-gcc-compat mingw-w64-clang-x86_64-clang make` to install the required tools
+  - Assuming you used the default install prefix for msys2 above, add `C:\msys64\clang64\bin` and `c:\msys64\usr\bin` to your environment variable `PATH` where you will perform the build steps below (e.g. system-wide, account-level, powershell, cmd, etc.)
+
+> [!NOTE]  
+> Due to bugs in the GCC C++ library for unicode support, Ollama should be built with clang on windows.
+
+```
+make -j 5
 ```

-## Windows (ARM)
+#### GPU Support

-Windows ARM does not support additional acceleration libraries at this time.
+The GPU tools require the Microsoft native build tools.  To build either CUDA or ROCm, you must first install MSVC via Visual Studio:

-## Linux
+- Make sure to select `Desktop development with C++` as a Workload during the Visual Studio install
+- You must complete the Visual Studio install and run it once **BEFORE** installing CUDA or ROCm for the tools to properly register
+- Add the location of the **64 bit (x64)** compiler (`cl.exe`) to your `PATH`
+- Note: the default Developer Shell may configure the 32 bit (x86) compiler which will lead to build failures.  Ollama requires a 64 bit toolchain.

-Install prerequisites:
+#### Windows CUDA (NVIDIA)

- [CMake](https://cmake.org/download/) or `sudo apt install cmake` or `sudo dnf install cmake`
- (Optional) AMD GPU support
-    - [ROCm](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/quick-start.html)
- (Optional) NVIDIA GPU support
-    - [CUDA SDK](https://developer.nvidia.com/cuda-downloads)
+In addition to the common Windows development tools and MSVC described above:

-> [!IMPORTANT]
-> Ensure prerequisites are in `PATH` before running CMake.
+- [NVIDIA CUDA](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html)

+#### Windows ROCm (AMD Radeon)

-Then, configure and build the project:
+In addition to the common Windows development tools and MSVC described above:

-```shell
-cmake -B build
-cmake --build build
+- [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
+
+#### Windows arm64
+
+The default `Developer PowerShell for VS 2022` may default to x86 which is not what you want.  To ensure you get an arm64 development environment, start a plain PowerShell terminal and run:
+
+```powershell
+import-module 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\Common7\\Tools\\Microsoft.VisualStudio.DevShell.dll'
+Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community' -skipautomaticlocation
 ```

-Lastly, run Ollama:
+You can confirm with `write-host $env:VSCMD_ARG_TGT_ARCH`

-```shell
-go run . serve
+Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64 msys2 environment.  Ollama requires gcc and mingw32-make to compile, which is not currently available on Windows arm64, but a gcc compatibility adapter is available via `mingw-w64-clang-aarch64-gcc-compat`. At a minimum you will need to install the following:
+
+```
+pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
 ```

-## Docker
+You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)

-```shell
-docker build .
+
+## Advanced CPU Vector Settings
+
+On x86, running `make` will compile several CPU runners which can run on different CPU families. At runtime, Ollama will auto-detect the best variation to load.  If GPU libraries are present at build time, Ollama also compiles GPU runners with the `AVX` CPU vector feature enabled.  This provides a good performance balance when loading large models that split across GPU and CPU with broad compatibility.  Some users may prefer no vector extensions (e.g. older Xeon/Celeron processors, or hypervisors that mask the vector features) while other users may prefer turning on many more vector extensions to further improve performance for split model loads.
+
+To customize the set of CPU vector features enabled for a CPU runner and all GPU runners, use CUSTOM_CPU_FLAGS during the build.
+
+To build without any vector flags:
+
+```
+make CUSTOM_CPU_FLAGS=""
 ```

-### ROCm
-
-```shell
-docker build --build-arg FLAVOR=rocm .
+To build with both AVX and AVX2:
+```
+make CUSTOM_CPU_FLAGS=avx,avx2
 ```

-## Running tests
+To build with AVX512 features turned on:

-To run tests, use `go test`:
-
-```shell
-go test ./...
+```
+make CUSTOM_CPU_FLAGS=avx,avx2,avx512,avx512vbmi,avx512vnni,avx512bf16
 ```

-## Library detection
-
-Ollama looks for acceleration libraries in the following paths relative to the `ollama` executable:
-
-* `./lib/ollama` (Windows)
-* `../lib/ollama` (Linux)
-* `.` (macOS)
-* `build/lib/ollama` (for development)
-
-If the libraries are not found, Ollama will not run with any acceleration libraries.
+> [!NOTE]  
+> If you are experimenting with different flags, make sure to do a `make clean` between each change to ensure everything is rebuilt with the new compiler flags
--- a/docs/docker.md
+++ b/docs/docker.md
@@ -2,7 +2,7 @@

 ### CPU only

-```shell
+```bash
 docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
 ```

@@ -11,46 +11,42 @@ Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-

 #### Install with Apt
 1.  Configure the repository
-
-    ```shell
-    curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
-        | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
-    curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
-        | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
-        | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
-    sudo apt-get update
-    ```
-
+```bash
+curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
+    | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
+    | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
+    | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
+sudo apt-get update
+```
 2.  Install the NVIDIA Container Toolkit packages
-
-    ```shell
-    sudo apt-get install -y nvidia-container-toolkit
-    ```
+```bash
+sudo apt-get install -y nvidia-container-toolkit
+```

 #### Install with Yum or Dnf
 1.  Configure the repository

-    ```shell
-    curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
-        | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
-    ```
+```bash
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
+    | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
+```

 2. Install the NVIDIA Container Toolkit packages

-    ```shell
-    sudo yum install -y nvidia-container-toolkit
-    ```
+```bash
+sudo yum install -y nvidia-container-toolkit
+```

 #### Configure Docker to use Nvidia driver
-
-```shell
+```
 sudo nvidia-ctk runtime configure --runtime=docker
 sudo systemctl restart docker
 ```

 #### Start the container

-```shell
+```bash
 docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
 ```

@@ -61,7 +57,7 @@ docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ol

 To run Ollama using Docker with AMD GPUs, use the `rocm` tag and the following command:

-```shell
+```
 docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama:rocm
 ```

@@ -69,7 +65,7 @@ docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 114

 Now you can run a model:

-```shell
+```
 docker exec -it ollama ollama run llama3.2
 ```

--- a/docs/faq.md
+++ b/docs/faq.md
@@ -24,7 +24,7 @@ By default, Ollama uses a context window size of 2048 tokens.

 To change this when using `ollama run`, use `/set parameter`:

-```shell
+```
 /set parameter num_ctx 4096
 ```

@@ -46,15 +46,10 @@ Use the `ollama ps` command to see what models are currently loaded into memory.

 ```shell
 ollama ps
+NAME      	ID          	SIZE 	PROCESSOR	UNTIL
+llama3:70b	bcfb190ca3a7	42 GB	100% GPU 	4 minutes from now
 ```

-> **Output**:
->
-> ```
-> NAME      	ID          	SIZE 	PROCESSOR	UNTIL
-> llama3:70b	bcfb190ca3a7	42 GB	100% GPU 	4 minutes from now
-> ```
-
 The `Processor` column will show which memory the model was loaded in to:
 * `100% GPU` means the model was loaded entirely into the GPU
 * `100% CPU` means the model was loaded entirely in system memory
@@ -71,7 +66,7 @@ If Ollama is run as a macOS application, environment variables should be set usi
 1. For each environment variable, call `launchctl setenv`.

    ```bash
-    launchctl setenv OLLAMA_HOST "0.0.0.0:11434"
+    launchctl setenv OLLAMA_HOST "0.0.0.0"
    ```

 2. Restart Ollama application.
@@ -86,14 +81,14 @@ If Ollama is run as a systemd service, environment variables should be set using

    ```ini
    [Service]
-    Environment="OLLAMA_HOST=0.0.0.0:11434"
+    Environment="OLLAMA_HOST=0.0.0.0"
    ```

 3. Save and exit.

 4. Reload `systemd` and restart Ollama:

-   ```shell
+   ```bash
   systemctl daemon-reload
   systemctl restart ollama
   ```
@@ -226,19 +221,16 @@ properties.
 If you are using the API you can preload a model by sending the Ollama server an empty request. This works with both the `/api/generate` and `/api/chat` API endpoints.

 To preload the mistral model using the generate endpoint, use:
-
 ```shell
 curl http://localhost:11434/api/generate -d '{"model": "mistral"}'
 ```

 To use the chat completions endpoint, use:
-
 ```shell
 curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
 ```

 To preload a model using the CLI, use the command:
-
 ```shell
 ollama run llama3.2 ""
 ```
@@ -258,13 +250,11 @@ If you're using the API, use the `keep_alive` parameter with the `/api/generate`
 * '0' which will unload the model immediately after generating a response

 For example, to preload a model and leave it in memory use:
-
 ```shell
 curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": -1}'
 ```

 To unload the model and free up memory use:
-
 ```shell
 curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": 0}'
 ```
--- a/docs/gpu.md
+++ b/docs/gpu.md
@@ -7,7 +7,7 @@ Check your compute compatibility to see if your card is supported:

 | Compute Capability | Family              | Cards                                                                                                       |
 | ------------------ | ------------------- | ----------------------------------------------------------------------------------------------------------- |
-| 9.0                | NVIDIA              | `H200` `H100`                                                                                               |
+| 9.0                | NVIDIA              | `H100`                                                                                                      |
 | 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060`  |
 |                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
 | 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060` `RTX 3050 Ti` `RTX 3050`   |
@@ -38,7 +38,7 @@ Numeric IDs may be used, however ordering may vary, so UUIDs are more reliable.
 You can discover the UUID of your GPUs by running `nvidia-smi -L` If you want to
 ignore the GPUs and force CPU usage, use an invalid GPU ID (e.g., "-1")

-### Linux Suspend Resume
+### Laptop Suspend Resume

 On linux, after a suspend/resume cycle, sometimes Ollama will fail to discover
 your NVIDIA GPU, and fallback to running on the CPU.  You can workaround this
--- a/docs/import.md
+++ b/docs/import.md
@@ -20,13 +20,13 @@ Make sure that you use the same base model in the `FROM` command as you used to

 Now run `ollama create` from the directory where the `Modelfile` was created:

-```shell
+```bash
 ollama create my-model
 ```

 Lastly, test the model:

-```shell
+```bash
 ollama run my-model
 ```

--- a/docs/linux.md
+++ b/docs/linux.md
@@ -119,7 +119,7 @@ sudo systemctl status ollama

 To customize the installation of Ollama, you can edit the systemd service file or the environment variables by running:

-```shell
+```
 sudo systemctl edit ollama
 ```

@@ -152,7 +152,7 @@ Use `OLLAMA_VERSION` environment variable with the install script to install a s
 For example:

 ```shell
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.5.7 sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.3.9 sh
 ```

 ## Viewing logs
@@ -186,9 +186,3 @@ sudo rm -r /usr/share/ollama
 sudo userdel ollama
 sudo groupdel ollama
 ```
-
-Remove installed libraries:
-
-```shell
-sudo rm -rf /usr/local/lib/ollama
-```
--- a/docs/modelfile.md
+++ b/docs/modelfile.md
@@ -28,7 +28,7 @@ A model file is the blueprint to create and share models with Ollama.

 The format of the `Modelfile`:

-```
+```modelfile
 # comment
 INSTRUCTION arguments
 ```
@@ -49,7 +49,7 @@ INSTRUCTION arguments

 An example of a `Modelfile` creating a mario blueprint:

-```
+```modelfile
 FROM llama3.2
 # sets the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -67,32 +67,28 @@ To use this:
 3. `ollama run choose-a-model-name`
 4. Start using the model!

+More examples are available in the [examples directory](../examples).
+
 To view the Modelfile of a given model, use the `ollama show --modelfile` command.

-```shell
-ollama show --modelfile llama3.2
-```
+  ```bash
+  > ollama show --modelfile llama3.2
+  # Modelfile generated by "ollama show"
+  # To build a new Modelfile based on this one, replace the FROM line with:
+  # FROM llama3.2:latest
+  FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
+  TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>

-> **Output**:
->
-> ```
-> # Modelfile generated by "ollama show"
-> # To build a new Modelfile based on this one, replace the FROM line with:
-> # FROM llama3.2:latest
-> FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
-> TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
->
-> {{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
->
-> {{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
->
-> {{ .Response }}<|eot_id|>"""
-> PARAMETER stop "<|start_header_id|>"
-> PARAMETER stop "<|end_header_id|>"
-> PARAMETER stop "<|eot_id|>"
-> PARAMETER stop "<|reserved_special_token"
-> ```
+  {{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>

+  {{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
+
+  {{ .Response }}<|eot_id|>"""
+  PARAMETER stop "<|start_header_id|>"
+  PARAMETER stop "<|end_header_id|>"
+  PARAMETER stop "<|eot_id|>"
+  PARAMETER stop "<|reserved_special_token"
+  ```

 ## Instructions

@@ -100,13 +96,13 @@ ollama show --modelfile llama3.2

 The `FROM` instruction defines the base model to use when creating a model.

-```
+```modelfile
 FROM <model name>:<tag>
 ```

 #### Build from existing model

-```
+```modelfile
 FROM llama3.2
 ```

@@ -117,7 +113,7 @@ Additional models can be found at:

 #### Build from a Safetensors model

-```
+```modelfile
 FROM <model directory>
 ```

@@ -131,7 +127,7 @@ Currently supported model architectures:

 #### Build from a GGUF file

-```
+```modelfile
 FROM ./ollama-model.gguf
 ```

@@ -142,7 +138,7 @@ The GGUF file location should be specified as an absolute path or relative to th

 The `PARAMETER` instruction defines a parameter that can be set when the model is run.

-```
+```modelfile
 PARAMETER <parameter> <parametervalue>
 ```

@@ -159,6 +155,7 @@ PARAMETER <parameter> <parametervalue>
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
 | seed           | Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt. (Default: 0)                                                                                       | int        | seed 42              |
 | stop           | Sets the stop sequences to use. When this pattern is encountered the LLM will stop generating text and return. Multiple stop patterns may be set by specifying multiple separate `stop` parameters in a modelfile.                                      | string     | stop "AI assistant:" |
+| tfs_z          | Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting. (default: 1)                                               | float      | tfs_z 1              |
 | num_predict    | Maximum number of tokens to predict when generating text. (Default: -1, infinite generation)                                                                                                                                   | int        | num_predict 42       |
 | top_k          | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)                                                                        | int        | top_k 40             |
 | top_p          | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)                                                                 | float      | top_p 0.9            |
@@ -189,7 +186,7 @@ TEMPLATE """{{ if .System }}<|im_start|>system

 The `SYSTEM` instruction specifies the system message to be used in the template, if applicable.

-```
+```modelfile
 SYSTEM """<system message>"""
 ```

@@ -199,7 +196,7 @@ The `ADAPTER` instruction specifies a fine tuned LoRA adapter that should apply

 #### Safetensor adapter

-```
+```modelfile
 ADAPTER <path to safetensor adapter>
 ```

@@ -210,7 +207,7 @@ Currently supported Safetensor adapters:

 #### GGUF adapter

-```
+```modelfile
 ADAPTER ./ollama-lora.gguf
 ```

@@ -218,7 +215,7 @@ ADAPTER ./ollama-lora.gguf

 The `LICENSE` instruction allows you to specify the legal license under which the model used with this Modelfile is shared or distributed.

-```
+```modelfile
 LICENSE """
 <license text>
 """
@@ -228,7 +225,7 @@ LICENSE """

 The `MESSAGE` instruction allows you to specify a message history for the model to use when responding. Use multiple iterations of the MESSAGE command to build up a conversation which will guide the model to answer in a similar way.

-```
+```modelfile
 MESSAGE <role> <message>
 ```

@@ -243,7 +240,7 @@ MESSAGE <role> <message>

 #### Example conversation

-```
+```modelfile
 MESSAGE user Is Toronto in Canada?
 MESSAGE assistant yes
 MESSAGE user Is Sacramento in Canada?
--- a/docs/openai.md
+++ b/docs/openai.md
@@ -1,7 +1,6 @@
 # OpenAI compatibility

-> [!NOTE]
-> OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/ollama/ollama/blob/main/docs/api.md).
+> **Note:** OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/ollama/ollama/blob/main/docs/api.md).

 Ollama provides experimental compatibility with parts of the [OpenAI API](https://platform.openai.com/docs/api-reference) to help connect existing applications to Ollama.

@@ -60,10 +59,8 @@ embeddings = client.embeddings.create(
    input=["why is the sky blue?", "why is the grass green?"],
 )
 ```
-
 #### Structured outputs
-
-```python
+```py
 from pydantic import BaseModel
 from openai import OpenAI

@@ -147,7 +144,7 @@ const embedding = await openai.embeddings.create({

 ### `curl`

-```shell
+``` shell
 curl http://localhost:11434/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
@@ -322,7 +319,7 @@ ollama pull llama3.2

 For tooling that relies on default OpenAI model names such as `gpt-3.5-turbo`, use `ollama cp` to copy an existing model name to a temporary name:

-```shell
+```
 ollama cp llama3.2 gpt-3.5-turbo
 ```

@@ -346,7 +343,7 @@ curl http://localhost:11434/v1/chat/completions \

 The OpenAI API does not have a way of setting the context size for a model. If you need to change the context size, create a `Modelfile` which looks like:

-```
+```modelfile
 FROM <some model>
 PARAMETER num_ctx <context size>
 ```
--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@@ -17,7 +17,6 @@ When you run Ollama in a **container**, the logs go to stdout/stderr in the cont
 ```shell
 docker logs <container-name>
 ```
-
 (Use `docker ps` to find the container name)

 If manually running `ollama serve` in a terminal, the logs will be on that terminal.
@@ -29,7 +28,6 @@ When you run Ollama on **Windows**, there are a few different locations. You can
 - `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories

 To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
-
 ```powershell
 $env:OLLAMA_DEBUG="1"
 & "ollama app.exe"
@@ -51,13 +49,12 @@ Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]

 You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to bypass autodetection, so for example, if you have a CUDA card, but want to force the CPU LLM library with AVX2 vector support, use:

-```shell
+```
 OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
 ```

 You can see what features your CPU has with the following.
-
-```shell
+```
 cat /proc/cpuinfo| grep flags | head -1
 ```

@@ -65,8 +62,8 @@ cat /proc/cpuinfo| grep flags | head -1

 If you run into problems on Linux and want to install an older version, or you'd like to try out a pre-release before it's officially released, you can tell the install script which version to install.

-```shell
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.5.7 sh
+```sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
 ```

 ## Linux tmp noexec 
--- a/docs/windows.md
+++ b/docs/windows.md
@@ -47,7 +47,6 @@ If Ollama is already running, Quit the tray application and relaunch it from the
 ## API Access

 Here's a quick example showing API access from `powershell`
-
 ```powershell
 (Invoke-WebRequest -method POST -Body '{"model":"llama3.2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
 ```
@@ -55,7 +54,7 @@ Here's a quick example showing API access from `powershell`
 ## Troubleshooting

 Ollama on Windows stores files in a few different locations.  You can view them in
-the explorer window by hitting `<Ctrl>+R` and type in:
+the explorer window by hitting `<cmd>+R` and type in:
 - `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
    - *app.log* contains most resent logs from the GUI application
    - *server.log* contains the most recent server logs
--- a/envconfig/config.go
+++ b/envconfig/config.go
@@ -165,8 +165,6 @@ var (
 	IntelGPU = Bool("OLLAMA_INTEL_GPU")
 	// MultiUserCache optimizes prompt caching for multi-user scenarios
 	MultiUserCache = Bool("OLLAMA_MULTIUSER_CACHE")
-	// Enable the new Ollama engine
-	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 )

 func String(s string) func() string {
@@ -252,7 +250,6 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
-		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},

 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
@@ -291,3 +288,12 @@ func Values() map[string]string {
 func Var(key string) string {
 	return strings.Trim(strings.TrimSpace(os.Getenv(key)), "\"'")
 }
+
+// On windows, we keep the binary at the top directory, but
+// other platforms use a "bin" directory, so this returns ".."
+func LibRelativeToExe() string {
+	if runtime.GOOS == "windows" {
+		return "."
+	}
+	return ".."
+}
--- a/format/bytes.go
+++ b/format/bytes.go
@@ -40,6 +40,8 @@ func HumanBytes(b int64) string {
 	}

 	switch {
+	case value >= 100:
+		return fmt.Sprintf("%d %s", int(value), unit)
 	case value >= 10:
 		return fmt.Sprintf("%d %s", int(value), unit)
 	case value != math.Trunc(value):
--- a/format/bytes_test.go
+++ b/format/bytes_test.go
@@ -1,91 +0,0 @@
-package format
-
-import (
-	"testing"
-)
-
-func TestHumanBytes(t *testing.T) {
-	type testCase struct {
-		input    int64
-		expected string
-	}
-
-	tests := []testCase{
-		// Test bytes (B)
-		{0, "0 B"},
-		{1, "1 B"},
-		{999, "999 B"},
-
-		// Test kilobytes (KB)
-		{1000, "1 KB"},
-		{1500, "1.5 KB"},
-		{999999, "999 KB"},
-
-		// Test megabytes (MB)
-		{1000000, "1 MB"},
-		{1500000, "1.5 MB"},
-		{999999999, "999 MB"},
-
-		// Test gigabytes (GB)
-		{1000000000, "1 GB"},
-		{1500000000, "1.5 GB"},
-		{999999999999, "999 GB"},
-
-		// Test terabytes (TB)
-		{1000000000000, "1 TB"},
-		{1500000000000, "1.5 TB"},
-		{1999999999999, "2.0 TB"},
-
-		// Test fractional values
-		{1234, "1.2 KB"},
-		{1234567, "1.2 MB"},
-		{1234567890, "1.2 GB"},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.expected, func(t *testing.T) {
-			result := HumanBytes(tc.input)
-			if result != tc.expected {
-				t.Errorf("Expected %s, got %s", tc.expected, result)
-			}
-		})
-	}
-}
-
-func TestHumanBytes2(t *testing.T) {
-	type testCase struct {
-		input    uint64
-		expected string
-	}
-
-	tests := []testCase{
-		// Test bytes (B)
-		{0, "0 B"},
-		{1, "1 B"},
-		{1023, "1023 B"},
-
-		// Test kibibytes (KiB)
-		{1024, "1.0 KiB"},
-		{1536, "1.5 KiB"},
-		{1048575, "1024.0 KiB"},
-
-		// Test mebibytes (MiB)
-		{1048576, "1.0 MiB"},
-		{1572864, "1.5 MiB"},
-		{1073741823, "1024.0 MiB"},
-
-		// Test gibibytes (GiB)
-		{1073741824, "1.0 GiB"},
-		{1610612736, "1.5 GiB"},
-		{2147483648, "2.0 GiB"},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.expected, func(t *testing.T) {
-			result := HumanBytes2(tc.input)
-			if result != tc.expected {
-				t.Errorf("Expected %s, got %s", tc.expected, result)
-			}
-		})
-	}
-}
--- a/format/format_test.go
+++ b/format/format_test.go
@@ -12,9 +12,6 @@ func TestHumanNumber(t *testing.T) {

 	testCases := []testCase{
 		{0, "0"},
-		{999, "999"},
-		{1000, "1K"},
-		{1001, "1K"},
 		{1000000, "1M"},
 		{125000000, "125M"},
 		{500500000, "500.50M"},
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -153,17 +153,19 @@ func (s Tensors) Items(prefix ...string) []*Tensor {
 	return items
 }

-func (ts Tensors) GroupLayers() map[string]Layer {
+func (ts Tensors) Layers() map[string]Layer {
 	layers := make(map[string]Layer)
 	for _, t := range ts.items {
 		parts := strings.Split(t.Name, ".")
-		if index := slices.IndexFunc(parts, func(s string) bool { return s == "blk" || s == "mm" }); index != -1 {
-			if len(parts) > index+2 {
-				// blk and mm should have a number after them, join it
-				parts = append(
-					[]string{strings.Join(parts[:index+2], ".")},
-					parts[index+2:]...)
-			}
+		if i := slices.Index(parts, "blk"); i > 0 {
+			parts = append([]string{
+				strings.Join(parts[:i], "."),
+				strings.Join(parts[i:i+2], "."),
+			}, parts[i+2:]...)
+		} else if i == 0 {
+			parts = append([]string{
+				strings.Join(parts[i:i+2], "."),
+			}, parts[i+2:]...)
 		}

 		if _, ok := layers[parts[0]]; !ok {
@@ -375,22 +377,22 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	}, offset, nil
 }

-func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
-	embedding := f.KV().EmbeddingLength()
-	heads := f.KV().HeadCount()
-	headsKV := f.KV().HeadCountKV()
-	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array).size)
+func (llm GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
+	embedding := llm.KV().EmbeddingLength()
+	heads := llm.KV().HeadCount()
+	headsKV := llm.KV().HeadCountKV()
+	vocab := uint64(llm.KV()["tokenizer.ggml.tokens"].(*array).size)

-	embeddingHeads := f.KV().EmbeddingHeadCount()
-	embeddingHeadsK := f.KV().EmbeddingHeadCountK()
-	embeddingHeadsV := f.KV().EmbeddingHeadCountV()
+	embeddingHeads := llm.KV().EmbeddingHeadCount()
+	embeddingHeadsK := llm.KV().EmbeddingHeadCountK()
+	embeddingHeadsV := llm.KV().EmbeddingHeadCountV()

-	layers := f.Tensors().GroupLayers()
+	layers := llm.Tensors().Layers()

 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	kv = uint64(float64(context*f.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	kv = uint64(float64(context*llm.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)

-	switch f.KV().Architecture() {
+	switch llm.KV().Architecture() {
 	case "llama":
 		fullOffload = max(
 			4*batch*(1+4*embedding+context*(1+heads)),
@@ -405,7 +407,7 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO

 		if ffnGateExpsWeight, ok := layers["blk.0"]["ffn_gate_exps.weight"]; ok {
 			// mixtral 8x22b
-			ff := uint64(f.KV()["llama.feed_forward_length"].(uint32))
+			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
 				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
 				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
@@ -422,11 +424,11 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	case "mllama":
 		var visionTokens, tiles uint64 = 1601, 4

-		if crossAttentionLayers, ok := f.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
+		if crossAttentionLayers, ok := llm.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
 			kv = headsKV *
 				(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
 				(2* // sizeof(float16)
-					(f.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
+					(llm.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
 					context +
 					4* // sizeof(float32)
 						uint64(crossAttentionLayers.size)* // num cross attention layers
@@ -441,7 +443,7 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 		)

 		var ropeFreqsCount uint64
-		if ropeFreqs, ok := f.Tensors().GroupLayers()["rope_freqs"]; ok {
+		if ropeFreqs, ok := llm.Tensors().Layers()["rope_freqs"]; ok {
 			if ropeFreqsWeights, ok := ropeFreqs["weights"]; ok {
 				ropeFreqsCount = ropeFreqsWeights.parameters()
 			}
@@ -545,20 +547,20 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 }

 // SupportsKVCacheType checks if the requested cache type is supported
-func (f GGML) SupportsKVCacheType(cacheType string) bool {
+func (llm GGML) SupportsKVCacheType(cacheType string) bool {
 	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
 }

 // SupportsFlashAttention checks if the model supports flash attention
-func (f GGML) SupportsFlashAttention() bool {
-	_, isEmbedding := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]
+func (llm GGML) SupportsFlashAttention() bool {
+	_, isEmbedding := llm.KV()[fmt.Sprintf("%s.pooling_type", llm.KV().Architecture())]
 	if isEmbedding {
 		return false
 	}

 	// Check head counts match and are non-zero
-	headCountK := f.KV().EmbeddingHeadCountK()
-	headCountV := f.KV().EmbeddingHeadCountV()
+	headCountK := llm.KV().EmbeddingHeadCountK()
+	headCountV := llm.KV().EmbeddingHeadCountV()
 	return headCountK != 0 && headCountV != 0 && headCountK == headCountV
 }

--- a/fs/ggml/ggml_test.go
+++ b/fs/ggml/ggml_test.go
@@ -1,159 +0,0 @@
-package ggml
-
-import (
-	"maps"
-	"slices"
-	"strings"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-)
-
-func TestTensorLayers(t *testing.T) {
-	tensors := make(map[string]*Tensor)
-	for _, name := range []string{
-		"token_embd.weight",
-		"blk.0.attn_k.weight",
-		"blk.0.attn_output.weight",
-		"blk.0.attn_q.weight",
-		"blk.0.attn_v.weight",
-		"blk.0.attn_norm.weight",
-		"blk.0.ffn_down.weight",
-		"blk.0.ffn_gate.weight",
-		"blk.0.ffn_up.weight",
-		"blk.0.ffn_norm.weight",
-		"output_norm.weight",
-		"mm.0.bias",
-		"mm.0.weight",
-		"v.blk.0.attn_k.weight",
-		"v.blk.0.attn_output.weight",
-		"v.blk.0.attn_q.weight",
-		"v.blk.0.attn_v.weight",
-		"v.blk.0.attn_norm.weight",
-		"v.blk.0.ffn_down.weight",
-		"v.blk.0.ffn_gate.weight",
-		"v.blk.0.ffn_up.weight",
-		"v.blk.0.ffn_norm.weight",
-		"v.patch_embd.weight",
-		"v.position_embd.gate",
-		"v.position_embd.weight",
-	} {
-		tensors[name] = &Tensor{Name: name}
-	}
-
-	cases := []struct {
-		name  string
-		items []*Tensor
-		want  map[string]Layer
-	}{
-		{
-			name: "text",
-			items: slices.Collect(func(yield func(*Tensor) bool) {
-				for k, v := range tensors {
-					if !strings.HasPrefix(k, "mm.") && !strings.HasPrefix(k, "v.") {
-						if !yield(v) {
-							return
-						}
-					}
-				}
-			}),
-			want: map[string]Layer{
-				"blk.0": {
-					"attn_k.weight":      tensors["blk.0.attn_k.weight"],
-					"attn_q.weight":      tensors["blk.0.attn_q.weight"],
-					"attn_v.weight":      tensors["blk.0.attn_v.weight"],
-					"attn_output.weight": tensors["blk.0.attn_output.weight"],
-					"attn_norm.weight":   tensors["blk.0.attn_norm.weight"],
-					"ffn_down.weight":    tensors["blk.0.ffn_down.weight"],
-					"ffn_gate.weight":    tensors["blk.0.ffn_gate.weight"],
-					"ffn_up.weight":      tensors["blk.0.ffn_up.weight"],
-					"ffn_norm.weight":    tensors["blk.0.ffn_norm.weight"],
-				},
-				"token_embd":  {"weight": tensors["token_embd.weight"]},
-				"output_norm": {"weight": tensors["output_norm.weight"]},
-			},
-		},
-		{
-			name: "vision",
-			items: slices.Collect(func(yield func(*Tensor) bool) {
-				for k, v := range tensors {
-					if strings.HasPrefix(k, "mm.") || strings.HasPrefix(k, "v.") {
-						if !yield(v) {
-							return
-						}
-					}
-				}
-			}),
-			want: map[string]Layer{
-				"mm.0": {
-					"bias":   tensors["mm.0.bias"],
-					"weight": tensors["mm.0.weight"],
-				},
-				"v.blk.0": {
-					"attn_k.weight":      tensors["v.blk.0.attn_k.weight"],
-					"attn_q.weight":      tensors["v.blk.0.attn_q.weight"],
-					"attn_v.weight":      tensors["v.blk.0.attn_v.weight"],
-					"attn_output.weight": tensors["v.blk.0.attn_output.weight"],
-					"attn_norm.weight":   tensors["v.blk.0.attn_norm.weight"],
-					"ffn_down.weight":    tensors["v.blk.0.ffn_down.weight"],
-					"ffn_gate.weight":    tensors["v.blk.0.ffn_gate.weight"],
-					"ffn_up.weight":      tensors["v.blk.0.ffn_up.weight"],
-					"ffn_norm.weight":    tensors["v.blk.0.ffn_norm.weight"],
-				},
-				"v": {
-					"patch_embd.weight":    tensors["v.patch_embd.weight"],
-					"position_embd.gate":   tensors["v.position_embd.gate"],
-					"position_embd.weight": tensors["v.position_embd.weight"],
-				},
-			},
-		},
-		{
-			name:  "vision and text",
-			items: slices.Collect(maps.Values(tensors)),
-			want: map[string]Layer{
-				"blk.0": {
-					"attn_k.weight":      tensors["blk.0.attn_k.weight"],
-					"attn_q.weight":      tensors["blk.0.attn_q.weight"],
-					"attn_v.weight":      tensors["blk.0.attn_v.weight"],
-					"attn_output.weight": tensors["blk.0.attn_output.weight"],
-					"attn_norm.weight":   tensors["blk.0.attn_norm.weight"],
-					"ffn_down.weight":    tensors["blk.0.ffn_down.weight"],
-					"ffn_gate.weight":    tensors["blk.0.ffn_gate.weight"],
-					"ffn_up.weight":      tensors["blk.0.ffn_up.weight"],
-					"ffn_norm.weight":    tensors["blk.0.ffn_norm.weight"],
-				},
-				"token_embd":  {"weight": tensors["token_embd.weight"]},
-				"output_norm": {"weight": tensors["output_norm.weight"]},
-				"mm.0": {
-					"bias":   tensors["mm.0.bias"],
-					"weight": tensors["mm.0.weight"],
-				},
-				"v.blk.0": {
-					"attn_k.weight":      tensors["v.blk.0.attn_k.weight"],
-					"attn_q.weight":      tensors["v.blk.0.attn_q.weight"],
-					"attn_v.weight":      tensors["v.blk.0.attn_v.weight"],
-					"attn_output.weight": tensors["v.blk.0.attn_output.weight"],
-					"attn_norm.weight":   tensors["v.blk.0.attn_norm.weight"],
-					"ffn_down.weight":    tensors["v.blk.0.ffn_down.weight"],
-					"ffn_gate.weight":    tensors["v.blk.0.ffn_gate.weight"],
-					"ffn_up.weight":      tensors["v.blk.0.ffn_up.weight"],
-					"ffn_norm.weight":    tensors["v.blk.0.ffn_norm.weight"],
-				},
-				"v": {
-					"patch_embd.weight":    tensors["v.patch_embd.weight"],
-					"position_embd.gate":   tensors["v.position_embd.gate"],
-					"position_embd.weight": tensors["v.position_embd.weight"],
-				},
-			},
-		},
-	}
-
-	for _, tt := range cases {
-		t.Run(tt.name, func(t *testing.T) {
-			got := Tensors{items: tt.items}.GroupLayers()
-			if diff := cmp.Diff(got, tt.want); diff != "" {
-				t.Errorf("unexpected layers (-got +want):\n%s", diff)
-			}
-		})
-	}
-}
--- a/fs/ggml/type.go
+++ b/fs/ggml/type.go
@@ -32,10 +32,9 @@ const (
 	fileTypeIQ1_S
 	fileTypeIQ4_NL
 	fileTypeIQ3_S
-	fileTypeIQ3_M
 	fileTypeIQ2_S
-	fileTypeIQ2_M
 	fileTypeIQ4_XS
+	fileTypeIQ2_M
 	fileTypeIQ1_M
 	fileTypeBF16

@@ -94,14 +93,12 @@ func ParseFileType(s string) (fileType, error) {
 		return fileTypeIQ4_NL, nil
 	case "IQ3_S":
 		return fileTypeIQ3_S, nil
-	case "IQ3_M":
-		return fileTypeIQ3_M, nil
 	case "IQ2_S":
 		return fileTypeIQ2_S, nil
-	case "IQ2_M":
-		return fileTypeIQ2_M, nil
 	case "IQ4_XS":
 		return fileTypeIQ4_XS, nil
+	case "IQ2_M":
+		return fileTypeIQ2_M, nil
 	case "IQ1_M":
 		return fileTypeIQ1_M, nil
 	case "BF16":
@@ -163,8 +160,6 @@ func (t fileType) String() string {
 		return "IQ4_NL"
 	case fileTypeIQ3_S:
 		return "IQ3_S"
-	case fileTypeIQ3_M:
-		return "IQ3_M"
 	case fileTypeIQ2_S:
 		return "IQ2_S"
 	case fileTypeIQ4_XS:
--- a/go.mod
+++ b/go.mod
@@ -24,6 +24,7 @@ require (
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 	golang.org/x/image v0.22.0
+	golang.org/x/tools v0.28.0
 	gonum.org/v1/gonum v0.15.0
 )

@@ -71,7 +72,7 @@ require (
 	golang.org/x/arch v0.8.0 // indirect
 	golang.org/x/crypto v0.31.0
 	golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa
-	golang.org/x/net v0.25.0 // indirect
+	golang.org/x/net v0.32.0 // indirect
 	golang.org/x/sys v0.28.0
 	golang.org/x/term v0.27.0
 	golang.org/x/text v0.21.0
--- a/go.sum
+++ b/go.sum
@@ -257,8 +257,8 @@ golang.org/x/net v0.0.0-20200822124328-c89045814202/go.mod h1:/O7V0waA8r7cgGh81R
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.0.0-20210614182718-04defd469f4e/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
-golang.org/x/net v0.25.0 h1:d/OCCoBEUq33pjydKrGQhw7IlUPI2Oylr+8qLx49kac=
-golang.org/x/net v0.25.0/go.mod h1:JkAGAh7GEvH74S6FOH42FLoXpXbE/aqXSrIQjXgsiwM=
+golang.org/x/net v0.32.0 h1:ZqPmj8Kzc+Y6e0+skZsuACbx+wzMgo5MQsJh9Qd6aYI=
+golang.org/x/net v0.32.0/go.mod h1:CwU0IoeOlnQQWJ6ioyFrfRuomB8GKF6KbYXZVyeXNfs=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.0.0-20200107190931-bf48bf16ab8d/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -309,6 +309,8 @@ golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapK
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
 golang.org/x/tools v0.1.4/go.mod h1:o0xws9oXOQQZyjljx8fwUC0k7L1pTE6eaCbjGeHmOkk=
+golang.org/x/tools v0.28.0 h1:WuB6qZ4RPCQo5aP3WdKZS7i595EdWqWR8vqJTlwTVK8=
+golang.org/x/tools v0.28.0/go.mod h1:dcIOrVd3mfQKTgrDVQHqCPMWy6lnhfhtX3hLXYVLfRw=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
--- a/grammar/bench_test.go
+++ b/grammar/bench_test.go
@@ -0,0 +1,22 @@
+//go:build go1.24
+
+package grammar
+
+import "testing"
+
+func BenchmarkFromSchema(b *testing.B) {
+	for tt := range testCases(b) {
+		b.Run("", func(b *testing.B) {
+			s := []byte(tt.schema)
+
+			b.ReportAllocs()
+			for b.Loop() {
+				_, err := FromSchema(nil, s)
+				if err != nil {
+					b.Fatalf("GrammarFromSchema: %v", err)
+				}
+			}
+		})
+		return
+	}
+}
--- a/grammar/grammar.go
+++ b/grammar/grammar.go
@@ -0,0 +1,227 @@
+package grammar
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"iter"
+	"strconv"
+
+	"github.com/ollama/ollama/grammar/jsonschema"
+)
+
+const jsonTerms = `
+# Unicode
+#
+# Unicode characters can be specified directly in the grammar, for example
+# hiragana ::= [ぁ-ゟ], or with escapes: 8-bit (\xXX), 16-bit (\uXXXX) or 32-bit
+# (\UXXXXXXXX).
+unicode ::= \x{hex}{2} | \u{hex}{4} | \U{hex}{8}
+
+# JSON grammar from RFC 7159
+null    ::= "null"
+object  ::= "{" (kv ("," kv)*)? "}"
+array   ::= "[" (value ("," value)*)? "]"
+kv      ::= string ":" value
+integer ::= "0" | [1-9] [0-9]*
+number  ::= "-"? integer frac? exp?
+frac    ::= "." [0-9]+
+exp     ::= ("e" | "E") ("+" | "-") [0-9]+
+string  ::= "\"" char* "\""
+escape  ::= ["/" | "b" | "f" | "n" | "r" | "t" | unicode]
+char    ::= [^"\\] | escape
+space   ::= (" " | "\t" | "\n" | "\r")*
+hex     ::= [0-9] | [a-f] | [A-F]
+boolean ::= "true" | "false"
+value   ::= object | array | string | number | boolean | "null"
+
+# User-defined
+`
+
+// FromSchema generates a grammar from a JSON schema.
+func FromSchema(buf []byte, jsonSchema []byte) ([]byte, error) {
+	var s *jsonschema.Schema
+	if err := json.Unmarshal(jsonSchema, &s); err != nil {
+		return nil, err
+	}
+
+	var g builder
+
+	// "root" is the only rule that is guaranteed to exist, so we start
+	// with its length for padding, and then adjust it as we go.
+	g.pad = len("root")
+	for id := range dependencies("root", s) {
+		g.pad = max(g.pad, len(id))
+	}
+
+	g.b.WriteString(jsonTerms)
+
+	ids := make(map[*jsonschema.Schema]string)
+	for id, s := range dependencies("root", s) {
+		ids[s] = id
+		g.define(id)
+		if err := fromSchema(&g, ids, s); err != nil {
+			return nil, err
+		}
+	}
+	g.define("root")
+	if err := fromSchema(&g, ids, s); err != nil {
+		return nil, err
+	}
+	g.define("") // finalize the last rule
+	return g.b.Bytes(), nil
+}
+
+func fromSchema(g *builder, ids map[*jsonschema.Schema]string, s *jsonschema.Schema) error {
+	switch typ := s.EffectiveType(); typ {
+	case "array":
+		if len(s.PrefixItems) == 0 && s.Items == nil {
+			g.u("array")
+		} else {
+			g.q("[")
+			for i, s := range s.PrefixItems {
+				if i > 0 {
+					g.q(",")
+				}
+				g.u(ids[s])
+			}
+			if s.Items != nil {
+				g.u("(")
+				if len(s.PrefixItems) > 0 {
+					g.q(",")
+				}
+				g.u(ids[s.Items])
+				g.u(")*")
+			}
+			g.q("]")
+		}
+	case "object":
+		if len(s.Properties) == 0 {
+			g.u("object")
+		} else {
+			g.q("{")
+			for i, p := range s.Properties {
+				name := ids[p]
+				if i > 0 {
+					g.q(",")
+				}
+				g.q(p.Name)
+				g.q(":")
+				g.u(name)
+			}
+			g.q("}")
+		}
+	case "number":
+		buildConstrainedNumber(g, s)
+	case "string":
+		if len(s.Enum) == 0 {
+			g.u("string")
+		} else {
+			g.u("(")
+			for i, e := range s.Enum {
+				if i > 0 {
+					g.q("|")
+				}
+				g.q(string(e))
+			}
+			g.u(")")
+		}
+	case "boolean", "value", "null", "integer":
+		g.u(typ)
+	default:
+		return fmt.Errorf("%s: unsupported type %q", s.Name, typ)
+	}
+	return nil
+}
+
+// dependencies returns a sequence of all child dependencies of the schema in
+// post-order.
+//
+// The first value is the id/pointer to the dependency, and the second value
+// is the schema.
+func dependencies(id string, s *jsonschema.Schema) iter.Seq2[string, *jsonschema.Schema] {
+	return func(yield func(string, *jsonschema.Schema) bool) {
+		for i, p := range s.Properties {
+			id := fmt.Sprintf("%s_%d", id, i)
+			for did, d := range dependencies(id, p) {
+				if !yield(did, d) {
+					return
+				}
+			}
+			if !yield(id, p) {
+				return
+			}
+		}
+		for i, p := range s.PrefixItems {
+			id := fmt.Sprintf("tuple_%d", i)
+			for did, d := range dependencies(id, p) {
+				id := fmt.Sprintf("%s_%s", id, did)
+				if !yield(id, d) {
+					return
+				}
+			}
+			if !yield(id, p) {
+				return
+			}
+		}
+		if s.Items != nil {
+			id := fmt.Sprintf("%s_tuple_%d", id, len(s.PrefixItems))
+			for did, d := range dependencies(id, s.Items) {
+				if !yield(did, d) {
+					return
+				}
+			}
+			if !yield(id, s.Items) {
+				return
+			}
+		}
+	}
+}
+
+type builder struct {
+	b     bytes.Buffer
+	pad   int
+	rules int
+	items int
+}
+
+// define terminates the current rule, if any, and then either starts a new
+// rule or does nothing else if the name is empty.
+func (b *builder) define(name string) {
+	if b.rules > 0 {
+		b.b.WriteString(";\n")
+	}
+	if name == "" {
+		return
+	}
+	fmt.Fprintf(&b.b, "% -*s", b.pad, name)
+	b.b.WriteString(" ::=")
+	b.rules++
+	b.items = 0
+}
+
+// quote appends a terminal to the current rule.
+func (b *builder) q(s string) {
+	if b.items > 0 {
+		b.b.WriteString(" ")
+	}
+	b.b.WriteString(" ")
+	b.b.WriteString(strconv.Quote(s))
+}
+
+// u appends a non-terminal to the current rule.
+func (b *builder) u(s string) {
+	if b.items > 0 {
+		b.b.WriteString(" ")
+	}
+	b.b.WriteString(" ")
+	b.b.WriteString(s)
+}
+
+func buildConstrainedNumber(b *builder, s *jsonschema.Schema) {
+	if s.Minimum == 0 && s.Maximum == 0 {
+		b.u("TODO")
+	} else {
+		b.u("number")
+	}
+}
--- a/grammar/grammar_test.go
+++ b/grammar/grammar_test.go
@@ -0,0 +1,75 @@
+package grammar
+
+import (
+	"bufio"
+	"cmp"
+	"iter"
+	"strings"
+	"testing"
+
+	_ "embed"
+
+	"github.com/ollama/ollama/grammar/internal/diff"
+)
+
+func TestFromSchema(t *testing.T) {
+	for tt := range testCases(t) {
+		t.Run(tt.name, func(t *testing.T) {
+			g, err := FromSchema(nil, []byte(tt.schema))
+			if err != nil {
+				t.Fatalf("FromSchema: %v", err)
+			}
+			got := string(g)
+			got = strings.TrimPrefix(got, jsonTerms)
+			if got != tt.want {
+				t.Logf("schema:\n%s", tt.schema)
+				t.Fatal(string(diff.Diff("got", []byte(got), "want", []byte(tt.want))))
+			}
+		})
+	}
+}
+
+type testCase struct {
+	name   string
+	schema string
+	want   string
+}
+
+//go:embed testdata/schemas.txt
+var tests string
+
+func testCases(t testing.TB) iter.Seq[testCase] {
+	t.Helper()
+	return func(yield func(testCase) bool) {
+		t.Helper()
+		sc := bufio.NewScanner(strings.NewReader(tests))
+		name := ""
+		for sc.Scan() {
+			line := strings.TrimSpace(sc.Text())
+			if line == "" {
+				name = ""
+				continue
+			}
+			if line[0] == '#' {
+				name = cmp.Or(name, strings.TrimSpace(line[1:]))
+				continue
+			}
+			s := sc.Text()
+			g := ""
+			for sc.Scan() {
+				line = strings.TrimSpace(sc.Text())
+				if line == "" || line[0] == '#' {
+					break
+				}
+				g += sc.Text() + "\n"
+			}
+			if !yield(testCase{name, s, g}) {
+				return
+			}
+			name = strings.TrimSpace(strings.TrimPrefix(line, "#"))
+		}
+		if err := sc.Err(); err != nil {
+			t.Fatalf("error reading tests: %v", err)
+		}
+	}
+}
--- a/grammar/internal/diff/diff.go
+++ b/grammar/internal/diff/diff.go
@@ -0,0 +1,261 @@
+// Copyright 2022 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package diff
+
+import (
+	"bytes"
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// A pair is a pair of values tracked for both the x and y side of a diff.
+// It is typically a pair of line indexes.
+type pair struct{ x, y int }
+
+// Diff returns an anchored diff of the two texts old and new
+// in the “unified diff” format. If old and new are identical,
+// Diff returns a nil slice (no output).
+//
+// Unix diff implementations typically look for a diff with
+// the smallest number of lines inserted and removed,
+// which can in the worst case take time quadratic in the
+// number of lines in the texts. As a result, many implementations
+// either can be made to run for a long time or cut off the search
+// after a predetermined amount of work.
+//
+// In contrast, this implementation looks for a diff with the
+// smallest number of “unique” lines inserted and removed,
+// where unique means a line that appears just once in both old and new.
+// We call this an “anchored diff” because the unique lines anchor
+// the chosen matching regions. An anchored diff is usually clearer
+// than a standard diff, because the algorithm does not try to
+// reuse unrelated blank lines or closing braces.
+// The algorithm also guarantees to run in O(n log n) time
+// instead of the standard O(n²) time.
+//
+// Some systems call this approach a “patience diff,” named for
+// the “patience sorting” algorithm, itself named for a solitaire card game.
+// We avoid that name for two reasons. First, the name has been used
+// for a few different variants of the algorithm, so it is imprecise.
+// Second, the name is frequently interpreted as meaning that you have
+// to wait longer (to be patient) for the diff, meaning that it is a slower algorithm,
+// when in fact the algorithm is faster than the standard one.
+func Diff(oldName string, old []byte, newName string, new []byte) []byte {
+	if bytes.Equal(old, new) {
+		return nil
+	}
+	x := lines(old)
+	y := lines(new)
+
+	// Print diff header.
+	var out bytes.Buffer
+	fmt.Fprintf(&out, "diff %s %s\n", oldName, newName)
+	fmt.Fprintf(&out, "--- %s\n", oldName)
+	fmt.Fprintf(&out, "+++ %s\n", newName)
+
+	// Loop over matches to consider,
+	// expanding each match to include surrounding lines,
+	// and then printing diff chunks.
+	// To avoid setup/teardown cases outside the loop,
+	// tgs returns a leading {0,0} and trailing {len(x), len(y)} pair
+	// in the sequence of matches.
+	var (
+		done  pair     // printed up to x[:done.x] and y[:done.y]
+		chunk pair     // start lines of current chunk
+		count pair     // number of lines from each side in current chunk
+		ctext []string // lines for current chunk
+	)
+	for _, m := range tgs(x, y) {
+		if m.x < done.x {
+			// Already handled scanning forward from earlier match.
+			continue
+		}
+
+		// Expand matching lines as far as possible,
+		// establishing that x[start.x:end.x] == y[start.y:end.y].
+		// Note that on the first (or last) iteration we may (or definitely do)
+		// have an empty match: start.x==end.x and start.y==end.y.
+		start := m
+		for start.x > done.x && start.y > done.y && x[start.x-1] == y[start.y-1] {
+			start.x--
+			start.y--
+		}
+		end := m
+		for end.x < len(x) && end.y < len(y) && x[end.x] == y[end.y] {
+			end.x++
+			end.y++
+		}
+
+		// Emit the mismatched lines before start into this chunk.
+		// (No effect on first sentinel iteration, when start = {0,0}.)
+		for _, s := range x[done.x:start.x] {
+			ctext = append(ctext, "-"+s)
+			count.x++
+		}
+		for _, s := range y[done.y:start.y] {
+			ctext = append(ctext, "+"+s)
+			count.y++
+		}
+
+		// If we're not at EOF and have too few common lines,
+		// the chunk includes all the common lines and continues.
+		const C = 3 // number of context lines
+		if (end.x < len(x) || end.y < len(y)) &&
+			(end.x-start.x < C || (len(ctext) > 0 && end.x-start.x < 2*C)) {
+			for _, s := range x[start.x:end.x] {
+				ctext = append(ctext, " "+s)
+				count.x++
+				count.y++
+			}
+			done = end
+			continue
+		}
+
+		// End chunk with common lines for context.
+		if len(ctext) > 0 {
+			n := end.x - start.x
+			if n > C {
+				n = C
+			}
+			for _, s := range x[start.x : start.x+n] {
+				ctext = append(ctext, " "+s)
+				count.x++
+				count.y++
+			}
+			done = pair{start.x + n, start.y + n}
+
+			// Format and emit chunk.
+			// Convert line numbers to 1-indexed.
+			// Special case: empty file shows up as 0,0 not 1,0.
+			if count.x > 0 {
+				chunk.x++
+			}
+			if count.y > 0 {
+				chunk.y++
+			}
+			fmt.Fprintf(&out, "@@ -%d,%d +%d,%d @@\n", chunk.x, count.x, chunk.y, count.y)
+			for _, s := range ctext {
+				out.WriteString(s)
+			}
+			count.x = 0
+			count.y = 0
+			ctext = ctext[:0]
+		}
+
+		// If we reached EOF, we're done.
+		if end.x >= len(x) && end.y >= len(y) {
+			break
+		}
+
+		// Otherwise start a new chunk.
+		chunk = pair{end.x - C, end.y - C}
+		for _, s := range x[chunk.x:end.x] {
+			ctext = append(ctext, " "+s)
+			count.x++
+			count.y++
+		}
+		done = end
+	}
+
+	return out.Bytes()
+}
+
+// lines returns the lines in the file x, including newlines.
+// If the file does not end in a newline, one is supplied
+// along with a warning about the missing newline.
+func lines(x []byte) []string {
+	l := strings.SplitAfter(string(x), "\n")
+	if l[len(l)-1] == "" {
+		l = l[:len(l)-1]
+	} else {
+		// Treat last line as having a message about the missing newline attached,
+		// using the same text as BSD/GNU diff (including the leading backslash).
+		l[len(l)-1] += "\n\\ No newline at end of file\n"
+	}
+	return l
+}
+
+// tgs returns the pairs of indexes of the longest common subsequence
+// of unique lines in x and y, where a unique line is one that appears
+// once in x and once in y.
+//
+// The longest common subsequence algorithm is as described in
+// Thomas G. Szymanski, “A Special Case of the Maximal Common
+// Subsequence Problem,” Princeton TR #170 (January 1975),
+// available at https://research.swtch.com/tgs170.pdf.
+func tgs(x, y []string) []pair {
+	// Count the number of times each string appears in a and b.
+	// We only care about 0, 1, many, counted as 0, -1, -2
+	// for the x side and 0, -4, -8 for the y side.
+	// Using negative numbers now lets us distinguish positive line numbers later.
+	m := make(map[string]int)
+	for _, s := range x {
+		if c := m[s]; c > -2 {
+			m[s] = c - 1
+		}
+	}
+	for _, s := range y {
+		if c := m[s]; c > -8 {
+			m[s] = c - 4
+		}
+	}
+
+	// Now unique strings can be identified by m[s] = -1+-4.
+	//
+	// Gather the indexes of those strings in x and y, building:
+	//	xi[i] = increasing indexes of unique strings in x.
+	//	yi[i] = increasing indexes of unique strings in y.
+	//	inv[i] = index j such that x[xi[i]] = y[yi[j]].
+	var xi, yi, inv []int
+	for i, s := range y {
+		if m[s] == -1+-4 {
+			m[s] = len(yi)
+			yi = append(yi, i)
+		}
+	}
+	for i, s := range x {
+		if j, ok := m[s]; ok && j >= 0 {
+			xi = append(xi, i)
+			inv = append(inv, j)
+		}
+	}
+
+	// Apply Algorithm A from Szymanski's paper.
+	// In those terms, A = J = inv and B = [0, n).
+	// We add sentinel pairs {0,0}, and {len(x),len(y)}
+	// to the returned sequence, to help the processing loop.
+	J := inv
+	n := len(xi)
+	T := make([]int, n)
+	L := make([]int, n)
+	for i := range T {
+		T[i] = n + 1
+	}
+	for i := range n {
+		k := sort.Search(n, func(k int) bool {
+			return T[k] >= J[i]
+		})
+		T[k] = J[i]
+		L[i] = k + 1
+	}
+	k := 0
+	for _, v := range L {
+		if k < v {
+			k = v
+		}
+	}
+	seq := make([]pair, 2+k)
+	seq[1+k] = pair{len(x), len(y)} // sentinel at end
+	lastj := n
+	for i := n - 1; i >= 0; i-- {
+		if L[i] == k && J[i] < lastj {
+			seq[k] = pair{xi[i], yi[J[i]]}
+			k--
+		}
+	}
+	seq[0] = pair{0, 0} // sentinel at start
+	return seq
+}
--- a/grammar/internal/diff/diff_test.go
+++ b/grammar/internal/diff/diff_test.go
@@ -0,0 +1,44 @@
+// Copyright 2022 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package diff
+
+import (
+	"bytes"
+	"path/filepath"
+	"testing"
+
+	"golang.org/x/tools/txtar"
+)
+
+func clean(text []byte) []byte {
+	text = bytes.ReplaceAll(text, []byte("$\n"), []byte("\n"))
+	text = bytes.TrimSuffix(text, []byte("^D\n"))
+	return text
+}
+
+func Test(t *testing.T) {
+	files, _ := filepath.Glob("testdata/*.txt")
+	if len(files) == 0 {
+		t.Fatalf("no testdata")
+	}
+
+	for _, file := range files {
+		t.Run(filepath.Base(file), func(t *testing.T) {
+			a, err := txtar.ParseFile(file)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if len(a.Files) != 3 || a.Files[2].Name != "diff" {
+				t.Fatalf("%s: want three files, third named \"diff\"", file)
+			}
+			diffs := Diff(a.Files[0].Name, clean(a.Files[0].Data), a.Files[1].Name, clean(a.Files[1].Data))
+			want := clean(a.Files[2].Data)
+			if !bytes.Equal(diffs, want) {
+				t.Fatalf("%s: have:\n%s\nwant:\n%s\n%s", file,
+					diffs, want, Diff("have", diffs, "want", want))
+			}
+		})
+	}
+}
--- a/grammar/internal/diff/testdata/allnew.txt
+++ b/grammar/internal/diff/testdata/allnew.txt
@@ -0,0 +1,13 @@
+-- old --
+-- new --
+a
+b
+c
+-- diff --
+diff old new
+--- old
+++ new
+@@ -0,0 +1,3 @@
+a
+b
+c
--- a/grammar/internal/diff/testdata/allold.txt
+++ b/grammar/internal/diff/testdata/allold.txt
@@ -0,0 +1,13 @@
+-- old --
+a
+b
+c
+-- new --
+-- diff --
+diff old new
+--- old
+++ new
+@@ -1,3 +0,0 @@
+-a
+-b
+-c
--- a/grammar/internal/diff/testdata/basic.txt
+++ b/grammar/internal/diff/testdata/basic.txt
@@ -0,0 +1,35 @@
+Example from Hunt and McIlroy, “An Algorithm for Differential File Comparison.”
+https://www.cs.dartmouth.edu/~doug/diff.pdf
+
+-- old --
+a
+b
+c
+d
+e
+f
+g
+-- new --
+w
+a
+b
+x
+y
+z
+e
+-- diff --
+diff old new
+--- old
+++ new
+@@ -1,7 +1,7 @@
+w
+ a
+ b
+-c
+-d
+x
+y
+z
+ e
+-f
+-g
--- a/grammar/internal/diff/testdata/dups.txt
+++ b/grammar/internal/diff/testdata/dups.txt
@@ -0,0 +1,40 @@
+-- old --
+a
+
+b
+
+c
+
+d
+
+e
+
+f
+-- new --
+a
+
+B
+
+C
+
+d
+
+e
+
+f
+-- diff --
+diff old new
+--- old
+++ new
+@@ -1,8 +1,8 @@
+ a
+ $
+-b
+-
+-c
+B
+
+C
+ $
+ d
+ $
--- a/grammar/internal/diff/testdata/end.txt
+++ b/grammar/internal/diff/testdata/end.txt
@@ -0,0 +1,38 @@
+-- old --
+1
+2
+3
+4
+5
+6
+7
+eight
+nine
+ten
+eleven
+-- new --
+1
+2
+3
+4
+5
+6
+7
+8
+9
+10
+-- diff --
+diff old new
+--- old
+++ new
+@@ -5,7 +5,6 @@
+ 5
+ 6
+ 7
+-eight
+-nine
+-ten
+-eleven
+8
+9
+10
--- a/grammar/internal/diff/testdata/eof.txt
+++ b/grammar/internal/diff/testdata/eof.txt
@@ -0,0 +1,9 @@
+-- old --
+a
+b
+c^D
+-- new --
+a
+b
+c^D
+-- diff --
--- a/grammar/internal/diff/testdata/eof1.txt
+++ b/grammar/internal/diff/testdata/eof1.txt
@@ -0,0 +1,18 @@
+-- old --
+a
+b
+c
+-- new --
+a
+b
+c^D
+-- diff --
+diff old new
+--- old
+++ new
+@@ -1,3 +1,3 @@
+ a
+ b
+-c
+c
+\ No newline at end of file
--- a/grammar/internal/diff/testdata/eof2.txt
+++ b/grammar/internal/diff/testdata/eof2.txt
@@ -0,0 +1,18 @@
+-- old --
+a
+b
+c^D
+-- new --
+a
+b
+c
+-- diff --
+diff old new
+--- old
+++ new
+@@ -1,3 +1,3 @@
+ a
+ b
+-c
+\ No newline at end of file
+c
--- a/grammar/internal/diff/testdata/long.txt
+++ b/grammar/internal/diff/testdata/long.txt
@@ -0,0 +1,62 @@
+-- old --
+1
+2
+3
+4
+5
+6
+7
+8
+9
+10
+11
+12
+13
+14
+14½
+15
+16
+17
+18
+19
+20
+-- new --
+1
+2
+3
+4
+5
+6
+8
+9
+10
+11
+12
+13
+14
+17
+18
+19
+20
+-- diff --
+diff old new
+--- old
+++ new
+@@ -4,7 +4,6 @@
+ 4
+ 5
+ 6
+-7
+ 8
+ 9
+ 10
+@@ -12,9 +11,6 @@
+ 12
+ 13
+ 14
+-14½
+-15
+-16
+ 17
+ 18
+ 19
--- a/grammar/internal/diff/testdata/same.txt
+++ b/grammar/internal/diff/testdata/same.txt
@@ -0,0 +1,5 @@
+-- old --
+hello world
+-- new --
+hello world
+-- diff --
--- a/grammar/internal/diff/testdata/start.txt
+++ b/grammar/internal/diff/testdata/start.txt
@@ -0,0 +1,34 @@
+-- old --
+e
+pi
+4
+5
+6
+7
+8
+9
+10
+-- new --
+1
+2
+3
+4
+5
+6
+7
+8
+9
+10
+-- diff --
+diff old new
+--- old
+++ new
+@@ -1,5 +1,6 @@
+-e
+-pi
+1
+2
+3
+ 4
+ 5
+ 6
--- a/grammar/internal/diff/testdata/triv.txt
+++ b/grammar/internal/diff/testdata/triv.txt
@@ -0,0 +1,40 @@
+Another example from Hunt and McIlroy,
+“An Algorithm for Differential File Comparison.”
+https://www.cs.dartmouth.edu/~doug/diff.pdf
+
+Anchored diff gives up on finding anything,
+since there are no unique lines.
+
+-- old --
+a
+b
+c
+a
+b
+b
+a
+-- new --
+c
+a
+b
+a
+b
+c
+-- diff --
+diff old new
+--- old
+++ new
+@@ -1,7 +1,6 @@
+-a
+-b
+-c
+-a
+-b
+-b
+-a
+c
+a
+b
+a
+b
+c
--- a/grammar/jsonschema/decode.go
+++ b/grammar/jsonschema/decode.go
@@ -0,0 +1,171 @@
+package jsonschema
+
+import (
+	"bytes"
+	"encoding/json"
+	"errors"
+)
+
+// Schema holds a JSON schema.
+type Schema struct {
+	// Name is the name of the property. For the parent/root property, this
+	// is "root". For child properties, this is the name of the property.
+	Name string `json:"-"`
+
+	// Type is the type of the property.
+	//
+	// TODO: Union types (e.g. make this a []string).
+	Type string
+
+	// PrefixItems is a list of schemas for each item in a tuple. By
+	// default, the tuple is "closed." unless Items is set to true or a
+	// valid Schema.
+	PrefixItems []*Schema
+
+	// Items is the schema for each item in a list.
+	//
+	// If it is missing, or its JSON value is "null" or "false", it is nil.
+	// If the JSON value is "true", it is set to the empty Schema. If the
+	// JSON value is an object, it will be decoded as a Schema.
+	Items *Schema
+
+	// MinItems specifies the minimum number of items allowed in a list.
+	MinItems int
+
+	// MaxItems specifies the maximum number of items allowed in a list.
+	MaxItems int
+
+	// Properties is the schema for each property of an object.
+	Properties []*Schema
+
+	// Format is the format of the property. This is used to validate the
+	// property against a specific format.
+	//
+	// It is the callers responsibility to validate the property against
+	// the format.
+	Format string
+
+	// Minimum specifies the minimum value for numeric properties.
+	Minimum float64
+
+	// Maximum specifies the maximum value for numeric properties.
+	Maximum float64
+
+	// Enum is a list of valid values for the property.
+	Enum []json.RawMessage
+}
+
+func (s *Schema) UnmarshalJSON(data []byte) error {
+	type S Schema
+	w := struct {
+		Properties props
+		Items      items
+		*S
+	}{
+		S: (*S)(s),
+	}
+	if err := json.Unmarshal(data, &w); err != nil {
+		return err
+	}
+	if w.Items.set {
+		s.Items = &w.Items.Schema
+	}
+	s.Properties = w.Properties
+	return nil
+}
+
+type items struct {
+	Schema
+	set bool
+}
+
+func (s *items) UnmarshalJSON(data []byte) error {
+	switch b := data[0]; b {
+	case 't':
+		*s = items{set: true}
+	case '{':
+		type I items
+		if err := json.Unmarshal(data, (*I)(s)); err != nil {
+			return err
+		}
+		s.set = true
+	case 'n', 'f':
+	default:
+		return errors.New("invalid Items")
+	}
+	return nil
+}
+
+// EffectiveType returns the effective type of the schema. If the Type field is
+// not empty, it is returned; otherwise:
+//
+//   - If the schema has both Properties and Items, it returns an empty string.
+//   - If the schema has Properties, it returns "object".
+//   - If the schema has Items, it returns "array".
+//   - If the schema has neither Properties nor Items, it returns "value".
+//
+// The returned string is never empty.
+func (d *Schema) EffectiveType() string {
+	if d.Type == "" {
+		if len(d.Properties) > 0 {
+			return "object"
+		}
+		if len(d.PrefixItems) > 0 || d.Items != nil {
+			return "array"
+		}
+		return "value"
+	}
+	return d.Type
+}
+
+// props is an ordered list of properties. The order of the properties
+// is the order in which they were defined in the schema.
+type props []*Schema
+
+var _ json.Unmarshaler = (*props)(nil)
+
+func (v *props) UnmarshalJSON(data []byte) error {
+	if len(data) == 0 {
+		return nil
+	}
+	if data[0] != '{' {
+		return errors.New("expected object")
+	}
+
+	d := json.NewDecoder(bytes.NewReader(data))
+
+	// TODO(bmizerany): Consider DisallowUnknownFields. Currently, we, like
+	// llama.cpp, ignore unknown fields, which could be lead to unexpected
+	// behavior for clients of this package, since they may not be aware
+	// that "additionalFields", "itemsPrefix", etc, are being ignored.
+	//
+	// For now, just do what llama.cpp does.
+
+	t, err := d.Token()
+	if err != nil {
+		return err
+	}
+	if t != json.Delim('{') {
+		return errors.New("expected object")
+	}
+	for d.More() {
+		// Use the first token (map key) as the property name, then
+		// decode the rest of the object fields into a Schema and
+		// append.
+		t, err := d.Token()
+		if err != nil {
+			return err
+		}
+		if t == json.Delim('}') {
+			return nil
+		}
+		s := &Schema{
+			Name: t.(string),
+		}
+		if err := d.Decode(s); err != nil {
+			return err
+		}
+		*v = append(*v, s)
+	}
+	return nil
+}
--- a/grammar/jsonschema/decode_test.go
+++ b/grammar/jsonschema/decode_test.go
@@ -0,0 +1,104 @@
+package jsonschema
+
+import (
+	"encoding/json"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+const testSchemaBasic = `
+{
+  "properties": {
+    "tupleClosedEmpty":   { "prefixItems": [] },
+    "tupleClosedMissing": { "prefixItems": [{}] },
+    "tupleClosedNull":    { "prefixItems": [{}], "items": null },
+    "tupleClosedFalse":   { "prefixItems": [{}], "items": false },
+    "tupleOpenTrue":      { "prefixItems": [{}], "items": true },
+    "tupleOpenEmpty":     { "prefixItems": [{}], "items": {} },
+    "tupleOpenTyped":     { "prefixItems": [{}], "items": {"type": "boolean"} },
+    "tupleOpenMax":       { "prefixItems": [{}], "items": true, "maxItems": 3},
+
+    "array": { "items": {"type": "number"} },
+
+    "null": { "type": "null" },
+    "string": { "type": "string" },
+    "boolean": { "type": "boolean" }
+  }
+}
+`
+
+func TestSchemaUnmarshal(t *testing.T) {
+	var got *Schema
+	if err := json.Unmarshal([]byte(testSchemaBasic), &got); err != nil {
+		t.Fatalf("Unmarshal: %v", err)
+	}
+	want := &Schema{
+		Properties: []*Schema{
+			{Name: "tupleClosedEmpty", PrefixItems: []*Schema{}, Items: nil},
+			{Name: "tupleClosedMissing", PrefixItems: []*Schema{{}}, Items: nil},
+			{Name: "tupleClosedNull", PrefixItems: []*Schema{{}}, Items: nil},
+			{Name: "tupleClosedFalse", PrefixItems: []*Schema{{}}, Items: nil},
+
+			{Name: "tupleOpenTrue", PrefixItems: []*Schema{{}}, Items: &Schema{}},
+			{Name: "tupleOpenEmpty", PrefixItems: []*Schema{{}}, Items: &Schema{}},
+			{Name: "tupleOpenTyped", PrefixItems: []*Schema{{}}, Items: &Schema{Type: "boolean"}},
+			{Name: "tupleOpenMax", PrefixItems: []*Schema{{}}, Items: &Schema{}, MaxItems: 3},
+
+			{Name: "array", Items: &Schema{Type: "number"}},
+
+			{Name: "null", Type: "null"},
+			{Name: "string", Type: "string"},
+			{Name: "boolean", Type: "boolean"},
+		},
+	}
+
+	if diff := cmp.Diff(want, got); diff != "" {
+		t.Errorf("(-want, +got)\n%s", diff)
+	}
+}
+
+func TestEffectiveType(t *testing.T) {
+	const schema = `
+		{"properties": {
+			"o": {"type": "object"},
+			"a": {"type": "array"},
+			"n": {"type": "number"},
+			"s": {"type": "string"},
+			"z": {"type": "null"},
+			"b": {"type": "boolean"},
+
+			"t0": {"prefixItems": [{}], "items": {"type": "number"}},
+			"t1": {"items": {"type": "number"}, "maxItems": 3},
+
+			"v": {"maxItems": 3}
+		}}
+	`
+
+	var s *Schema
+	if err := json.Unmarshal([]byte(schema), &s); err != nil {
+		t.Fatalf("json.Unmarshal: %v", err)
+	}
+
+	var got []string
+	for _, p := range s.Properties {
+		got = append(got, p.EffectiveType())
+	}
+
+	want := strings.Fields(`
+		object
+		array
+		number
+		string
+		null
+		boolean
+		array
+		array
+		value
+	`)
+	if !reflect.DeepEqual(want, got) {
+		t.Errorf("\ngot:\n\t%v\nwant:\n\t%v", got, want)
+	}
+}
--- a/grammar/testdata/schemas.txt
+++ b/grammar/testdata/schemas.txt
@@ -0,0 +1,76 @@
+# This file holds tests for JSON schema to EBNF grammar conversions.
+#
+# The format is a JSON schema, followed by the expected EBNF grammar. Each test
+# MAY be preceded by a comment that describes the test (e.g. the test name), followed by
+# the JSON schema and the expected EBNF grammar. If no comment is present, the test
+# name the tests number in the file (e.g. "#0", "#1", etc.)
+#
+# Blank lines signify the end or start of a new test. Comments can be added
+# anywhere in the file, but they must be preceded by a '#' character and start at
+# the beginning of the line.
+
+# default
+{}
+root ::= value;
+
+{"properties": {}}
+root ::= value;
+
+# array
+{"properties": {"a": {"type": "array", "items": {"type": "string"}}}}
+root_0_tuple_0 ::= string;
+root_0         ::= "[" ( root_0_tuple_0 )* "]";
+root           ::= "{" "a" ":" root_0 "}";
+
+# array with nested array
+{"type": "array", "items": {"type": "array", "items": {"type": "string"}}}
+root_tuple_0_tuple_0 ::= string;
+root_tuple_0         ::= "[" ( root_tuple_0_tuple_0 )* "]";
+root                 ::= "[" ( root_tuple_0 )* "]";
+
+# object
+{"properties": {"e": {}}}
+root_0 ::= value;
+root   ::= "{" "e" ":" root_0 "}";
+
+# object with nested object
+{"properties": {"o": {"type": "object", "properties": {"e": {}}}}}
+root_0_0 ::= value;
+root_0   ::= "{" "e" ":" root_0_0 "}";
+root     ::= "{" "o" ":" root_0 "}";
+
+# boolean
+{"type": "boolean"}
+root ::= boolean;
+
+# number
+{"properties": {"n": {"type": "number", "minimum": 123, "maximum": 4567}}}
+root_0 ::= number;
+root   ::= "{" "n" ":" root_0 "}";
+
+# string
+{"type": "string"}
+root ::= string;
+
+# string with enum
+{"type": "string", "enum": ["a", "b", "c"]}
+root ::= ( "\"a\"" "|" "\"b\"" "|" "\"c\"" );
+
+# spaces in key
+{"properties": {"a b": {}}}
+root_0 ::= value;
+root   ::= "{" "a b" ":" root_0 "}";
+
+# issue7978
+{ "type": "object", "properties": { "steps": { "type": "array", "items": { "type": "object", "properties": { "explanation": { "type": "string" }, "output": { "type": "string" } }, "required": [ "explanation", "output" ], "additionalProperties": false } }, "final_answer": { "type": "string" } }, "required": [ "steps", "final_answer" ], "additionalProperties": false }
+root_0_tuple_0_0 ::= string;
+root_0_tuple_0_1 ::= string;
+root_0_tuple_0   ::= "{" "explanation" ":" root_0_tuple_0_0 "," "output" ":" root_0_tuple_0_1 "}";
+root_0           ::= "[" ( root_0_tuple_0 )* "]";
+root_1           ::= string;
+root             ::= "{" "steps" ":" root_0 "," "final_answer" ":" root_1 "}";
+
+# !! # special characters in key
+# !! {"properties": {"a!b": {}}}
+# !! !invalid character '!' in key
+# !! 
--- a/kvcache/cache.go
+++ b/kvcache/cache.go
@@ -1,54 +0,0 @@
-package kvcache
-
-import (
-	"errors"
-
-	"github.com/ollama/ollama/ml"
-)
-
-var (
-	ErrKvCacheFull  = errors.New("could not find a kv cache slot")
-	ErrNotSupported = errors.New("model does not support operation")
-)
-
-type Cache interface {
-	// ** used by model implementations **
-
-	// SetLayer sets the active layer of the cache
-	SetLayer(layer int)
-
-	// Get returns the history of key and value tensors plus a mask
-	//
-	// The shape of the tensors is documented in the specific
-	// cache implementation used.
-	Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor)
-
-	// Put stores a batch of key and value in the cache
-	//
-	// The shape of the tensors is documented in the specific
-	// cache implementation used.
-	Put(ctx ml.Context, key, value ml.Tensor)
-
-	// ** cache management **
-
-	// Init sets up runtime parameters
-	Init(backend ml.Backend, dtype ml.DType, capacity int32)
-
-	// Close closes the cache and frees resources associated with it
-	Close()
-
-	// StartForward is called before the start of the model's forward pass.
-	// For each token in the coming batch, there must be a corresponding
-	// entry in positions and seqs.
-	StartForward(ctx ml.Context, positions []int32, seqs []int) error
-
-	// CopyPrefix copies tokens in the range [0, len) from srcSeq to dstSeq
-	CopyPrefix(srcSeq, dstSeq int, len int32)
-
-	// Remove deletes tokens in the range [beginIndex, endIndex) from seq. Set
-	// endIndex to math.MaxInt32 to remove everything starting at beginIndex.
-	//
-	// If an error occurs, the entire context for the sequence should be
-	// removed by calling Remove(seq, 0, math.MaxInt32)
-	Remove(seq int, beginIndex, endIndex int32) error
-}
--- a/kvcache/causal.go
+++ b/kvcache/causal.go
@@ -1,455 +0,0 @@
-package kvcache
-
-import (
-	"errors"
-	"fmt"
-	"log/slog"
-	"math"
-	"slices"
-
-	"github.com/ollama/ollama/ml"
-)
-
-type shiftFn func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error)
-
-// Causal cache stores K and V tensors according to their position in the
-// sequence. Returns the history and a mask for attending to past tokens
-//
-// The tensors are of shape embed dim, kv heads, batch size
-// The mask is of shape history size, batch size
-type Causal struct {
-	DType      ml.DType
-	Capacity   int32
-	windowSize int32
-
-	// ** current forward pass **
-
-	// the active layer for Get and Put
-	curLayer int
-
-	// starting location for data storage for this batch
-	curLoc int
-
-	// size of the current batch
-	curBatchSize int
-
-	// mask of the cache as used by this batch
-	curMask ml.Tensor
-
-	// locations in the cache that are needed for this batch
-	curCellRange cellRange
-
-	// ** cache metadata **
-
-	// for each possible location in the cache, stores the position and set of sequences
-	// that reference the data there
-	cells []cacheCell
-
-	// maps from sequence to the range of locations where it is stored in the cache
-	cellRanges map[int]cellRange
-
-	// ** cache data storage **
-
-	shiftFn      shiftFn
-	backend      ml.Backend
-	cacheCtx     ml.Context
-	keys, values []ml.Tensor
-}
-
-type cacheCell struct {
-	pos       int32
-	sequences []int
-}
-
-type cellRange struct {
-	min int
-	max int
-}
-
-func NewCausalCache(shift shiftFn) *Causal {
-	return &Causal{windowSize: math.MaxInt32, shiftFn: shift}
-}
-
-func NewSWACache(windowSize int32, shift shiftFn) *Causal {
-	return &Causal{windowSize: windowSize, shiftFn: shift}
-}
-
-func (c *Causal) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
-	c.DType = dtype
-	c.Capacity = capacity
-	c.cells = make([]cacheCell, capacity)
-	c.cellRanges = make(map[int]cellRange)
-	c.backend = backend
-	c.cacheCtx = backend.NewContext()
-}
-
-func (c *Causal) Close() {
-	c.cacheCtx.Close()
-}
-
-func (c *Causal) StartForward(ctx ml.Context, positions []int32, seqs []int) error {
-	c.curBatchSize = len(positions)
-
-	var err error
-	c.curLoc, err = c.findStartLoc()
-	if errors.Is(err, ErrKvCacheFull) {
-		c.defrag()
-		c.curLoc, err = c.findStartLoc()
-	}
-	if err != nil {
-		return err
-	}
-
-	c.curCellRange = newRange()
-	for i, pos := range positions {
-		seq := seqs[i]
-
-		c.cells[c.curLoc+i] = cacheCell{pos: pos, sequences: []int{seq}}
-
-		seqRange, ok := c.cellRanges[seq]
-		if !ok {
-			seqRange = newRange()
-		}
-
-		if c.curLoc+i > seqRange.max {
-			seqRange.max = c.curLoc + i
-		}
-		if seqRange.max > c.curCellRange.max {
-			c.curCellRange.max = seqRange.max
-		}
-
-		if c.curLoc+i < seqRange.min {
-			seqRange.min = c.curLoc + i
-		}
-		if seqRange.min < c.curCellRange.min {
-			c.curCellRange.min = seqRange.min
-		}
-		c.cellRanges[seq] = seqRange
-	}
-
-	c.curMask, err = c.buildMask(ctx, positions, seqs)
-
-	return err
-}
-
-func newRange() cellRange {
-	return cellRange{
-		min: math.MaxInt,
-		max: 0,
-	}
-}
-
-// Find the first contiguous block of at least curBatchSize
-func (c *Causal) findStartLoc() (int, error) {
-	var start, count int
-	for i := range c.cells {
-		if len(c.cells[i].sequences) == 0 {
-			count++
-			if count >= c.curBatchSize {
-				return start, nil
-			}
-		} else {
-			start = i + 1
-			count = 0
-		}
-	}
-
-	return 0, fmt.Errorf("%w (length: %v)", ErrKvCacheFull, c.Capacity)
-}
-
-// Builds a mask of history x batch indicating whether for each token in the batch the
-// token in the history should apply. This is based on both the sequence and causality (the
-// position of the history is not ahead of the token in the batch).
-func (c *Causal) buildMask(ctx ml.Context, positions []int32, seqs []int) (ml.Tensor, error) {
-	// TODO(jessegross): This does not do padding, which is required for flash attention
-	len := c.curCellRange.max - c.curCellRange.min + 1
-	mask := make([]float32, c.curBatchSize*len)
-
-	for i := range c.curBatchSize {
-		for j := c.curCellRange.min; j <= c.curCellRange.max; j++ {
-			if !slices.Contains(c.cells[j].sequences, seqs[i]) || c.cells[j].pos > positions[i] ||
-				c.cells[j].pos < positions[i]-c.windowSize {
-				mask[i*len+(j-c.curCellRange.min)] = float32(math.Inf(-1))
-			}
-		}
-	}
-
-	return ctx.FromFloatSlice(mask, len, c.curBatchSize)
-}
-
-func moveCell(ctx ml.Context, objs []ml.Tensor, src, dst, len int) {
-	for _, obj := range objs {
-		if obj == nil {
-			continue
-		}
-
-		srcView := obj.View(ctx, obj.Stride(2)*src, obj.Dim(0)*obj.Dim(1)*len)
-		dstView := obj.View(ctx, obj.Stride(2)*dst, obj.Dim(0)*obj.Dim(1)*len)
-
-		ctx.Forward(srcView.Copy(ctx, dstView))
-	}
-}
-
-func (c *Causal) defrag() {
-	slog.Debug("defragmenting kv cache")
-
-	// Defrag strategy:
-	// - Search for empty holes at the beginning of the cache,
-	//   filling them with active data starting at the end
-	// - If there are contiguous elements that need to be moved,
-	//   combine them into a single operation by holding new moves
-	//   until we see that the next one is non-contiguous
-	// - Fill up the context with the maximum number of operations it
-	//   can hold then compute that and continue with a new context
-	//
-	// We could try to optimize placement by grouping blocks from
-	// the same sequences together but most likely the next forward
-	// pass will disrupt this anyways, so the real world benefit
-	// seems limited as this time.
-
-	ctx := c.backend.NewContext()
-
-	// For every move, 6 tensors are required per layer (2 views and a
-	// copy for each of k and v).
-	layers := 0
-	for _, key := range c.keys {
-		if key == nil {
-			continue
-		}
-		layers++
-	}
-
-	maxMoves := ctx.MaxTensors() / (6 * layers)
-	moves := 0
-
-	var pendingSrc, pendingDst, pendingLen int
-	src := len(c.cells) - 1
-
-	for dst := 0; dst < src; dst++ {
-		if len(c.cells[dst].sequences) == 0 {
-			for ; src > dst; src-- {
-				if len(c.cells[src].sequences) != 0 {
-					c.cells[dst] = c.cells[src]
-					c.cells[src] = cacheCell{}
-
-					if pendingLen > 0 {
-						if src == pendingSrc-pendingLen && dst == pendingDst+pendingLen {
-							pendingSrc = src
-							pendingLen++
-							break
-						} else {
-							moveCell(ctx, c.keys, pendingSrc, pendingDst, pendingLen)
-							moveCell(ctx, c.values, pendingSrc, pendingDst, pendingLen)
-							moves++
-						}
-					}
-
-					pendingSrc = src
-					pendingDst = dst
-					pendingLen = 1
-
-					break
-				}
-			}
-		}
-
-		if moves >= maxMoves {
-			ctx.Compute()
-			ctx.Close()
-			ctx = c.backend.NewContext()
-
-			moves = 0
-		}
-	}
-
-	if pendingLen > 0 {
-		moveCell(ctx, c.keys, pendingSrc, pendingDst, pendingLen)
-		moveCell(ctx, c.values, pendingSrc, pendingDst, pendingLen)
-		moves++
-	}
-
-	if moves > 0 {
-		ctx.Compute()
-	}
-	ctx.Close()
-
-	// Reset range metadata
-	for seq := range c.cellRanges {
-		seqRange := newRange()
-
-		for i, cell := range c.cells {
-			if slices.Contains(cell.sequences, seq) {
-				if i < seqRange.min {
-					seqRange.min = i
-				}
-				if i > seqRange.max {
-					seqRange.max = i
-				}
-			}
-		}
-
-		c.cellRanges[seq] = seqRange
-	}
-}
-
-func (c *Causal) SetLayer(layer int) {
-	if layer >= len(c.keys) {
-		c.keys = append(c.keys, make([]ml.Tensor, layer-len(c.keys)+1)...)
-		c.values = append(c.values, make([]ml.Tensor, layer-len(c.values)+1)...)
-	}
-
-	c.curLayer = layer
-}
-
-func (c *Causal) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
-	key := c.keys[c.curLayer]
-	value := c.values[c.curLayer]
-
-	key = key.View(ctx, key.Stride(2)*c.curCellRange.min,
-		key.Dim(0), key.Stride(1),
-		key.Dim(1), key.Stride(2),
-		c.curMask.Dim(0),
-	)
-
-	value = value.View(ctx, key.Stride(2)*c.curCellRange.min,
-		value.Dim(0), value.Stride(1),
-		value.Dim(1), value.Stride(2),
-		c.curMask.Dim(0),
-	)
-
-	return key, value, c.curMask
-}
-
-func (c *Causal) Put(ctx ml.Context, key, value ml.Tensor) {
-	if c.curBatchSize != key.Dim(2) {
-		panic(fmt.Errorf("inconsistent batch sizes (layer: %v, batch size: %v layer batch size: %v)", c.curLayer, c.curBatchSize, key.Dim(2)))
-	}
-
-	if c.keys[c.curLayer] == nil || c.values[c.curLayer] == nil {
-		c.keys[c.curLayer] = c.cacheCtx.Zeros(c.DType, key.Dim(0), key.Dim(1), int(c.Capacity))
-		c.values[c.curLayer] = c.cacheCtx.Zeros(c.DType, value.Dim(0), value.Dim(1), int(c.Capacity))
-	}
-
-	ctx.Forward(key.Copy(ctx, c.keys[c.curLayer].View(ctx, c.keys[c.curLayer].Stride(2)*c.curLoc, key.Dim(0)*key.Dim(1)*key.Dim(2))))
-	ctx.Forward(value.Copy(ctx, c.values[c.curLayer].View(ctx, c.values[c.curLayer].Stride(2)*c.curLoc, value.Dim(0)*value.Dim(1)*value.Dim(2))))
-}
-
-func (c *Causal) CopyPrefix(srcSeq, dstSeq int, len int32) {
-	seqRange := newRange()
-
-	for i := range c.cells {
-		// Remove the contents of dstSeq so that we only have the copied prefix, metadata will be reset at the end
-		if slices.Contains(c.cells[i].sequences, dstSeq) {
-			c.cells[i].sequences = slices.DeleteFunc(c.cells[i].sequences, func(s int) bool { return s == dstSeq })
-		}
-
-		if slices.Contains(c.cells[i].sequences, srcSeq) && c.cells[i].pos < len {
-			c.cells[i].sequences = append(c.cells[i].sequences, dstSeq)
-			if i < seqRange.min {
-				seqRange.min = i
-			}
-			if i > seqRange.max {
-				seqRange.max = i
-			}
-		}
-	}
-
-	c.cellRanges[dstSeq] = seqRange
-}
-
-func (c *Causal) shift(seq int, beginIndex, offset int32) error {
-	if c.shiftFn == nil {
-		return ErrNotSupported
-	}
-
-	ctx := c.backend.NewContext()
-	defer ctx.Close()
-
-	seqRange := c.cellRanges[seq]
-	size := seqRange.max - seqRange.min + 1
-
-	offsets := make([]int32, size)
-	for i := range offsets {
-		cell := c.cells[seqRange.min+i]
-
-		if slices.Contains(cell.sequences, seq) && cell.pos >= beginIndex {
-			offsets[i] = offset
-		}
-	}
-
-	kShift, err := ctx.FromIntSlice(offsets, len(offsets))
-	if err != nil {
-		return err
-	}
-
-	for i, key := range c.keys {
-		if key == nil {
-			continue
-		}
-
-		key = key.View(ctx, key.Stride(2)*seqRange.min,
-			key.Dim(0), key.Stride(1),
-			key.Dim(1), key.Stride(2),
-			size,
-		)
-
-		roped, err := c.shiftFn(ctx, i, key, kShift)
-		if err != nil {
-			return err
-		}
-
-		ctx.Forward(roped.Copy(ctx, key))
-	}
-
-	ctx.Compute()
-
-	return nil
-}
-
-func (c *Causal) Remove(seq int, beginIndex, endIndex int32) error {
-	var offset int32
-	if endIndex != math.MaxInt32 {
-		offset = beginIndex - endIndex
-	}
-
-	seqRange := newRange()
-
-	for i := range c.cells {
-		if slices.Contains(c.cells[i].sequences, seq) {
-			if c.cells[i].pos >= beginIndex && c.cells[i].pos < endIndex {
-				c.cells[i].sequences = slices.DeleteFunc(c.cells[i].sequences, func(s int) bool { return s == seq })
-			} else {
-				if c.cells[i].pos >= endIndex {
-					if slices.ContainsFunc(c.cells[i].sequences, func(s int) bool { return s != seq }) {
-						// TODO(jessegross): Need to be careful about data shared between sequences
-						return errors.New("shifting on cells shared by multiple sequences not yet implemented")
-					}
-
-					c.cells[i].pos += offset
-				}
-				if i < seqRange.min {
-					seqRange.min = i
-				}
-				if i > seqRange.max {
-					seqRange.max = i
-				}
-			}
-		}
-	}
-
-	if seqRange == newRange() {
-		delete(c.cellRanges, seq)
-		return nil
-	}
-
-	c.cellRanges[seq] = seqRange
-
-	if endIndex != math.MaxInt32 {
-		err := c.shift(seq, endIndex+offset, offset)
-		if err != nil {
-			return err
-		}
-	}
-
-	return nil
-}
--- a/kvcache/causal_test.go
+++ b/kvcache/causal_test.go
@@ -1,510 +0,0 @@
-package kvcache
-
-import (
-	"math"
-	"slices"
-	"testing"
-
-	"github.com/ollama/ollama/ml"
-)
-
-type testCase struct {
-	name          string
-	in            []float32
-	inShape       []int
-	seqs          []int
-	pos           []int32
-	expected      []float32
-	expectedShape []int
-	expectedMask  []float32
-}
-
-func TestStore(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(nil)
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{111, 211, 121, 221, 131, 231, 112, 212, 122, 222, 132, 232, 113, 213, 123, 223, 133, 233, 114, 214, 124, 224, 134, 234},
-			inShape:       []int{2, 3, 4},
-			seqs:          []int{0, 0, 0, 0},
-			pos:           []int32{0, 1, 2, 3},
-			expected:      []float32{111, 211, 121, 221, 131, 231, 112, 212, 122, 222, 132, 232, 113, 213, 123, 223, 133, 233, 114, 214, 124, 224, 134, 234},
-			expectedShape: []int{2, 3, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), 0, 0, 0, 0},
-		},
-		{
-			name:          "SecondBatch",
-			in:            []float32{115, 215, 125, 225, 135, 235},
-			inShape:       []int{2, 3, 1},
-			seqs:          []int{0},
-			pos:           []int32{4},
-			expected:      []float32{111, 211, 121, 221, 131, 231, 112, 212, 122, 222, 132, 232, 113, 213, 123, 223, 133, 233, 114, 214, 124, 224, 134, 234, 115, 215, 125, 225, 135, 235},
-			expectedShape: []int{2, 3, 5},
-			expectedMask:  []float32{0, 0, 0, 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestSWA(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewSWACache(1, nil)
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF32, 16)
-
-	tests := []testCase{
-		{
-			name:          "SlidingWindow",
-			in:            []float32{1, 2, 3, 4},
-			inShape:       []int{1, 1, 4},
-			seqs:          []int{0, 0, 0, 0},
-			pos:           []int32{0, 1, 2, 3},
-			expected:      []float32{1, 2, 3, 4},
-			expectedShape: []int{1, 1, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestSequences(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(nil)
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{1, 2, 3, 4},
-			inShape:       []int{1, 1, 4},
-			seqs:          []int{0, 0, 1, 1},
-			pos:           []int32{0, 1, 0, 1},
-			expected:      []float32{1, 2, 3, 4},
-			expectedShape: []int{1, 1, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
-		},
-		{
-			name:          "SecondBatch",
-			in:            []float32{5, 6},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{0, 1},
-			pos:           []int32{2, 2},
-			expected:      []float32{1, 2, 3, 4, 5, 6},
-			expectedShape: []int{1, 1, 6},
-			expectedMask:  []float32{0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestRemove(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-		return key.Add(ctx, shift), nil
-	})
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{1, 2, 3, 4},
-			inShape:       []int{1, 1, 4},
-			seqs:          []int{0, 0, 1, 1},
-			pos:           []int32{0, 1, 0, 1},
-			expected:      []float32{1, 2, 3, 4},
-			expectedShape: []int{1, 1, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-
-	err := cache.Remove(0, 1, math.MaxInt32)
-	if err != nil {
-		panic(err)
-	}
-
-	tests = []testCase{
-		{
-			name:          "RemoveEnd",
-			in:            []float32{5, 6},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{0, 1},
-			pos:           []int32{1, 2},
-			expected:      []float32{1, 2, 3, 4, 5, 6},
-			expectedShape: []int{1, 1, 6},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-
-	err = cache.Remove(0, 0, 1)
-	if err != nil {
-		panic(err)
-	}
-
-	tests = []testCase{
-		{
-			name:          "RemoveMiddle",
-			in:            []float32{7, 8},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{0, 0},
-			pos:           []int32{1, 2},
-			expected:      []float32{7, 8, 3, 4, 4},
-			expectedShape: []int{1, 1, 5},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestDefrag(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-		return key.Add(ctx, shift), nil
-	})
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
-			inShape:       []int{1, 1, 16},
-			seqs:          []int{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
-			pos:           []int32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
-			expected:      []float32{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
-			expectedShape: []int{1, 1, 16},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-
-	err := cache.Remove(0, 2, 4)
-	if err != nil {
-		panic(err)
-	}
-
-	err = cache.Remove(0, 13, math.MaxInt32)
-	if err != nil {
-		panic(err)
-	}
-
-	tests = []testCase{
-		{
-			name:          "Defrag",
-			in:            []float32{17, 18, 19},
-			inShape:       []int{1, 1, 3},
-			seqs:          []int{0, 0, 0},
-			pos:           []int32{16, 17, 18},
-			expected:      []float32{1, 2, 12, 13, 3, 4, 5, 6, 7, 8, 9, 10, 11, 17, 18, 19},
-			expectedShape: []int{1, 1, 16},
-			expectedMask:  []float32{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestCopy(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) { return key, nil })
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{1, 2, 3, 4},
-			inShape:       []int{1, 1, 4},
-			seqs:          []int{0, 0, 0, 0},
-			pos:           []int32{0, 1, 2, 3},
-			expected:      []float32{1, 2, 3, 4},
-			expectedShape: []int{1, 1, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), 0, 0, 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-
-	cache.CopyPrefix(0, 1, 2)
-
-	tests = []testCase{
-		{
-			name:          "Copy",
-			in:            []float32{5, 6},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{1, 1},
-			pos:           []int32{3, 4},
-			expected:      []float32{1, 2, 3, 4, 5, 6},
-			expectedShape: []int{1, 1, 6},
-			expectedMask:  []float32{0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase) {
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			context := backend.NewContext()
-			defer context.Close()
-
-			err := cache.StartForward(context, test.pos, test.seqs)
-			if err != nil {
-				panic(err)
-			}
-
-			cache.SetLayer(0)
-			tensor, _ := context.FromFloatSlice(test.in, test.inShape...)
-			cache.Put(context, tensor, tensor)
-
-			out, _, mask := cache.Get(context)
-
-			context.Forward(out)
-			context.Forward(mask)
-			context.Compute(out, mask)
-
-			if !slices.Equal(out.Floats(), test.expected) || !slices.Equal(out.Shape(), test.expectedShape) || !slices.Equal(mask.Floats(), test.expectedMask) {
-				t.Errorf("TestCache: have %v (shape %v); want %v (shape %v); mask: have %v (shape %v) want %v", out.Floats(), out.Shape(), test.expected, test.expectedShape, mask.Floats(), mask.Shape(), test.expectedMask)
-			}
-		})
-	}
-}
-
-type testBackend struct{}
-
-func (b *testBackend) Config() ml.Config {
-	panic("not implemented")
-}
-
-func (b *testBackend) Get(name string) ml.Tensor {
-	panic("not implemented")
-}
-
-func (b *testBackend) NewContext() ml.Context {
-	return &testContext{}
-}
-
-func (b *testBackend) SystemInfo() string {
-	return "not implemented"
-}
-
-type testContext struct{}
-
-func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
-	total := 0
-
-	if len(shape) > 0 {
-		total = 1
-		for _, s := range shape {
-			total *= s
-		}
-	}
-
-	return &testTensor{dtype: dtype, elementSize: 4, data: make([]float32, total), shape: shape}
-}
-
-func (c *testContext) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
-	t := c.Zeros(ml.DTypeF32, shape...).(*testTensor)
-
-	copy(t.data, s)
-
-	return t, nil
-}
-
-func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
-	f := make([]float32, len(s))
-	for i := range f {
-		f[i] = float32(s[i])
-	}
-
-	out, _ := c.FromFloatSlice(f, shape...)
-	out.(*testTensor).dtype = ml.DTypeI32
-
-	return out, nil
-}
-
-func (c *testContext) Forward(ml.Tensor) {}
-
-func (c *testContext) Compute(...ml.Tensor) {}
-
-func (c *testContext) MaxTensors() int {
-	return 10
-}
-
-func (c *testContext) Close() {}
-
-type testTensor struct {
-	dtype       ml.DType
-	elementSize int
-	data        []float32
-	shape       []int
-}
-
-func (t *testTensor) Dim(n int) int {
-	return t.shape[n]
-}
-
-func (t *testTensor) Stride(n int) int {
-	stride := t.elementSize
-	for i := range n {
-		stride *= t.shape[i]
-	}
-
-	return stride
-}
-
-func (t *testTensor) Shape() []int {
-	return t.shape
-}
-
-func (t *testTensor) DType() ml.DType {
-	return t.dtype
-}
-
-func (t *testTensor) Bytes() []byte {
-	panic("not implemented")
-}
-
-func (t *testTensor) Floats() []float32 {
-	out := make([]float32, len(t.data))
-	copy(out, t.data)
-	return out
-}
-
-func (t *testTensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	out := ctx.Zeros(t.DType(), t.Shape()...).(*testTensor)
-
-	for i := range out.data {
-		out.data[i] = t.data[i] + t2.(*testTensor).data[i]
-	}
-
-	return out
-}
-
-func (t *testTensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) MulmatFullPrec(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Softmax(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) LayerNorm(ctx ml.Context, weight, bias ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RMSNorm(ctx ml.Context, weight ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Scale(ctx ml.Context, s float64) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim uint32, base, scale float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Tanh(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) GELU(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) SILU(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
-	offset /= t.elementSize
-
-	var s []int
-
-	switch len(shape) {
-	case 1:
-		s = []int{shape[0]}
-	case 5:
-		s = []int{shape[0], shape[2], shape[4]}
-	default:
-		panic("unsupported number of dimensions")
-	}
-
-	context := &testContext{}
-
-	view := context.Zeros(t.dtype, s...).(*testTensor)
-	view.data = t.data[offset : offset+len(view.data)]
-
-	return view
-}
-
-func (t *testTensor) Permute(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Contiguous(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Unpad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Stack(ctx ml.Context, dim int, s ...ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Concat(ctx ml.Context, t2 ml.Tensor, dim int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Rows(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Copy(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	copy(t2.(*testTensor).data, t.data)
-	return nil
-}
--- a/kvcache/encoder.go
+++ b/kvcache/encoder.go
@@ -1,97 +0,0 @@
-package kvcache
-
-import (
-	"github.com/ollama/ollama/ml"
-)
-
-// Encoder cache stores K and V tensors that are position independent
-//
-// The tensors can be of any shape and will be returned as they were stored
-// The mask is currently always nil
-//
-// Not currently safe for multiple sequences
-type EncoderCache struct {
-	// ** current forward pass **
-
-	// the active layer for Get and Put
-	curLayer int
-
-	// if something is stored during this pass, this
-	// will be the position (but there is no guarantee
-	// anything will be stored)
-	curPos int32
-
-	// ** cache metadata **
-
-	// was something stored in the cache?
-	encoderCached bool
-
-	// position of the cached data
-	encoderPos int32
-
-	// ** cache data storage **
-
-	cacheCtx     ml.Context
-	keys, values []ml.Tensor
-}
-
-func NewEncoderCache() *EncoderCache {
-	return &EncoderCache{}
-}
-
-func (c *EncoderCache) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
-	c.cacheCtx = backend.NewContext()
-}
-
-func (c *EncoderCache) Close() {
-	c.cacheCtx.Close()
-}
-
-func (c *EncoderCache) StartForward(ctx ml.Context, positions []int32, seqs []int) error {
-	// The image is always in the first position
-	c.curPos = positions[0]
-
-	return nil
-}
-
-func (c *EncoderCache) SetLayer(layer int) {
-	if layer >= len(c.keys) {
-		c.keys = append(c.keys, make([]ml.Tensor, layer-len(c.keys)+1)...)
-		c.values = append(c.values, make([]ml.Tensor, layer-len(c.values)+1)...)
-	}
-
-	c.curLayer = layer
-}
-
-func (c *EncoderCache) EncoderCached() bool {
-	return c.encoderCached
-}
-
-func (c *EncoderCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
-	return c.keys[c.curLayer], c.values[c.curLayer], nil
-}
-
-func (c *EncoderCache) Put(ctx ml.Context, key, value ml.Tensor) {
-	c.encoderPos = c.curPos
-	c.encoderCached = true
-
-	if c.keys[c.curLayer] == nil || c.values[c.curLayer] == nil {
-		c.keys[c.curLayer] = c.cacheCtx.Zeros(key.DType(), key.Shape()...)
-		c.values[c.curLayer] = c.cacheCtx.Zeros(value.DType(), value.Shape()...)
-	}
-
-	ctx.Forward(key.Copy(ctx, c.keys[c.curLayer]))
-	ctx.Forward(value.Copy(ctx, c.values[c.curLayer]))
-}
-
-func (c *EncoderCache) CopyPrefix(srcSeq, dstSeq int, len int32) {
-	panic("encoder cache does not support multiple sequences")
-}
-
-func (c *EncoderCache) Remove(seq int, beginIndex, endIndex int32) error {
-	if c.encoderPos >= beginIndex && c.encoderPos < endIndex {
-		c.encoderCached = false
-	}
-
-	return nil
-}
--- a/kvcache/wrapper.go
+++ b/kvcache/wrapper.go
@@ -1,93 +0,0 @@
-package kvcache
-
-import (
-	"math"
-
-	"github.com/ollama/ollama/ml"
-)
-
-// Wrapper cache is a container for multiple types of caches,
-// such as for the encoding and decoding portions of a model.
-type WrapperCache struct {
-	// caches we are wrapping
-	caches []Cache
-
-	// cache to be used for this layer
-	curType int
-}
-
-func NewWrapperCache(caches ...Cache) *WrapperCache {
-	return &WrapperCache{
-		caches: caches,
-	}
-}
-
-func (c *WrapperCache) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
-	for _, cache := range c.caches {
-		cache.Init(backend, dtype, capacity)
-	}
-}
-
-func (c *WrapperCache) Close() {
-	for _, cache := range c.caches {
-		cache.Close()
-	}
-}
-
-func (c *WrapperCache) StartForward(ctx ml.Context, positions []int32, seqs []int) error {
-	for i, cache := range c.caches {
-		err := cache.StartForward(ctx, positions, seqs)
-		if err != nil {
-			// unwind on error - Remove with endIndex set to math.MaxInt32 does not fail
-			for j := i - 1; j >= 0; j-- {
-				for k := range positions {
-					_ = c.caches[j].Remove(seqs[k], positions[k], math.MaxInt32)
-				}
-			}
-			return err
-		}
-	}
-
-	c.curType = 0
-	return nil
-}
-
-func (c *WrapperCache) SetLayer(layer int) {
-	for _, cache := range c.caches {
-		cache.SetLayer(layer)
-	}
-}
-
-func (c *WrapperCache) SetLayerType(layerType int) {
-	c.curType = layerType
-}
-
-func (c *WrapperCache) UnderlyingCache() Cache {
-	return c.caches[c.curType]
-}
-
-func (c *WrapperCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
-	return c.caches[c.curType].Get(ctx)
-}
-
-func (c *WrapperCache) Put(ctx ml.Context, key, value ml.Tensor) {
-	c.caches[c.curType].Put(ctx, key, value)
-}
-
-func (c *WrapperCache) CopyPrefix(srcSeq, dstSeq int, len int32) {
-	for _, cache := range c.caches {
-		cache.CopyPrefix(srcSeq, dstSeq, len)
-	}
-}
-
-func (c *WrapperCache) Remove(seq int, beginIndex, endIndex int32) error {
-	// If the one of these fails, the caller is supposed to retry with endIndex set to math.MaxInt32, which should not fail
-	for _, cache := range c.caches {
-		err := cache.Remove(seq, beginIndex, endIndex)
-		if err != nil {
-			return err
-		}
-	}
-
-	return nil
-}
--- a/llama/README.md
+++ b/llama/README.md
@@ -1,52 +1,156 @@
 # `llama`

-This package provides Go bindings to [llama.cpp](https://github.com/ggerganov/llama.cpp).
+This package integrates the [llama.cpp](https://github.com/ggerganov/llama.cpp) library as a Go package and makes it easy to build it with tags for different CPU and GPU processors.
+
+Supported:
+
+- [x] CPU
+- [x] avx, avx2
+- [x] macOS Metal
+- [x] Windows CUDA
+- [x] Windows ROCm
+- [x] Linux CUDA
+- [x] Linux ROCm
+- [x] Llava
+
+Extra build steps are required for CUDA and ROCm on Windows since `nvcc` and `hipcc` both require using msvc as the host compiler. For these shared libraries are created:
+
+- `ggml_cuda.dll` on Windows or `ggml_cuda.so` on Linux
+- `ggml_hipblas.dll` on Windows or `ggml_hipblas.so` on Linux
+
+> Note: it's important that memory is allocated and freed by the same compiler (e.g. entirely by code compiled with msvc or mingw). Issues from this should be rare, but there are some places where pointers are returned by the CUDA or HIP runtimes and freed elsewhere, causing a a crash. In a future change the same runtime should be used in both cases to avoid crashes.
+
+## Building
+
+```
+go build .
+```
+
+### AVX
+
+```shell
+go build -tags avx .
+```
+
+### AVX2
+
+```shell
+# go doesn't recognize `-mfma` as a valid compiler flag
+# see https://github.com/golang/go/issues/17895
+go env -w "CGO_CPPFLAGS_ALLOW=-mfma|-mf16c"
+go build -tags=avx,avx2 .
+```
+
+## Linux
+
+### CUDA
+
+Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive):
+
+```shell
+make ggml_cuda.so
+go build -tags avx,cuda .
+```
+
+### ROCm
+
+Install [ROCm](https://rocm.docs.amd.com/en/latest/).
+
+```shell
+make ggml_hipblas.so
+go build -tags avx,rocm .
+```
+
+## Windows
+
+Download [w64devkit](https://github.com/skeeto/w64devkit/releases/latest) for a simple MinGW development environment.
+
+### CUDA
+
+Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive) then build the cuda code:
+
+```shell
+make ggml_cuda.dll
+go build -tags avx,cuda .
+```
+
+### ROCm
+
+Install [ROCm](https://rocm.docs.amd.com/en/latest/).
+
+```shell
+make ggml_hipblas.dll
+go build -tags avx,rocm .
+```
+
+## Building runners
+
+```shell
+# build all runners for this platform
+make -j
+```

 ## Vendoring

-Ollama vendors [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [ggml](https://github.com/ggerganov/llama.cpp/tree/master/ggml/src). While we generally strive to contribute changes back upstream to avoid drift, we carry a small set of patches which are applied to the tracking commit.
+Ollama currently vendors [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [ggml](https://github.com/ggerganov/ggml) through a vendoring model. While we generally strive to contribute changes back upstream to avoid drift, we cary a small set of patches which are applied to the tracking commit. A set of make targets are available to aid developers in updating to a newer tracking commit, or to work on changes.

 If you update the vendoring code, start by running the following command to establish the tracking llama.cpp repo in the `./vendor/` directory.

-```shell
-make -f Makefile.sync apply-patches
+```
+make apply-patches
 ```

 ### Updating Base Commit

 **Pin to new base commit**

-To change the base commit, update `FETCH_HEAD` in Makefile.sync.
+To update to a newer base commit, select the upstream git tag or commit and update `llama/vendoring`
+
+#### Applying patches

 When updating to a newer base commit, the existing patches may not apply cleanly and require manual merge resolution.

 Start by applying the patches. If any of the patches have conflicts, the `git am` will stop at the first failure.

-```shell
-make -f Makefile.sync apply-patches
+```
+make apply-patches
 ```

-If there are conflicts, you will see an error message. Resolve the conflicts in `./vendor/`, and continue the patch series with `git am --continue` and rerun `make -f Makefile.sync apply-patches`. Repeat until all patches are successfully applied.
+If you see an error message about a conflict, go into the `./vendor/` directory, and perform merge resolution using your preferred tool to the patch commit which failed. Save the file(s) and continue the patch series with `git am --continue` . If any additional patches fail, follow the same pattern until the full patch series is applied. Once finished, run a final `create-patches` and `sync` target to ensure everything is updated.

-Once all patches are applied, commit the changes to the tracking repository.
-
-```shell
-make -f Makefile.sync format-patches sync
 ```
+make create-patches sync
+```
+
+Build and test Ollama, and make any necessary changes to the Go code based on the new base commit. Submit your PR to the Ollama repo.

 ### Generating Patches

 When working on new fixes or features that impact vendored code, use the following model. First get a clean tracking repo with all current patches applied:

-```shell
-make -f Makefile.sync clean apply-patches
 ```
+make apply-patches
+```
+
+Now edit the upstream native code in the `./vendor/` directory. You do not need to commit every change in order to build, a dirty working tree in the tracking repo is OK while developing. Simply save in your editor, and run the following to refresh the vendored code with your changes, build the backend(s) and build ollama:
+
+```
+make sync
+make -j 8
+go build .
+```
+
+> [!IMPORTANT]
+> Do **NOT** run `apply-patches` while you're iterating as that will reset the tracking repo. It will detect a dirty tree and abort, but if your tree is clean and you accidentally ran this target, use `git reflog` to recover your commit(s).

 Iterate until you're ready to submit PRs. Once your code is ready, commit a change in the `./vendor/` directory, then generate the patches for ollama with

-```shell
-make -f Makefile.sync format-patches
 ```
+make create-patches
+```
+
+> [!IMPORTANT]
+> Once you have completed this step, it is safe to run `apply-patches` since your change is preserved in the patches.

 In your `./vendor/` directory, create a branch, and cherry-pick the new commit to that branch, then submit a PR upstream to llama.cpp.

--- a/llama/build-info.cpp
+++ b/llama/build-info.cpp
@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "46e3556e01b824e52395fb050b29804b6cff2a7c";
+char const *LLAMA_COMMIT = "ba1cb19cdd0d92e012e0f6e009e0620f854b6afd";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";
--- a/llama/build-info.cpp.in
+++ b/llama/build-info.cpp.in
@@ -1,4 +0,0 @@
-int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "@FETCH_HEAD@";
-char const *LLAMA_COMPILER = "";
-char const *LLAMA_BUILD_TARGET = "";
--- a/llama/llama.cpp/examples/llava/clip.cpp
+++ b/llama/llama.cpp/examples/llava/clip.cpp
@@ -1235,15 +1235,35 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
        }
    }

-    ggml_backend_t backend = ggml_backend_init_best();
-    if (backend == nullptr) {
-        LOG_ERR("%s: failed to initialize backend\n", __func__);
-        clip_free(new_clip);
-        gguf_free(ctx);
-        return nullptr;
+#ifdef GGML_USE_CUDA
+   new_clip->backend = ggml_backend_cuda_init(0);
+   LOG_INF("%s: CLIP using CUDA backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_METAL
+   new_clip->backend = ggml_backend_metal_init();
+   LOG_INF("%s: CLIP using Metal backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_CANN
+   new_clip->backend = ggml_backend_cann_init(0);
+   LOG_INF("%s: CLIP using CANN backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_VULKAN
+   new_clip->backend = ggml_backend_vk_init(0);
+   LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_SYCL
+   new_clip->backend = ggml_backend_sycl_init(0);
+   LOG_INF("%s: CLIP using SYCL backend\n", __func__);
+#endif
+
+    if (!new_clip->backend) {
+        new_clip->backend = ggml_backend_cpu_init();
+        LOG_INF("%s: CLIP using CPU backend\n", __func__);
    }
-    LOG_INF("%s: using %s backend\n", __func__, ggml_backend_name(backend));
-    new_clip->backend = backend;

    // model size and capabilities
    {
--- a/llama/llama.cpp/src/llama.go
+++ b/llama/llama.cpp/src/llama.go
@@ -3,6 +3,5 @@ package llama
 // #cgo CXXFLAGS: -std=c++17
 // #cgo CPPFLAGS: -I${SRCDIR}/../include
 // #cgo CPPFLAGS: -I${SRCDIR}/../../../ml/backend/ggml/ggml/include
-// #cgo windows CPPFLAGS: -D_WIN32_WINNT=0x0602
 import "C"
 import _ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
--- a/llama/llama.go
+++ b/llama/llama.go
@@ -199,25 +199,21 @@ func (c *Context) KvCacheDefrag() {

 // Get the embeddings for a sequence id
 func (c *Context) GetEmbeddingsSeq(seqId int) []float32 {
-	e := unsafe.Pointer(C.llama_get_embeddings_seq(c.c, C.int(seqId)))
-	if e == nil {
+	embeddings := unsafe.Pointer(C.llama_get_embeddings_seq(c.c, C.int(seqId)))
+	if embeddings == nil {
 		return nil
 	}

-	embeddings := make([]float32, c.Model().NEmbd())
-	_ = copy(embeddings, unsafe.Slice((*float32)(e), c.Model().NEmbd()))
-	return embeddings
+	return unsafe.Slice((*float32)(embeddings), c.Model().NEmbd())
 }

 func (c *Context) GetEmbeddingsIth(i int) []float32 {
-	e := unsafe.Pointer(C.llama_get_embeddings_ith(c.c, C.int32_t(i)))
-	if e == nil {
+	embeddings := unsafe.Pointer(C.llama_get_embeddings_ith(c.c, C.int32_t(i)))
+	if embeddings == nil {
 		return nil
 	}

-	embeddings := make([]float32, c.Model().NEmbd())
-	_ = copy(embeddings, unsafe.Slice((*float32)(e), c.Model().NEmbd()))
-	return embeddings
+	return unsafe.Slice((*float32)(embeddings), c.Model().NEmbd())
 }

 type ModelParams struct {
--- a/llama/mllama.cpp
+++ b/llama/mllama.cpp
@@ -558,15 +558,30 @@ struct mllama_ctx *mllama_model_load(const char *fname, const int verbosity = 1)

    mllama_ctx *new_mllama = new mllama_ctx{};

-    ggml_backend_t backend = ggml_backend_init_best();
-    if (backend == nullptr) {
-        LOG("%s: failed to initialize backend\n", __func__);
-        mllama_free(new_mllama);
-        gguf_free(ctx);
-        return nullptr;
+#ifdef GGML_USE_CUDA
+    new_mllama->backend = ggml_backend_cuda_init(0);
+    LOG("vision using CUDA backend");
+#endif
+
+#ifdef GGML_USE_METAL
+    new_mllama->backend = ggml_backend_metal_init();
+    LOG("vision using Metal backend");
+#endif
+
+#ifdef GGML_USE_CANN
+    new_mllama->backend = ggml_backend_cann_init(0);
+    LOG("vision using CANN backend");
+#endif
+
+#ifdef GGML_USE_VULKAN
+    new_mllama->backend = ggml_backend_vk_init(0);
+    LOG("vision using Vulkan backend");
+#endif
+
+    if (!new_mllama->backend) {
+        new_mllama->backend = ggml_backend_cpu_init();
+        LOG("vision using CPU backend");
    }
-    LOG("%s: using %s backend\n", __func__, ggml_backend_name(backend));
-    new_mllama->backend = backend;

    // load tensors
    {
--- a/llama/patches/0013-use-dynamic-backend-loading-for-clip.patch
+++ b/llama/patches/0013-use-dynamic-backend-loading-for-clip.patch
@@ -1,14 +1,14 @@
 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
 From: jmorganca <jmorganca@gmail.com>
 Date: Sat, 4 Jan 2025 22:52:48 -0800
-Subject: [PATCH] use dynamic backend loading for clip
+Subject: [PATCH] re-enable gpu for clip

 ---
- examples/llava/clip.cpp | 74 +++++++++++++++--------------------------
- 1 file changed, 27 insertions(+), 47 deletions(-)
+ examples/llava/clip.cpp | 86 ++++++++++++++++++++---------------------
+ 1 file changed, 43 insertions(+), 43 deletions(-)

 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index b3c1829f..86b91d5c 100644
+index b3c1829f..718052e1 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
@@ -8,25 +8,25 @@
@@ -56,7 +56,7 @@ index b3c1829f..86b91d5c 100644
 
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
-@@ -1235,35 +1235,15 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
+@@ -1235,30 +1235,30 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
@@ -84,19 +84,30 @@ index b3c1829f..86b91d5c 100644
 -//    new_clip->backend = ggml_backend_sycl_init(0);
 -//    LOG_INF("%s: CLIP using SYCL backend\n", __func__);
 -//#endif
-
-    if (!new_clip->backend) {
-        new_clip->backend = ggml_backend_cpu_init();
-        LOG_INF("%s: CLIP using CPU backend\n", __func__);
-+    ggml_backend_t backend = ggml_backend_init_best();
-+    if (backend == nullptr) {
-+        LOG_ERR("%s: failed to initialize backend\n", __func__);
-+        clip_free(new_clip);
-+        gguf_free(ctx);
-+        return nullptr;
-     }
-+    LOG_INF("%s: using %s backend\n", __func__, ggml_backend_name(backend));
-+    new_clip->backend = backend;
+#ifdef GGML_USE_CUDA
+   new_clip->backend = ggml_backend_cuda_init(0);
+   LOG_INF("%s: CLIP using CUDA backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_METAL
+   new_clip->backend = ggml_backend_metal_init();
+   LOG_INF("%s: CLIP using Metal backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_CANN
+   new_clip->backend = ggml_backend_cann_init(0);
+   LOG_INF("%s: CLIP using CANN backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_VULKAN
+   new_clip->backend = ggml_backend_vk_init(0);
+   LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_SYCL
+   new_clip->backend = ggml_backend_sycl_init(0);
+   LOG_INF("%s: CLIP using SYCL backend\n", __func__);
+#endif
 
-     // model size and capabilities
-     {
+     if (!new_clip->backend) {
+         new_clip->backend = ggml_backend_cpu_init();
--- a/llama/patches/0014-sort-devices-by-score.patch
+++ b/llama/patches/0014-sort-devices-by-score.patch
@@ -8,7 +8,7 @@ Subject: [PATCH] sort devices by score
 1 file changed, 13 insertions(+), 8 deletions(-)

 diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
-index 899d16f2..135f7df0 100644
+index 899d16f2..ac5cda07 100644
 --- a/ggml/src/ggml-backend-reg.cpp
 +++ b/ggml/src/ggml-backend-reg.cpp
@@ -150,7 +150,7 @@ struct ggml_backend_reg_entry {
@@ -29,7 +29,7 @@ index 899d16f2..135f7df0 100644
         if (!reg) {
             return;
         }
-@@ -206,15 +206,20 @@ struct ggml_backend_registry {
+@@ -206,15 +206,15 @@ struct ggml_backend_registry {
 #endif
         backends.push_back({ reg, std::move(handle) });
         for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
@@ -45,15 +45,10 @@ index 899d16f2..135f7df0 100644
 #endif
 -        devices.push_back(device);
 +        devices.push_back({device, score});
-+        std::stable_sort(devices.begin(), devices.end(),
-+            [](const auto & a, const auto & b) {
-+                return a.second > b.second;
-+            }
-+        );
     }
 
     ggml_backend_reg_t load_backend(const std::wstring & path, bool silent) {
-@@ -257,7 +262,7 @@ struct ggml_backend_registry {
+@@ -257,7 +257,7 @@ struct ggml_backend_registry {
 
         GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), utf16_to_utf8(path).c_str());
 
@@ -62,7 +57,7 @@ index 899d16f2..135f7df0 100644
 
         return reg;
     }
-@@ -280,7 +285,7 @@ struct ggml_backend_registry {
+@@ -280,7 +280,7 @@ struct ggml_backend_registry {
         // remove devices
         devices.erase(
             std::remove_if(devices.begin(), devices.end(),
@@ -71,12 +66,17 @@ index 899d16f2..135f7df0 100644
             devices.end());
 
         // remove backend
-@@ -338,7 +343,7 @@ size_t ggml_backend_dev_count() {
+@@ -338,7 +338,12 @@ size_t ggml_backend_dev_count() {
 
 ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
     GGML_ASSERT(index < ggml_backend_dev_count());
 -    return get_reg().devices[index];
-+    return get_reg().devices[index].first;
+    auto devices = get_reg().devices;
+    if (!std::is_heap(devices.begin(), devices.end())) {
+        std::make_heap(devices.begin(), devices.end(), [](const auto & a, const auto & b) { return a.second < b.second; });
+    }
+
+    return devices[index].first;
 }
 
 ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
--- a/llama/patches/0015-add-phony-target-ggml-cpu-for-all-cpu-variants.patch
+++ b/llama/patches/0015-add-phony-target-ggml-cpu-for-all-cpu-variants.patch
@@ -1,29 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Tue, 14 Jan 2025 15:59:04 -0800
-Subject: [PATCH] add phony target ggml-cpu for all cpu variants
-
---
- ggml/src/CMakeLists.txt | 2 ++
- 1 file changed, 2 insertions(+)
-
-diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index 84101c32..72b488dd 100644
--- a/ggml/src/CMakeLists.txt
-+++ b/ggml/src/CMakeLists.txt
-@@ -278,6 +278,7 @@ function(ggml_add_cpu_backend_variant tag_name)
-     endforeach()
- 
-     ggml_add_cpu_backend_variant_impl(${tag_name})
-+    add_dependencies(ggml-cpu ggml-cpu-${tag_name})
- endfunction()
- 
- ggml_add_backend(CPU)
-@@ -286,6 +287,7 @@ if (GGML_CPU_ALL_VARIANTS)
-     if (NOT GGML_BACKEND_DL)
-         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
-     endif()
-+    add_custom_target(ggml-cpu)
-     ggml_add_cpu_backend_variant(sandybridge    AVX)
-     ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 FMA)
-     ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 FMA AVX512)
--- a/llama/patches/0016-remove-sgemm-global-variables.patch
+++ b/llama/patches/0016-remove-sgemm-global-variables.patch
@@ -1,55 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: jmorganca <jmorganca@gmail.com>
-Date: Sun, 9 Feb 2025 17:22:15 -0800
-Subject: [PATCH] remove sgemm global variables
-
-removes the 'iq4nlt' global variable in sgemm.cpp that causes
-a runtime crash when calling dlopen on ggml-cpu libraries as
-its initialization depends on AVX instructions the host machine
-may not have
---
- ggml/src/ggml-cpu/llamafile/sgemm.cpp | 17 +++++++++--------
- 1 file changed, 9 insertions(+), 8 deletions(-)
-
-diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
-index 8fce576c..3f260ce5 100644
--- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp
-+++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
-@@ -279,14 +279,6 @@ template <> inline __m256bh load(const float *p) {
- }
- #endif
- 
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// CONSTANTS
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
-#endif
-
- ////////////////////////////////////////////////////////////////////////////////////////////////////
- // FLOATING POINT MATRIX MULTIPLICATION
- 
-@@ -613,6 +605,14 @@ class tinyBLAS_Q0_AVX {
-                     TC *C, int64_t ldc,
-                     int ith, int nth)
-         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
-+        const int8_t kvalues_iq4nl[16] = {
-+            -127, -104, -83, -65,
-+            -49,  -35,  -22, -10,
-+              1,   13,   25,  38,
-+             53,   69,   89, 113
-+        };
-+
-+        iq4nlt = _mm_loadu_si128((const __m128i *)kvalues_iq4nl);
-     }
- 
-     void matmul(int64_t m, int64_t n) {
-@@ -1037,6 +1037,7 @@ class tinyBLAS_Q0_AVX {
-     const int64_t ldc;
-     const int ith;
-     const int nth;
-+    __m128i iq4nlt;
- };
- #endif // __AVX__
- 
--- a/llama/patches/0017-try-catch-backend-load.patch
+++ b/llama/patches/0017-try-catch-backend-load.patch
@@ -1,69 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Tue, 11 Feb 2025 14:06:36 -0800
-Subject: [PATCH] try/catch backend load
-
---
- ggml/src/ggml-backend-reg.cpp | 45 ++++++++++++++++++-----------------
- 1 file changed, 23 insertions(+), 22 deletions(-)
-
-diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
-index 135f7df0..84b21dd8 100644
--- a/ggml/src/ggml-backend-reg.cpp
-+++ b/ggml/src/ggml-backend-reg.cpp
-@@ -512,32 +512,33 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
-         }
-         fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
-         for (const auto & entry : dir_it) {
-            if (entry.is_regular_file()) {
-                std::wstring filename = entry.path().filename().wstring();
-                std::wstring ext = entry.path().extension().wstring();
-                if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
-                    dl_handle_ptr handle { dl_load_library(entry.path().wstring()) };
-                    if (!handle && !silent) {
-                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
-                    }
-                    if (handle) {
-+            try {
-+                if (entry.is_regular_file()) {
-+                    std::wstring filename = entry.path().filename().wstring();
-+                    std::wstring ext = entry.path().extension().wstring();
-+                    if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
-+                        dl_handle_ptr handle { dl_load_library(entry.path().wstring()) };
-+                        if (!handle) {
-+                            GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
-+                            continue;
-+                        }
-+
-                         auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
-                        if (score_fn) {
-                            int s = score_fn();
-#ifndef NDEBUG
-                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s);
-#endif
-                            if (s > best_score) {
-                                best_score = s;
-                                best_path = entry.path().wstring();
-                            }
-                        } else {
-                            if (!silent) {
-                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
-                            }
-+                        if (!score_fn) {
-+                            GGML_LOG_DEBUG("%s: failed to find ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
-+                            continue;
-+                        }
-+
-+                        int s = score_fn();
-+                        GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s);
-+                        if (s > best_score) {
-+                            best_score = s;
-+                            best_path = entry.path().wstring();
-                         }
-                     }
-                 }
-+            } catch (const std::exception & e) {
-+                GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), e.what());
-             }
-         }
-     }
--- a/llama/runner/README.md
+++ b/llama/runner/README.md
@@ -4,18 +4,18 @@

 A minimial runner for loading a model and running inference via a http web server.

-```shell
+```
 ./runner -model <model binary>
 ```

 ### Completion

-```shell
+```
 curl -X POST -H "Content-Type: application/json" -d '{"prompt": "hi"}' http://localhost:8080/completion
 ```

 ### Embeddings

-```shell
+```
 curl -X POST -H "Content-Type: application/json" -d '{"prompt": "turn me into an embedding"}' http://localhost:8080/embedding
 ```
--- a/runner/llamarunner/cache.go
+++ b/runner/llamarunner/cache.go
@@ -1,4 +1,4 @@
-package llamarunner
+package runner

 import (
 	"errors"
--- a/runner/llamarunner/cache_test.go
+++ b/runner/llamarunner/cache_test.go
@@ -1,4 +1,4 @@
-package llamarunner
+package runner

 import (
 	"testing"
--- a/runner/llamarunner/image.go
+++ b/runner/llamarunner/image.go
@@ -1,4 +1,4 @@
-package llamarunner
+package runner

 import (
 	"errors"
--- a/runner/llamarunner/image_test.go
+++ b/runner/llamarunner/image_test.go
@@ -1,4 +1,4 @@
-package llamarunner
+package runner

 import (
 	"reflect"
--- a/runner/llamarunner/runner.go
+++ b/runner/llamarunner/runner.go
@@ -1,4 +1,4 @@
-package llamarunner
+package runner

 import (
 	"context"
@@ -24,7 +24,6 @@ import (

 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/llama"
-	"github.com/ollama/ollama/runner/common"
 )

 // input is an element of the prompt to process, either
@@ -444,6 +443,7 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 		s.lc.Synchronize()
 	}

+	var totalSamplingTime time.Duration
 	for i, seq := range s.seqs {
 		if seq == nil {
 			continue
@@ -478,8 +478,12 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 		}

 		// sample a token
+		samplingStart := time.Now()
 		token := seq.samplingCtx.Sample(s.lc, seq.iBatch)
 		seq.samplingCtx.Accept(token, true)
+		samplingTime := time.Since(samplingStart)
+		totalSamplingTime += samplingTime
+		slog.Info("sampling time", "time", samplingTime)
 		piece := s.model.TokenToPiece(token)

 		seq.numPredicted++
@@ -499,12 +503,12 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 		seq.pendingResponses = append(seq.pendingResponses, piece)
 		sequence := strings.Join(seq.pendingResponses, "")

-		if ok, stop := common.FindStop(sequence, seq.stop); ok {
+		if ok, stop := findStop(sequence, seq.stop); ok {
 			slog.Debug("hit stop token", "pending", seq.pendingResponses, "stop", stop)

 			var tokenTruncated bool
 			origLen := len(seq.pendingResponses)
-			seq.pendingResponses, tokenTruncated = common.TruncateStop(seq.pendingResponses, stop)
+			seq.pendingResponses, tokenTruncated = truncateStop(seq.pendingResponses, stop)
 			newLen := len(seq.pendingResponses)

 			// Update the cache based on the tokens that will be returned:
@@ -525,11 +529,11 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 			continue
 		}

-		if common.ContainsStopSuffix(sequence, seq.stop) {
+		if containsStopSuffix(sequence, seq.stop) {
 			continue
 		}

-		if common.IncompleteUnicode(sequence) {
+		if incompleteUnicode(sequence) {
 			continue
 		}

@@ -636,6 +640,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	samplingParams.Seed = uint32(req.Seed)
 	samplingParams.Grammar = req.Grammar

+	start := time.Now()
 	seq, err := s.NewSequence(req.Prompt, req.Images, NewSequenceParams{
 		numPredict:     req.NumPredict,
 		stop:           req.Stop,
@@ -643,6 +648,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		samplingParams: &samplingParams,
 		embedding:      false,
 	})
+	slog.Info("new sequence created", "duration", time.Since(start))
 	if err != nil {
 		http.Error(w, fmt.Sprintf("Failed to create new sequence: %v", err), http.StatusInternalServerError)
 		return
@@ -845,6 +851,8 @@ func (s *Server) loadModel(
 	threads int,
 	multiUserCache bool,
 ) {
+	llama.BackendInit()
+
 	var err error
 	s.model, err = llama.LoadModelFromFile(mpath, params)
 	if err != nil {
@@ -884,6 +892,9 @@ func (s *Server) loadModel(
 }

 func Execute(args []string) error {
+	if args[0] == "runner" {
+		args = args[1:]
+	}
 	fs := flag.NewFlagSet("runner", flag.ExitOnError)
 	mpath := fs.String("model", "", "Path to model binary file")
 	ppath := fs.String("mmproj", "", "Path to projector binary file")
@@ -930,8 +941,6 @@ func Execute(args []string) error {
 	})
 	slog.SetDefault(slog.New(handler))
 	slog.Info("starting go runner")
-
-	llama.BackendInit()
 	slog.Info("system", "info", llama.PrintSystemInfo(), "threads", *threads)

 	server := &Server{
--- a/runner/common/stop.go
+++ b/runner/common/stop.go
@@ -1,10 +1,10 @@
-package common
+package runner

 import (
 	"strings"
 )

-func FindStop(sequence string, stops []string) (bool, string) {
+func findStop(sequence string, stops []string) (bool, string) {
 	for _, stop := range stops {
 		if strings.Contains(sequence, stop) {
 			return true, stop
@@ -14,7 +14,7 @@ func FindStop(sequence string, stops []string) (bool, string) {
 	return false, ""
 }

-func ContainsStopSuffix(sequence string, stops []string) bool {
+func containsStopSuffix(sequence string, stops []string) bool {
 	for _, stop := range stops {
 		for i := 1; i <= len(stop); i++ {
 			if strings.HasSuffix(sequence, stop[:i]) {
@@ -29,7 +29,7 @@ func ContainsStopSuffix(sequence string, stops []string) bool {
 // truncateStop removes the provided stop string from pieces,
 // returning the partial pieces with stop removed, including truncating
 // the last piece if required (and signalling if this was the case)
-func TruncateStop(pieces []string, stop string) ([]string, bool) {
+func truncateStop(pieces []string, stop string) ([]string, bool) {
 	joined := strings.Join(pieces, "")

 	index := strings.Index(joined, stop)
@@ -65,7 +65,7 @@ func TruncateStop(pieces []string, stop string) ([]string, bool) {
 	return result, tokenTruncated
 }

-func IncompleteUnicode(token string) bool {
+func incompleteUnicode(token string) bool {
 	incomplete := false

 	// check if there is incomplete UTF-8 character at the end
--- a/runner/common/stop_test.go
+++ b/runner/common/stop_test.go
@@ -1,4 +1,4 @@
-package common
+package runner

 import (
 	"reflect"
@@ -52,7 +52,7 @@ func TestTruncateStop(t *testing.T) {

 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			result, resultTrunc := TruncateStop(tt.pieces, tt.stop)
+			result, resultTrunc := truncateStop(tt.pieces, tt.stop)
 			if !reflect.DeepEqual(result, tt.expected) || resultTrunc != tt.expectedTrunc {
 				t.Errorf("truncateStop(%v, %s): have %v (%v); want %v (%v)", tt.pieces, tt.stop, result, resultTrunc, tt.expected, tt.expectedTrunc)
 			}
@@ -120,7 +120,7 @@ func TestIncompleteUnicode(t *testing.T) {

 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			result := IncompleteUnicode(tt.input)
+			result := incompleteUnicode(tt.input)
 			if result != tt.expected {
 				t.Errorf("incompleteUnicode(%s): have %v; want %v", tt.input, result, tt.expected)
 			}
--- a/llm/memory.go
+++ b/llm/memory.go
@@ -116,7 +116,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		opts.NumCtx = max(opts.NumCtx, 2048)
 	}

-	layers := f.Tensors().GroupLayers()
+	layers := f.Tensors().Layers()
 	// add one layer worth of memory as a buffer
 	if blk0, ok := layers["blk.0"]; ok {
 		layerSize = blk0.Size()
@@ -410,7 +410,7 @@ func projectorMemoryRequirements(filename string) (weights, graphSize uint64) {
 		return 0, 0
 	}

-	for _, layer := range ggml.Tensors().GroupLayers() {
+	for _, layer := range ggml.Tensors().Layers() {
 		weights += layer.Size()
 	}

@@ -431,7 +431,7 @@ func projectorMemoryRequirements(filename string) (weights, graphSize uint64) {
 		headCount := kv("attention.head_count")

 		numPatches := (imageSize / kv("patch_size")) * (imageSize / kv("patch_size"))
-		if _, ok := ggml.Tensors().GroupLayers()["v"]["class_embd"]; ok {
+		if _, ok := ggml.Tensors().Layers()["v"]["class_embd"]; ok {
 			numPatches++
 		}

--- a/llm/server.go
+++ b/llm/server.go
@@ -29,6 +29,7 @@ import (
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/grammar"
 	"github.com/ollama/ollama/llama"
 )

@@ -101,8 +102,12 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, f *ggml.GGML, adapt
 		gpus = discover.GetCPUInfo()
 	}

-	estimate := EstimateGPULayers(gpus, f, projectors, opts)
-	if len(gpus) > 1 || gpus[0].Library != "cpu" {
+	var estimate MemoryEstimate
+	if len(gpus) == 1 && gpus[0].Library == "cpu" {
+		estimate = EstimateGPULayers(gpus, f, projectors, opts)
+	} else {
+		estimate = EstimateGPULayers(gpus, f, projectors, opts)
+
 		switch {
 		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:
 			// disable partial offloading when model is greater than total system memory as this
@@ -229,207 +234,149 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, f *ggml.GGML, adapt
 		params = append(params, "--multiuser-cache")
 	}

-	libs := make(map[string]string)
-	if entries, err := os.ReadDir(discover.LibOllamaPath); err == nil {
-		for _, entry := range entries {
-			libs[entry.Name()] = filepath.Join(discover.LibOllamaPath, entry.Name())
-		}
+	exe, err := os.Executable()
+	if err != nil {
+		return nil, err
 	}

-	lib := gpus[0].RunnerName()
-	requested := envconfig.LLMLibrary()
-	if libs[requested] != "" {
-		slog.Info("using requested gpu library", "requested", requested)
-		lib = requested
-	}
-
-	var compatible []string
-	for k := range libs {
-		// exact match first
-		if k == lib {
-			compatible = append([]string{k}, compatible...)
-			continue
-		}
-
-		// then match the family (e.g. 'cuda')
-		if strings.Split(k, "_")[0] == strings.Split(lib, "_")[0] {
-			compatible = append(compatible, k)
+	// Find an availableServers  port, retry on each iteration in case the failure was a port conflict race
+	port := 0
+	if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
+		var l *net.TCPListener
+		if l, err = net.ListenTCP("tcp", a); err == nil {
+			port = l.Addr().(*net.TCPAddr).Port
+			l.Close()
 		}
 	}
-	slog.Debug("compatible gpu libraries", "compatible", compatible)
+	if port == 0 {
+		slog.Debug("ResolveTCPAddr failed ", "error", err)
+		port = rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
+	}
+	finalParams := []string{"runner"}
+	finalParams = append(finalParams, params...)
+	finalParams = append(finalParams, "--port", strconv.Itoa(port))

-	// iterate through compatible GPU libraries such as 'cuda_v12', 'cuda_v11', 'rocm', etc.
-	// adding each library's respective path to the LD_LIBRARY_PATH, until finally running
-	// without any LD_LIBRARY_PATH flags
-	for {
-		port := 0
-		if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
-			var l *net.TCPListener
-			if l, err = net.ListenTCP("tcp", a); err == nil {
-				port = l.Addr().(*net.TCPAddr).Port
-				l.Close()
-			}
-		}
-		if port == 0 {
-			slog.Debug("ResolveTCPAddr failed, using random port")
-			port = rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
-		}
-		finalParams := []string{"runner"}
-		if envconfig.NewEngine() {
-			finalParams = append(finalParams, "--ollama-engine")
-		}
-		finalParams = append(finalParams, params...)
-		finalParams = append(finalParams, "--port", strconv.Itoa(port))
+	pathEnv := "LD_LIBRARY_PATH"
+	if runtime.GOOS == "windows" {
+		pathEnv = "PATH"
+	}
+	// Start with the server directory for the LD_LIBRARY_PATH/PATH
+	libraryPaths := []string{filepath.Dir(exe)}

-		var pathEnv string
-		switch runtime.GOOS {
-		case "windows":
-			pathEnv = "PATH"
-		case "darwin":
-			pathEnv = "DYLD_LIBRARY_PATH"
-		default:
-			pathEnv = "LD_LIBRARY_PATH"
-		}
+	if libraryPath, ok := os.LookupEnv(pathEnv); ok {
+		// favor our bundled library dependencies over system libraries
+		libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
+	}

-		var libraryPaths []string
-		if libraryPath, ok := os.LookupEnv(pathEnv); ok {
-			libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
-		}
+	// Note: we always put the dependency path first
+	// since this was the exact version we compiled/linked against
+	if gpus[0].DependencyPath != nil {
+		// assume gpus from the same library have the same dependency path
+		libraryPaths = append(gpus[0].DependencyPath, libraryPaths...)
+	}

-		if len(compatible) > 0 {
-			c := compatible[0]
-			if libpath, ok := libs[c]; ok {
-				slog.Debug("adding gpu library", "path", libpath)
-				libraryPaths = append(libraryPaths, libpath)
-			}
-		}
+	// TODO - once fully switched to the Go runner, load the model here for tokenize/detokenize cgo access
+	s := &llmServer{
+		port:        port,
+		cmd:         exec.Command(exe, finalParams...),
+		status:      NewStatusWriter(os.Stderr),
+		options:     opts,
+		modelPath:   model,
+		estimate:    estimate,
+		numParallel: numParallel,
+		sem:         semaphore.NewWeighted(int64(numParallel)),
+		totalLayers: f.KV().BlockCount() + 1,
+		gpus:        gpus,
+		done:        make(chan error, 1),
+	}

-		// Note: we always put the dependency path first
-		// since this was the exact version we compiled/linked against
-		if gpus[0].DependencyPath != nil {
-			slog.Debug("adding gpu dependency paths", "paths", gpus[0].DependencyPath)
-			// assume gpus from the same library have the same dependency path
-			libraryPaths = append(gpus[0].DependencyPath, libraryPaths...)
-		}
+	s.cmd.Env = os.Environ()
+	s.cmd.Stdout = os.Stdout
+	s.cmd.Stderr = s.status
+	s.cmd.SysProcAttr = LlamaServerSysProcAttr

-		// finally, add the root library path
-		libraryPaths = append(libraryPaths, discover.LibOllamaPath)
+	envWorkarounds := [][2]string{}
+	for _, gpu := range gpus {
+		envWorkarounds = append(envWorkarounds, gpu.EnvWorkarounds...)
+	}
+	visibleDevicesEnv, visibleDevicesEnvVal := gpus.GetVisibleDevicesEnv()
+	pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))

-		exe, err := os.Executable()
-		if err != nil {
-			return nil, fmt.Errorf("unable to lookup executable path: %w", err)
-		}
-
-		if eval, err := filepath.EvalSymlinks(exe); err == nil {
-			exe = eval
-		}
-
-		// TODO - once fully switched to the Go runner, load the model here for tokenize/detokenize cgo access
-		s := &llmServer{
-			port:        port,
-			cmd:         exec.Command(exe, finalParams...),
-			status:      NewStatusWriter(os.Stderr),
-			options:     opts,
-			modelPath:   model,
-			estimate:    estimate,
-			numParallel: numParallel,
-			sem:         semaphore.NewWeighted(int64(numParallel)),
-			totalLayers: f.KV().BlockCount() + 1,
-			gpus:        gpus,
-			done:        make(chan error, 1),
-		}
-
-		s.cmd.Env = os.Environ()
-		s.cmd.Stdout = os.Stdout
-		s.cmd.Stderr = s.status
-		s.cmd.SysProcAttr = LlamaServerSysProcAttr
-
-		envWorkarounds := [][2]string{}
-		for _, gpu := range gpus {
-			envWorkarounds = append(envWorkarounds, gpu.EnvWorkarounds...)
-		}
-		visibleDevicesEnv, visibleDevicesEnvVal := gpus.GetVisibleDevicesEnv()
-		pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))
-
-		// Update or add the path and visible devices variable with our adjusted version
-		pathNeeded := true
-		devicesNeeded := visibleDevicesEnv != ""
-		for i := range s.cmd.Env {
-			cmp := strings.SplitN(s.cmd.Env[i], "=", 2)
-			if strings.EqualFold(cmp[0], pathEnv) {
-				s.cmd.Env[i] = pathEnv + "=" + pathEnvVal
-				pathNeeded = false
-			} else if devicesNeeded && strings.EqualFold(cmp[0], visibleDevicesEnv) {
-				s.cmd.Env[i] = visibleDevicesEnv + "=" + visibleDevicesEnvVal
-				devicesNeeded = false
-			} else if len(envWorkarounds) != 0 {
-				for _, kv := range envWorkarounds {
-					if strings.EqualFold(cmp[0], kv[0]) {
-						s.cmd.Env[i] = kv[0] + "=" + kv[1]
-					}
+	// Update or add the path and visible devices variable with our adjusted version
+	pathNeeded := true
+	devicesNeeded := visibleDevicesEnv != ""
+	for i := range s.cmd.Env {
+		cmp := strings.SplitN(s.cmd.Env[i], "=", 2)
+		if strings.EqualFold(cmp[0], pathEnv) {
+			s.cmd.Env[i] = pathEnv + "=" + pathEnvVal
+			pathNeeded = false
+		} else if devicesNeeded && strings.EqualFold(cmp[0], visibleDevicesEnv) {
+			s.cmd.Env[i] = visibleDevicesEnv + "=" + visibleDevicesEnvVal
+			devicesNeeded = false
+		} else if len(envWorkarounds) != 0 {
+			for _, kv := range envWorkarounds {
+				if strings.EqualFold(cmp[0], kv[0]) {
+					s.cmd.Env[i] = kv[0] + "=" + kv[1]
 				}
 			}
 		}
-		if pathNeeded {
-			s.cmd.Env = append(s.cmd.Env, pathEnv+"="+pathEnvVal)
-		}
-		if devicesNeeded {
-			s.cmd.Env = append(s.cmd.Env, visibleDevicesEnv+"="+visibleDevicesEnvVal)
-		}
-
-		slog.Info("starting llama server", "cmd", s.cmd.String())
-		if envconfig.Debug() {
-			filteredEnv := []string{}
-			for _, ev := range s.cmd.Env {
-				if strings.HasPrefix(ev, "CUDA_") ||
-					strings.HasPrefix(ev, "ROCR_") ||
-					strings.HasPrefix(ev, "ROCM_") ||
-					strings.HasPrefix(ev, "HIP_") ||
-					strings.HasPrefix(ev, "GPU_") ||
-					strings.HasPrefix(ev, "HSA_") ||
-					strings.HasPrefix(ev, "GGML_") ||
-					strings.HasPrefix(ev, "PATH=") ||
-					strings.HasPrefix(ev, "LD_LIBRARY_PATH=") ||
-					strings.HasPrefix(ev, "DYLD_LIBRARY_PATH=") {
-					filteredEnv = append(filteredEnv, ev)
-				}
-			}
-			// Log at debug as the environment is inherited and might contain sensitive information
-			slog.Debug("subprocess", "environment", filteredEnv)
-		}
-
-		if err = s.cmd.Start(); err != nil {
-			var msg string
-			if s.status != nil && s.status.LastErrMsg != "" {
-				msg = s.status.LastErrMsg
-			}
-			err := fmt.Errorf("error starting runner: %v %s", err, msg)
-			if len(compatible) == 0 {
-				return nil, err
-			}
-
-			slog.Warn("unable to start runner with compatible gpu", "error", err, "compatible", compatible)
-			compatible = compatible[1:]
-			continue
-		}
-
-		// reap subprocess when it exits
-		go func() {
-			err := s.cmd.Wait()
-			// Favor a more detailed message over the process exit status
-			if err != nil && s.status != nil && s.status.LastErrMsg != "" {
-				slog.Error("llama runner terminated", "error", err)
-				if strings.Contains(s.status.LastErrMsg, "unknown model") {
-					s.status.LastErrMsg = "this model is not supported by your version of Ollama. You may need to upgrade"
-				}
-				s.done <- errors.New(s.status.LastErrMsg)
-			} else {
-				s.done <- err
-			}
-		}()
-
-		return s, nil
 	}
+	if pathNeeded {
+		s.cmd.Env = append(s.cmd.Env, pathEnv+"="+pathEnvVal)
+	}
+	if devicesNeeded {
+		s.cmd.Env = append(s.cmd.Env, visibleDevicesEnv+"="+visibleDevicesEnvVal)
+	}
+
+	slog.Info("starting llama server", "cmd", s.cmd.String())
+	if envconfig.Debug() {
+		filteredEnv := []string{}
+		for _, ev := range s.cmd.Env {
+			if strings.HasPrefix(ev, "CUDA_") ||
+				strings.HasPrefix(ev, "ROCR_") ||
+				strings.HasPrefix(ev, "ROCM_") ||
+				strings.HasPrefix(ev, "HIP_") ||
+				strings.HasPrefix(ev, "GPU_") ||
+				strings.HasPrefix(ev, "HSA_") ||
+				strings.HasPrefix(ev, "GGML_") ||
+				strings.HasPrefix(ev, "PATH=") ||
+				strings.HasPrefix(ev, "LD_LIBRARY_PATH=") {
+				filteredEnv = append(filteredEnv, ev)
+			}
+		}
+		// Log at debug as the environment is inherited and might contain sensitive information
+		slog.Debug("subprocess", "environment", filteredEnv)
+	}
+
+	if err = s.cmd.Start(); err != nil {
+		// Detect permission denied and augment the message about noexec
+		if errors.Is(err, os.ErrPermission) {
+			return nil, fmt.Errorf("unable to start server %w.  %s may have noexec set.  Set OLLAMA_TMPDIR for server to a writable executable directory", err, exe)
+		}
+
+		msg := ""
+		if s.status != nil && s.status.LastErrMsg != "" {
+			msg = s.status.LastErrMsg
+		}
+		return nil, fmt.Errorf("error starting the external llama server: %v %s", err, msg)
+	}
+
+	// reap subprocess when it exits
+	go func() {
+		err := s.cmd.Wait()
+		// Favor a more detailed message over the process exit status
+		if err != nil && s.status != nil && s.status.LastErrMsg != "" {
+			slog.Debug("llama runner terminated", "error", err)
+			if strings.Contains(s.status.LastErrMsg, "unknown model") {
+				s.status.LastErrMsg = "this model is not supported by your version of Ollama. You may need to upgrade"
+			}
+			s.done <- errors.New(s.status.LastErrMsg)
+		} else {
+			s.done <- err
+		}
+	}()
+
+	return s, nil
 }

 type ServerStatus int
@@ -714,9 +661,9 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 			}

 			// User provided a JSON schema
-			g := llama.SchemaToGrammar(req.Format)
-			if g == nil {
-				return fmt.Errorf("invalid JSON schema in format")
+			g, err := grammar.FromSchema(nil, req.Format)
+			if err != nil {
+				return fmt.Errorf("invalid JSON schema in format: %w", err)
 			}
 			request["grammar"] = string(g)
 		}
@@ -736,7 +683,6 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	if req.Options.NumPredict < 0 || req.Options.NumPredict > 10*s.options.NumCtx {
 		req.Options.NumPredict = 10 * s.options.NumCtx
 	}
-
 	// Make sure the server is ready
 	status, err := s.getServerStatusRetry(ctx)
 	if err != nil {
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
ParthSareen	a4265c278a	wip	2025-03-24 11:02:02 -07:00
ParthSareen	ffd6428758	err handling + fixing scope issue	2025-02-11 19:29:48 -08:00
ParthSareen	aa6d5151df	wip with json stuff and cleanup	2025-02-11 16:40:40 -08:00
ParthSareen	25edfa6fdb	first pass so working	2025-02-04 22:09:22 -08:00
ParthSareen	cc2e44b885	Improved sampler interface	2025-02-03 16:05:15 -08:00
ParthSareen	d5f8670f0a	WIP simple SO working	2025-02-03 11:40:06 -08:00
ParthSareen	524029cd6d	tested with so	2025-01-31 17:12:39 -08:00
ParthSareen	b973dedb4b	saved state	2025-01-30 15:17:42 -08:00
ParthSareen	c56a8b7749	wip	2025-01-30 15:05:25 -08:00
ParthSareen	198fde82aa	Enable array type json	2025-01-29 14:54:39 -08:00
ParthSareen	77f709ebd5	cleanup state machine	2025-01-29 14:28:02 -08:00
ParthSareen	73098a2973	WIP working with other stuff	2025-01-29 10:28:22 -08:00
ParthSareen	e93db4d20e	WIP	2025-01-27 16:33:55 -08:00
ParthSareen	a2a73ce5e0	wip!	2025-01-23 20:21:50 -08:00
ParthSareen	6ba557f25b	checkpoint	2025-01-23 09:46:14 -08:00
ParthSareen	a7c8cc06da	json checkpoint	2025-01-21 17:48:12 -08:00
Blake Mizerany	089bbb537d	grammar: introduce new grammar package This package provides a way to convert JSON schemas to equivalent EBNF. It is intended to be a replacement to llama.cpp's schema_to_grammar. This is still an early version and does not yet support all JSON schema features. The to-do list includes: - minumum/maximum constraints on integer types - minLength/maxLength constraints on string types - defs and refs	2025-01-20 14:40:50 -08:00
ParthSareen	7cd9fbbbb1	improve temperature sampler	2025-01-20 13:53:07 -08:00
ParthSareen	b91487f289	Working json sampler	2025-01-15 14:08:54 -08:00
ParthSareen	5b19d4941a	addressing comments + cleanup	2025-01-14 16:23:48 -08:00
ParthSareen	5e73f24e16	sampling package	2025-01-14 16:23:38 -08:00
Michael Yang	4aac178cac	update ci	2025-01-14 14:25:56 -08:00
Michael Yang	800e21d060	sort devices by type	2025-01-14 14:25:56 -08:00
Michael Yang	5827999e9e	next ollama runner implement llama and mllama model architectures in go using ggml (through cgo)	2025-01-14 14:25:56 -08:00