model: framework for testing forward pass

2026-01-21 05:48:35 -05:00 · 2025-05-08 09:25:12 -07:00
465 changed files with 47222 additions and 301684 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -23,7 +23,7 @@ jobs:
          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_OUTPUT

  darwin-build:
-    runs-on: macos-13-xlarge
+    runs-on: macos-13
    environment: release
    needs: setup-environment
    strategy:
@@ -54,6 +54,48 @@ jobs:
          name: build-${{ matrix.os }}-${{ matrix.arch }}
          path: dist/*

+  darwin-sign:
+    runs-on: macos-13
+    environment: release
+    needs: darwin-build
+    steps:
+      - uses: actions/checkout@v4
+      - run: |
+          echo $MACOS_SIGNING_KEY | base64 --decode > certificate.p12
+          security create-keychain -p password build.keychain
+          security default-keychain -s build.keychain
+          security unlock-keychain -p password build.keychain
+          security import certificate.p12 -k build.keychain -P $MACOS_SIGNING_KEY_PASSWORD -T /usr/bin/codesign
+          security set-key-partition-list -S apple-tool:,apple:,codesign: -s -k password build.keychain
+          security set-keychain-settings -lut 3600 build.keychain
+        env:
+          MACOS_SIGNING_KEY: ${{ secrets.MACOS_SIGNING_KEY }}
+          MACOS_SIGNING_KEY_PASSWORD: ${{ secrets.MACOS_SIGNING_KEY_PASSWORD }}
+      - uses: actions/download-artifact@v4
+        with:
+          name: build-darwin-amd64
+          path: dist/darwin-amd64
+      - uses: actions/download-artifact@v4
+        with:
+          name: build-darwin-arm64
+          path: dist/darwin-arm64
+      - run: |
+          export VERSION=${GITHUB_REF_NAME#v}
+          ./scripts/build_darwin.sh sign macapp
+        env:
+          APPLE_IDENTITY: ${{ secrets.APPLE_IDENTITY }}
+          APPLE_PASSWORD: ${{ secrets.APPLE_PASSWORD }}
+          APPLE_TEAM_ID: ${{ vars.APPLE_TEAM_ID }}
+          APPLE_ID: ${{ vars.APPLE_ID }}
+          SDKROOT: /Applications/Xcode_14.1.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
+          DEVELOPER_DIR: /Applications/Xcode_14.1.0.app/Contents/Developer
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-darwin
+          path: |
+            dist/Ollama-darwin.zip
+            dist/ollama-darwin.tgz
+
  windows-depends:
    strategy:
      matrix:
@@ -66,13 +108,11 @@ jobs:
            preset: 'CUDA 12'
            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
            cuda-version: '12.8'
-            flags: ''
          - os: windows
            arch: amd64
            preset: 'ROCm 6'
            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
            rocm-version: '6.2'
-            flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
    environment: release
    env:
@@ -115,9 +155,6 @@ jobs:
          echo "$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 $env:GITHUB_ENV -Append
-          echo "HIPCXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "HIP_PLATFORM=amd" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "CMAKE_PREFIX_PATH=$hipPath" | Out-File -FilePath $env:GITHUB_ENV -Append
      - if: matrix.preset == 'CPU'
        run: |
          echo "CC=clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
@@ -136,9 +173,9 @@ jobs:
          key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}
      - name: Build target "${{ matrix.preset }}"
        run: |
-          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
+          Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
+          cmake --preset "${{ matrix.preset }}"
          cmake --build --parallel --preset "${{ matrix.preset }}"
          cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || 'CPU' }}" --strip --parallel 8
        env:
@@ -188,11 +225,61 @@ jobs:
          go-version-file: go.mod
      - run: |
          go build -o dist/${{ matrix.os }}-${{ matrix.arch }}/ .
+      - if: matrix.arch == 'arm64'
+        run: |
+          Invoke-WebRequest -Uri "https://aka.ms/vs/17/release/vc_redist.arm64.exe" -OutFile "dist\windows-arm64\vc_redist.arm64.exe"
+      - run: |
+          $env:VERSION='${{ github.ref_name }}' -Replace "v(.*)", '$1'
+          & .\scripts\build_windows.ps1 buildApp
+        env:
+          VCToolsRedistDir: stub
      - uses: actions/upload-artifact@v4
        with:
          name: build-${{ matrix.os }}-${{ matrix.arch }}
          path: |
            dist\${{ matrix.os }}-${{ matrix.arch }}\*.exe
+            dist\${{ matrix.os }}-${{ matrix.arch }}-app.exe
+
+  windows-sign:
+    runs-on: windows-2022
+    environment: release
+    needs: [windows-depends, windows-build]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: google-github-actions/auth@v2
+        with:
+          project_id: ollama
+          credentials_json: ${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}
+      - run: |
+          $ErrorActionPreference = "Stop"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${{ runner.temp }}\sdksetup.exe"
+          Start-Process "${{ runner.temp }}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${{ runner.temp }}\plugin.zip"
+          Expand-Archive -Path "${{ runner.temp }}\plugin.zip" -DestinationPath "${{ runner.temp }}\plugin\"
+          & "${{ runner.temp }}\plugin\*\kmscng.msi" /quiet
+
+          echo "${{ vars.OLLAMA_CERT }}" >ollama_inc.crt
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: build-windows-*
+          path: dist\
+          merge-multiple: true
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: depends-windows-amd64-*
+          path: dist\windows-amd64\
+          merge-multiple: true
+      - run: |
+          & .\scripts\build_windows.ps1 gatherDependencies sign buildInstaller distZip
+        env:
+          KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-windows
+          path: |
+            dist\OllamaSetup.exe
+            dist\ollama-windows-*.zip

  linux-build:
    strategy:
@@ -225,26 +312,20 @@ jobs:
            CGO_CFLAGS=${{ env.CGO_CFLAGS }}
            CGO_CXXFLAGS=${{ env.CGO_CXXFLAGS }}
          outputs: type=local,dest=dist/${{ matrix.os }}-${{ matrix.arch }}
-          cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          cache-from: type=registry,ref=ollama/ollama:latest
          cache-to: type=inline
      - run: |
          for COMPONENT in bin/* lib/ollama/*; do
            case "$COMPONENT" in
-              bin/ollama)                echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_sbsa)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
-              lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
-              lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
+              bin/ollama)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/*.so)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v12)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_jetpack5) echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
+              lib/ollama/cuda_jetpack6) echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
+              lib/ollama/rocm)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
            esac
          done
        working-directory: dist/${{ matrix.os }}-${{ matrix.arch }}
-      - run: |
-          echo "Manifests"
-          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in ; do
-            echo $ARCHIVE
-            cat $ARCHIVE
-          done
      - run: |
          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | pigz -9vc >$(basename ${ARCHIVE//.*/}.tgz);
@@ -298,8 +379,8 @@ jobs:
          context: .
          platforms: ${{ matrix.os }}/${{ matrix.arch }}
          build-args: ${{ matrix.build-args }}
-          outputs: type=image,name=${{ vars.DOCKER_REPO }},push-by-digest=true,name-canonical=true,push=true
-          cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          outputs: type=image,name=ollama/ollama,push-by-digest=true,name-canonical=true,push=true
+          cache-from: type=registry,ref=ollama/ollama:latest
          cache-to: type=inline
      - run: |
          mkdir -p ${{ matrix.os }}-${{ matrix.arch }}
@@ -331,7 +412,7 @@ jobs:
            latest=false
            suffix=${{ matrix.suffix }}
          images: |
-            ${{ vars.DOCKER_REPO }}
+            ollama/ollama
          tags: |
            type=ref,enable=true,priority=600,prefix=pr-,event=pr
            type=semver,pattern={{version}}
@@ -341,24 +422,56 @@ jobs:
          path: ${{ runner.temp }}
          merge-multiple: true
      - run: |
-          docker buildx imagetools create $(echo '${{ steps.metadata.outputs.json }}' | jq -cr '.tags | map("-t", .) | join(" ")') $(cat *-${{ matrix.suffix }}.txt | xargs printf '${{ vars.DOCKER_REPO }}@%s ')
-          docker buildx imagetools inspect ${{ vars.DOCKER_REPO }}:${{ steps.metadata.outputs.version }}
+          docker buildx imagetools create $(echo '${{ steps.metadata.outputs.json }}' | jq -cr '.tags | map("-t", .) | join(" ")') $(cat *-${{ matrix.suffix }}.txt | xargs printf 'ollama/ollama@%s ')
+          docker buildx imagetools inspect ollama/ollama:${{ steps.metadata.outputs.version }}
        working-directory: ${{ runner.temp }}

  # Trigger downstream release process
  trigger:
    runs-on: ubuntu-latest
    environment: release
-    needs: [darwin-build, windows-build, windows-depends, linux-build]
+    needs: [darwin-build, windows-build, windows-depends]
+    steps:
+      - name: Trigger downstream release process
+        run: |
+          curl -L \
+            -X POST \
+            -H "Accept: application/vnd.github+json" \
+            -H "Authorization: Bearer ${{ secrets.RELEASE_TOKEN }}" \
+            -H "X-GitHub-Api-Version: 2022-11-28" \
+            https://api.github.com/repos/ollama/${{ vars.RELEASE_REPO }}/dispatches \
+            -d "{\"event_type\": \"trigger-workflow\", \"client_payload\": {\"run_id\": \"${GITHUB_RUN_ID}\", \"version\": \"${GITHUB_REF_NAME#v}\"}}"
+
+  # Aggregate all the assets and ship a release
+  release:
+    needs: [darwin-sign, windows-sign, linux-build]
+    runs-on: linux
+    environment: release
    permissions:
      contents: write
    env:
      GH_TOKEN: ${{ github.token }}
    steps:
      - uses: actions/checkout@v4
-      - name: Create or update Release for tag
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist-darwin
+          path: dist
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist-windows
+          path: dist
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: dist-linux-*
+          path: dist
+          merge-multiple: true
+      - run: find . -type f -not -name 'sha256sum.txt' | xargs sha256sum | tee sha256sum.txt
+        working-directory: dist
+      - name: Create or update Release
        run: |
          RELEASE_VERSION="$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)"
+
          echo "Looking for existing release for ${RELEASE_VERSION}"
          OLD_TAG=$(gh release ls --json name,tagName | jq -r ".[] | select(.name == \"${RELEASE_VERSION}\") | .tagName")
          if [ -n "$OLD_TAG" ]; then
@@ -372,12 +485,5 @@ jobs:
              --generate-notes \
              --prerelease
          fi
-      - name: Trigger downstream release process
-        run: |
-          curl -L \
-            -X POST \
-            -H "Accept: application/vnd.github+json" \
-            -H "Authorization: Bearer ${{ secrets.RELEASE_TOKEN }}" \
-            -H "X-GitHub-Api-Version: 2022-11-28" \
-            https://api.github.com/repos/ollama/${{ vars.RELEASE_REPO }}/dispatches \
-            -d "{\"event_type\": \"trigger-workflow\", \"client_payload\": {\"run_id\": \"${GITHUB_RUN_ID}\", \"version\": \"${GITHUB_REF_NAME#v}\", \"origin\": \"${GITHUB_REPOSITORY}\", \"publish\": \"1\"}}"
+          echo "Uploading artifacts for tag ${GITHUB_REF_NAME}"
+          gh release upload ${GITHUB_REF_NAME} dist/* --clobber
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -36,7 +36,7 @@ jobs:
              | xargs python3 -c "import sys; from pathlib import Path; print(any(Path(x).match(glob) for x in sys.argv[1:] for glob in '$*'.split(' ')))"
          }

-          echo changed=$(changed 'llama/llama.cpp/**/*' 'ml/backend/ggml/ggml/**/*') | tee -a $GITHUB_OUTPUT
+          echo changed=$(changed 'llama/llama.cpp/**' 'ml/backend/ggml/ggml/**') | tee -a $GITHUB_OUTPUT

  linux:
    needs: [changes]
@@ -82,7 +82,7 @@ jobs:
            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
          - preset: ROCm
            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
-            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
+            flags: '-DAMDGPU_TARGETS=gfx1010'
    runs-on: windows
    steps:
      - run: |
@@ -120,9 +120,6 @@ jobs:
          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 $env:GITHUB_ENV -Append
-          echo "HIPCXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "HIP_PLATFORM=amd" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "CMAKE_PREFIX_PATH=$hipPath" | Out-File -FilePath $env:GITHUB_ENV -Append
      - if: ${{ !cancelled() && steps.cache-install.outputs.cache-hit != 'true' }}
        uses: actions/cache/save@v4
        with:
@@ -136,8 +133,8 @@ jobs:
          path: ${{ github.workspace }}\.ccache
          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
      - run: |
-          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
+          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:
--- a/.gitignore
+++ b/.gitignore
@@ -14,3 +14,4 @@ __debug_bin*
 llama/build
 llama/vendor
 /ollama
+/model/testdata/forward
--- a/.golangci.yaml
+++ b/.golangci.yaml
@@ -19,8 +19,8 @@ linters:
    - nolintlint
    - nosprintfhostport
    - staticcheck
+    - tenv
    - unconvert
-    - usetesting
    - wastedassign
    - whitespace
  disable:
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -3,7 +3,6 @@ cmake_minimum_required(VERSION 3.21)
 project(Ollama C CXX)

 include(CheckLanguage)
-include(GNUInstallDirs)

 find_package(Threads REQUIRED)

@@ -52,8 +51,6 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx)

-add_compile_definitions(NDEBUG GGML_VERSION=0x0 GGML_COMMIT=0x0)
-
 set(GGML_CPU ON)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)
@@ -79,13 +76,14 @@ if(CMAKE_CUDA_COMPILER)

    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_INSTALL_DIR} COMPONENT CUDA
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
+        RUNTIME DESTINATION ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
+        LIBRARY DESTINATION ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
    )
 endif()

@@ -116,11 +114,7 @@ if(CMAKE_HIP_COMPILER)

        set(OLLAMA_HIP_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/rocm)
        install(TARGETS ggml-hip
-            RUNTIME_DEPENDENCY_SET rocm
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-        )
-        install(RUNTIME_DEPENDENCY_SET rocm
+            RUNTIME_DEPENDENCIES
                DIRECTORIES ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR}
                PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register drm drm_amdgpu numa elf
                PRE_EXCLUDE_REGEXES ".*"
--- a/CMakePresets.json
+++ b/CMakePresets.json
@@ -6,8 +6,7 @@
      "binaryDir": "${sourceDir}/build",
      "installDir": "${sourceDir}/dist",
      "cacheVariables": {
-        "CMAKE_BUILD_TYPE": "Release",
-        "CMAKE_MSVC_RUNTIME_LIBRARY": "MultiThreaded"
+        "CMAKE_BUILD_TYPE": "Release"
      }
    },
    {
@@ -23,7 +22,7 @@
      "inherits": [ "CUDA" ],
      "cacheVariables": {
        "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
      }
    },
    {
@@ -51,7 +50,6 @@
      "name": "ROCm 6",
      "inherits": [ "ROCm" ],
      "cacheVariables": {
-        "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
      }
    }
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -65,8 +65,7 @@ continuation of the sentence:
 Examples:

      llm/backend/mlx: support the llama architecture
-      CONTRIBUTING: provide clarity on good commit messages, and bad
-      docs: simplify manual installation with shorter curl commands
+      CONTRIBUTING: provide clairity on good commit messages, and bad

 Bad Examples:

--- a/19
+++ b/19
@@ -7,15 +7,10 @@ ARG JETPACK5VERSION=r35.4.1
 ARG JETPACK6VERSION=r36.4.0
 ARG CMAKEVERSION=3.31.2

-# We require gcc v10 minimum.  v10.3 has regressions, so the rockylinux 8.5 AppStream has the latest compatible version
 FROM --platform=linux/amd64 rocm/dev-almalinux-8:${ROCMVERSION}-complete AS base-amd64
 RUN yum install -y yum-utils \
-    && yum-config-manager --add-repo https://dl.rockylinux.org/vault/rocky/8.5/AppStream/\$basearch/os/ \
-    && rpm --import https://dl.rockylinux.org/pub/rocky/RPM-GPG-KEY-Rocky-8 \
-    && dnf install -y yum-utils ccache gcc-toolset-10-gcc-10.2.1-8.2.el8 gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 gcc-toolset-10-binutils-2.35-11.el8 \
    && dnf install -y ccache \
    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH

 FROM --platform=linux/arm64 almalinux:8 AS base-arm64
 # install epel-release for ccache
@@ -86,27 +81,25 @@ RUN go mod download
 COPY . .
 ARG GOFLAGS="'-ldflags=-w -s'"
 ENV CGO_ENABLED=1
-ARG CGO_CFLAGS
-ARG CGO_CXXFLAGS
 RUN --mount=type=cache,target=/root/.cache/go-build \
    go build -trimpath -buildmode=pie -o /bin/ollama .

 FROM --platform=linux/amd64 scratch AS amd64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama
+COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12

 FROM --platform=linux/arm64 scratch AS arm64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/cuda_sbsa
-COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/cuda_jetpack5
-COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/cuda_jetpack6
+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 scratch AS rocm
-COPY --from=rocm-6 dist/lib/ollama /lib/ollama
+COPY --from=rocm-6 dist/lib/ollama/rocm /lib/ollama/rocm

 FROM ${FLAVOR} AS archive
 COPY --from=cpu dist/lib/ollama /lib/ollama
 COPY --from=build /bin/ollama /bin/ollama

-FROM ubuntu:24.04
+FROM ubuntu:20.04
 RUN apt-get update \
    && apt-get install -y ca-certificates \
    && apt-get clean \
--- a/Makefile.sync
+++ b/Makefile.sync
@@ -1,6 +1,6 @@
-UPSTREAM=https://github.com/ggml-org/llama.cpp.git
+UPSTREAM=https://github.com/ggerganov/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=e54d41befcc1575f4c898c5ff4ef43970cead75f
+FETCH_HEAD=e1e8e0991ffd9e99a445c6812bb519d5bac9f4b5

 .PHONY: help
 help:
@@ -12,42 +12,31 @@ help:
 	@echo "    clean                Clean local repository"
 	@echo
 	@echo "Example:"
-	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean apply-patches sync"
+	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean sync"

 .PHONY: sync
-sync: llama/build-info.cpp ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal
+sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml

-llama/build-info.cpp: llama/build-info.cpp.in llama/llama.cpp
-	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' <$< >$@
-
-ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal: ml/backend/ggml/ggml
-	go generate ./$(@D)
+.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
-	rsync -arvzc --delete -f "include LICENSE" -f "merge $@/.rsync-filter" $(addprefix $<,/LICENSE /) $@
+llama/llama.cpp: llama/vendor/
+	rsync -arvzc -f "merge $@/.rsync-filter" $< $@

 .PHONY: ml/backend/ggml/ggml
-ml/backend/ggml/ggml: llama/vendor
-	rsync -arvzc --delete -f "include LICENSE" -f "merge $@/.rsync-filter" $(addprefix $<,/LICENSE /ggml/) $@
+ml/backend/ggml/ggml: llama/vendor/ggml/
+	rsync -arvzc -f "merge $@/.rsync-filter" $< $@

 PATCHES=$(wildcard llama/patches/*.patch)
-PATCHED=$(join $(dir $(PATCHES)), $(addsuffix ed, $(addprefix ., $(notdir $(PATCHES)))))

 .PHONY: apply-patches
 .NOTPARALLEL:
-apply-patches: $(PATCHED)
+apply-patches: $(addsuffix ed, $(PATCHES))

-llama/patches/.%.patched: llama/patches/%.patch
-	@if git -c user.name=nobody -c 'user.email=<>' -C $(WORKDIR) am -3 $(realpath $<); then \
-		touch $@;                                                                           \
-	else                                                                                    \
-		echo "Patch failed. Resolve any conflicts then continue.";                          \
-		echo "1. Run 'git -C $(WORKDIR) am --continue'";                                    \
-		echo "2. Run 'make -f $(lastword $(MAKEFILE_LIST)) format-patches'";                \
-		echo "3. Run 'make -f $(lastword $(MAKEFILE_LIST)) clean apply-patches'";           \
-		exit 1;                                                                             \
-	fi
+%.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)
@@ -68,5 +57,4 @@ format-patches: llama/patches

 .PHONE: clean
 clean: checkout
-	@git -C $(WORKDIR) am --abort || true
-	$(RM) llama/patches/.*.patched
+	$(RM) $(addsuffix ed, $(PATCHES))
--- a/README.md
+++ b/README.md
@@ -1,6 +1,6 @@
 <div align="center">
   <a href="https://ollama.com">
-    <img alt="ollama" width="240" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
+    <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
  </a>
 </div>

@@ -10,7 +10,7 @@ Get up and running with large language models.

 ### macOS

-[Download](https://ollama.com/download/Ollama.dmg)
+[Download](https://ollama.com/download/Ollama-darwin.zip)

 ### Windows

@@ -40,10 +40,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla

 ## Quickstart

-To run and chat with [Gemma 3](https://ollama.com/library/gemma3):
+To run and chat with [Llama 3.2](https://ollama.com/library/llama3.2):

 ```shell
-ollama run gemma3
+ollama run llama3.2
 ```

 ## Model library
@@ -315,7 +315,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
 - [IntelliBar](https://intellibar.app/) (AI-powered assistant for macOS)
 - [Jirapt](https://github.com/AliAhmedNada/jirapt) (Jira Integration to generate issues, tasks, epics)
- [ojira](https://github.com/AliAhmedNada/ojira) (Jira chrome plugin to easily generate descriptions for tasks)
 - [QA-Pilot](https://github.com/reid41/QA-Pilot) (Interactive chat tool that can leverage Ollama models for rapid understanding and navigation of GitHub code repositories)
 - [ChatOllama](https://github.com/sugarforever/chat-ollama) (Open Source Chatbot based on Ollama with Knowledge Bases)
 - [CRAG Ollama Chat](https://github.com/Nagi-ovo/CRAG-Ollama-Chat) (Simple Web Search with Corrective RAG)
@@ -360,7 +359,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Tkinter-based client](https://github.com/chyok/ollama-gui) (Python tkinter-based Client for Ollama)
 - [LLMChat](https://github.com/trendy-design/llmchat) (Privacy focused, 100% local, intuitive all-in-one chat interface)
 - [Local Multimodal AI Chat](https://github.com/Leon-Sander/Local-Multimodal-AI-Chat) (Ollama-based LLM Chat with support for multiple features, including PDF RAG, voice chat, image-based interactions, and integration with OpenAI.)
- [ARGO](https://github.com/xark-argo/argo) (Locally download and run Ollama and Huggingface models with RAG and deep research on Mac/Windows/Linux)
+- [ARGO](https://github.com/xark-argo/argo) (Locally download and run Ollama and Huggingface models with RAG on Mac/Windows/Linux)
 - [OrionChat](https://github.com/EliasPereirah/OrionChat) - OrionChat is a web interface for chatting with different AI providers
 - [G1](https://github.com/bklieger-groq/g1) (Prototype of using prompting strategies to improve the LLM's reasoning through o1-like reasoning chains.)
 - [Web management](https://github.com/lemonit-eric-mao/ollama-web-management) (Web management page)
@@ -405,12 +404,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Writeopia](https://github.com/Writeopia/Writeopia) (Text editor with integration with Ollama)
 - [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
- [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
- [GPTranslate](https://github.com/philberndt/GPTranslate) (A fast and lightweight, AI powered desktop translation application written with Rust and Tauri. Features real-time translation with OpenAI/Azure/Ollama.)
- [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
- [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
- [Mayan EDMS](https://gitlab.com/mayan-edms/mayan-edms) (Open source document management system to organize, tag, search, and automate your files with powerful Ollama driven workflows.)

 ### Cloud

@@ -454,9 +447,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [orbiton](https://github.com/xyproto/orbiton) Configuration-free text editor and IDE with support for tab completion with Ollama.
 - [orca-cli](https://github.com/molbal/orca-cli) Ollama Registry CLI Application - Browse, pull, and download models from Ollama Registry in your terminal.
 - [GGUF-to-Ollama](https://github.com/jonathanhecl/gguf-to-ollama) - Importing GGUF to Ollama made easy (multiplatform)
- [AWS-Strands-With-Ollama](https://github.com/rapidarchitect/ollama_strands) - AWS Strands Agents with Ollama Examples
- [ollama-multirun](https://github.com/attogram/ollama-multirun) - A bash shell script to run a single prompt against any or all of your locally installed ollama models, saving the output and performance statistics as easily navigable web pages. ([Demo](https://attogram.github.io/ai_test_zone/))
- [ollama-bash-toolshed](https://github.com/attogram/ollama-bash-toolshed) - Bash scripts to chat with tool using models. Add new tools to your shed with ease. Runs on Ollama.

 ### Apple Vision Pro

@@ -536,7 +526,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Abso](https://github.com/lunary-ai/abso) (OpenAI-compatible TypeScript SDK for any LLM provider)
 - [Nichey](https://github.com/goodreasonai/nichey) is a Python package for generating custom wikis for your research topic
 - [Ollama for D](https://github.com/kassane/ollama-d)
- [OllamaPlusPlus](https://github.com/HardCodeDev777/OllamaPlusPlus) (Very simple C++ library for Ollama)

 ### Mobile

@@ -593,14 +582,10 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
 - [mcp-llm](https://github.com/sammcj/mcp-llm) (MCP Server to allow LLMs to call other LLMs)
- [SimpleOllamaUnity](https://github.com/HardCodeDev777/SimpleOllamaUnity) (Unity Engine extension for communicating with Ollama in a few lines of code. Also works at runtime)
- [UnityCodeLama](https://github.com/HardCodeDev777/UnityCodeLama) (Unity Edtior tool to analyze scripts via Ollama)
- [NativeMind](https://github.com/NativeMindBrowser/NativeMindExtension) (Private, on-device AI Assistant, no cloud dependencies)
- [GMAI - Gradle Managed AI](https://gmai.premex.se/) (Gradle plugin for automated Ollama lifecycle management during build phases)

 ### Supported backends

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

 ### Observability
 - [Opik](https://www.comet.com/docs/opik/cookbook/ollama) is an open-source platform to debug, evaluate, and monitor your LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Opik supports native intergration to Ollama.
--- a/api/client.go
+++ b/api/client.go
@@ -24,10 +24,7 @@ import (
 	"net/http"
 	"net/url"
 	"runtime"
-	"strconv"
-	"time"

-	"github.com/ollama/ollama/auth"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/version"
@@ -79,14 +76,6 @@ func NewClient(base *url.URL, http *http.Client) *Client {
 	}
 }

-func getAuthorizationToken(ctx context.Context, challenge string) (string, error) {
-	token, err := auth.Sign(ctx, []byte(challenge))
-	if err != nil {
-		return "", err
-	}
-	return token, nil
-}
-
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
 	var reqBody io.Reader
 	var data []byte
@@ -108,21 +97,6 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	}

 	requestURL := c.base.JoinPath(path)
-
-	var token string
-	if envconfig.UseAuth() || c.base.Hostname() == "ollama.com" {
-		now := strconv.FormatInt(time.Now().Unix(), 10)
-		chal := fmt.Sprintf("%s,%s?ts=%s", method, path, now)
-		token, err = getAuthorizationToken(ctx, chal)
-		if err != nil {
-			return err
-		}
-
-		q := requestURL.Query()
-		q.Set("ts", now)
-		requestURL.RawQuery = q.Encode()
-	}
-
 	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), reqBody)
 	if err != nil {
 		return err
@@ -132,10 +106,6 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	request.Header.Set("Accept", "application/json")
 	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))

-	if token != "" {
-		request.Header.Set("Authorization", token)
-	}
-
 	respObj, err := c.http.Do(request)
 	if err != nil {
 		return err
@@ -173,22 +143,6 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	}

 	requestURL := c.base.JoinPath(path)
-
-	var token string
-	if envconfig.UseAuth() || c.base.Hostname() == "ollama.com" {
-		var err error
-		now := strconv.FormatInt(time.Now().Unix(), 10)
-		chal := fmt.Sprintf("%s,%s?ts=%s", method, path, now)
-		token, err = getAuthorizationToken(ctx, chal)
-		if err != nil {
-			return err
-		}
-
-		q := requestURL.Query()
-		q.Set("ts", now)
-		requestURL.RawQuery = q.Encode()
-	}
-
 	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), buf)
 	if err != nil {
 		return err
@@ -198,10 +152,6 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	request.Header.Set("Accept", "application/x-ndjson")
 	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))

-	if token != "" {
-		request.Header.Set("Authorization", token)
-	}
-
 	response, err := c.http.Do(request)
 	if err != nil {
 		return err
@@ -222,6 +172,10 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			return fmt.Errorf("unmarshal: %w", err)
 		}

+		if errorResponse.Error != "" {
+			return errors.New(errorResponse.Error)
+		}
+
 		if response.StatusCode >= http.StatusBadRequest {
 			return StatusError{
 				StatusCode:   response.StatusCode,
@@ -230,10 +184,6 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			}
 		}

-		if errorResponse.Error != "" {
-			return errors.New(errorResponse.Error)
-		}
-
 		if err := fn(bts); err != nil {
 			return err
 		}
--- a/api/client_test.go
+++ b/api/client_test.go
@@ -1,6 +1,7 @@
 package api

 import (
+	"context"
 	"encoding/json"
 	"fmt"
 	"net/http"
@@ -89,16 +90,6 @@ func TestClientStream(t *testing.T) {
 			},
 			wantErr: "mid-stream error",
 		},
-		{
-			name: "http status error takes precedence over general error",
-			responses: []any{
-				testError{
-					message:    "custom error message",
-					statusCode: http.StatusInternalServerError,
-				},
-			},
-			wantErr: "500",
-		},
 		{
 			name: "successful stream completion",
 			responses: []any{
@@ -146,7 +137,7 @@ func TestClientStream(t *testing.T) {
 			client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)

 			var receivedChunks []ChatResponse
-			err := client.stream(t.Context(), http.MethodPost, "/v1/chat", nil, func(chunk []byte) error {
+			err := client.stream(context.Background(), http.MethodPost, "/v1/chat", nil, func(chunk []byte) error {
 				var resp ChatResponse
 				if err := json.Unmarshal(chunk, &resp); err != nil {
 					return fmt.Errorf("failed to unmarshal chunk: %w", err)
@@ -232,7 +223,7 @@ func TestClientDo(t *testing.T) {
 				ID      string `json:"id"`
 				Success bool   `json:"success"`
 			}
-			err := client.do(t.Context(), http.MethodPost, "/v1/messages", nil, &resp)
+			err := client.do(context.Background(), http.MethodPost, "/v1/messages", nil, &resp)

 			if tc.wantErr != "" {
 				if err == nil {
--- a/api/types.go
+++ b/api/types.go
@@ -83,13 +83,6 @@ type GenerateRequest struct {
 	// Options lists model-specific options. For example, temperature can be
 	// set through this field, if the model supports it.
 	Options map[string]any `json:"options"`
-
-	// Think controls whether thinking/reasoning models will think before
-	// responding. Can be a boolean (true/false) or a string ("high", "medium", "low")
-	// for supported models. Needs to be a pointer so we can distinguish between false
-	// (request that thinking _not_ be used) and unset (use the old behavior
-	// before this option was introduced)
-	Think *ThinkValue `json:"think,omitempty"`
 }

 // ChatRequest describes a request sent by [Client.Chat].
@@ -115,11 +108,6 @@ type ChatRequest struct {

 	// Options lists model-specific options.
 	Options map[string]any `json:"options"`
-
-	// Think controls whether thinking/reasoning models will think before
-	// responding. Can be a boolean (true/false) or a string ("high", "medium", "low")
-	// for supported models.
-	Think *ThinkValue `json:"think,omitempty"`
 }

 type Tools []Tool
@@ -138,14 +126,10 @@ func (t Tool) String() string {
 // role ("system", "user", or "assistant"), the content and an optional list
 // of images.
 type Message struct {
-	Role    string `json:"role"`
-	Content string `json:"content"`
-	// Thinking contains the text that was inside thinking tags in the
-	// original model output when ChatRequest.Think is enabled.
-	Thinking  string      `json:"thinking,omitempty"`
+	Role      string      `json:"role"`
+	Content   string      `json:"content"`
 	Images    []ImageData `json:"images,omitempty"`
 	ToolCalls []ToolCall  `json:"tool_calls,omitempty"`
-	ToolName  string      `json:"tool_name,omitempty"`
 }

 func (m *Message) UnmarshalJSON(b []byte) error {
@@ -225,68 +209,20 @@ func (pt PropertyType) String() string {
 	return fmt.Sprintf("%v", []string(pt))
 }

-type ToolProperty struct {
-	AnyOf       []ToolProperty `json:"anyOf,omitempty"`
-	Type        PropertyType   `json:"type"`
-	Items       any            `json:"items,omitempty"`
-	Description string         `json:"description"`
-	Enum        []any          `json:"enum,omitempty"`
-}
-
-// ToTypeScriptType converts a ToolProperty to a TypeScript type string
-func (tp ToolProperty) ToTypeScriptType() string {
-	if len(tp.AnyOf) > 0 {
-		var types []string
-		for _, anyOf := range tp.AnyOf {
-			types = append(types, anyOf.ToTypeScriptType())
-		}
-		return strings.Join(types, " | ")
-	}
-
-	if len(tp.Type) == 0 {
-		return "any"
-	}
-
-	if len(tp.Type) == 1 {
-		return mapToTypeScriptType(tp.Type[0])
-	}
-
-	var types []string
-	for _, t := range tp.Type {
-		types = append(types, mapToTypeScriptType(t))
-	}
-	return strings.Join(types, " | ")
-}
-
-// mapToTypeScriptType maps JSON Schema types to TypeScript types
-func mapToTypeScriptType(jsonType string) string {
-	switch jsonType {
-	case "string":
-		return "string"
-	case "number", "integer":
-		return "number"
-	case "boolean":
-		return "boolean"
-	case "array":
-		return "any[]"
-	case "object":
-		return "Record<string, any>"
-	case "null":
-		return "null"
-	default:
-		return "any"
-	}
-}
-
 type ToolFunction struct {
 	Name        string `json:"name"`
 	Description string `json:"description"`
 	Parameters  struct {
-		Type       string                  `json:"type"`
-		Defs       any                     `json:"$defs,omitempty"`
-		Items      any                     `json:"items,omitempty"`
-		Required   []string                `json:"required"`
-		Properties map[string]ToolProperty `json:"properties"`
+		Type       string   `json:"type"`
+		Defs       any      `json:"$defs,omitempty"`
+		Items      any      `json:"items,omitempty"`
+		Required   []string `json:"required"`
+		Properties map[string]struct {
+			Type        PropertyType `json:"type"`
+			Items       any          `json:"items,omitempty"`
+			Description string       `json:"description"`
+			Enum        []any        `json:"enum,omitempty"`
+		} `json:"properties"`
 	} `json:"parameters"`
 }

@@ -518,14 +454,13 @@ type ListModelResponse struct {

 // ProcessModelResponse is a single model description in [ProcessResponse].
 type ProcessModelResponse struct {
-	Name          string       `json:"name"`
-	Model         string       `json:"model"`
-	Size          int64        `json:"size"`
-	Digest        string       `json:"digest"`
-	Details       ModelDetails `json:"details,omitempty"`
-	ExpiresAt     time.Time    `json:"expires_at"`
-	SizeVRAM      int64        `json:"size_vram"`
-	ContextLength int          `json:"context_length"`
+	Name      string       `json:"name"`
+	Model     string       `json:"model"`
+	Size      int64        `json:"size"`
+	Digest    string       `json:"digest"`
+	Details   ModelDetails `json:"details,omitempty"`
+	ExpiresAt time.Time    `json:"expires_at"`
+	SizeVRAM  int64        `json:"size_vram"`
 }

 type TokenResponse struct {
@@ -543,10 +478,6 @@ type GenerateResponse struct {
 	// Response is the textual response itself.
 	Response string `json:"response"`

-	// Thinking contains the text that was inside thinking tags in the
-	// original model output when ChatRequest.Think is enabled.
-	Thinking string `json:"thinking,omitempty"`
-
 	// Done specifies if the response is complete.
 	Done bool `json:"done"`

@@ -558,8 +489,6 @@ type GenerateResponse struct {
 	Context []int `json:"context,omitempty"`

 	Metrics
-
-	ToolCalls []ToolCall `json:"tool_calls,omitempty"`
 }

 // ModelDetails provides details about a model.
@@ -729,113 +658,6 @@ func DefaultOptions() Options {
 	}
 }

-// ThinkValue represents a value that can be a boolean or a string ("high", "medium", "low")
-type ThinkValue struct {
-	// Value can be a bool or string
-	Value interface{}
-}
-
-// IsValid checks if the ThinkValue is valid
-func (t *ThinkValue) IsValid() bool {
-	if t == nil || t.Value == nil {
-		return true // nil is valid (means not set)
-	}
-
-	switch v := t.Value.(type) {
-	case bool:
-		return true
-	case string:
-		return v == "high" || v == "medium" || v == "low"
-	default:
-		return false
-	}
-}
-
-// IsBool returns true if the value is a boolean
-func (t *ThinkValue) IsBool() bool {
-	if t == nil || t.Value == nil {
-		return false
-	}
-	_, ok := t.Value.(bool)
-	return ok
-}
-
-// IsString returns true if the value is a string
-func (t *ThinkValue) IsString() bool {
-	if t == nil || t.Value == nil {
-		return false
-	}
-	_, ok := t.Value.(string)
-	return ok
-}
-
-// Bool returns the value as a bool (true if enabled in any way)
-func (t *ThinkValue) Bool() bool {
-	if t == nil || t.Value == nil {
-		return false
-	}
-
-	switch v := t.Value.(type) {
-	case bool:
-		return v
-	case string:
-		// Any string value ("high", "medium", "low") means thinking is enabled
-		return v == "high" || v == "medium" || v == "low"
-	default:
-		return false
-	}
-}
-
-// String returns the value as a string
-func (t *ThinkValue) String() string {
-	if t == nil || t.Value == nil {
-		return ""
-	}
-
-	switch v := t.Value.(type) {
-	case string:
-		return v
-	case bool:
-		if v {
-			return "medium" // Default level when just true
-		}
-		return ""
-	default:
-		return ""
-	}
-}
-
-// UnmarshalJSON implements json.Unmarshaler
-func (t *ThinkValue) UnmarshalJSON(data []byte) error {
-	// Try to unmarshal as bool first
-	var b bool
-	if err := json.Unmarshal(data, &b); err == nil {
-		t.Value = b
-		return nil
-	}
-
-	// Try to unmarshal as string
-	var s string
-	if err := json.Unmarshal(data, &s); err == nil {
-		// Validate string values
-		if s != "high" && s != "medium" && s != "low" {
-			return fmt.Errorf("invalid think value: %q (must be \"high\", \"medium\", \"low\", true, or false)", s)
-		}
-		t.Value = s
-		return nil
-	}
-
-	return fmt.Errorf("think must be a boolean or string (\"high\", \"medium\", \"low\")")
-}
-
-// MarshalJSON implements json.Marshaler
-func (t *ThinkValue) MarshalJSON() ([]byte, error) {
-	if t == nil || t.Value == nil {
-		return []byte("null"), nil
-	}
-	return json.Marshal(t.Value)
-}
-
 type Duration struct {
 	time.Duration
 }
--- a/api/types_test.go
+++ b/api/types_test.go
@@ -372,67 +372,3 @@ func TestPropertyType_MarshalJSON(t *testing.T) {
 		})
 	}
 }
-
-func TestThinking_UnmarshalJSON(t *testing.T) {
-	tests := []struct {
-		name             string
-		input            string
-		expectedThinking *ThinkValue
-		expectedError    bool
-	}{
-		{
-			name:             "true",
-			input:            `{ "think": true }`,
-			expectedThinking: &ThinkValue{Value: true},
-		},
-		{
-			name:             "false",
-			input:            `{ "think": false }`,
-			expectedThinking: &ThinkValue{Value: false},
-		},
-		{
-			name:             "unset",
-			input:            `{ }`,
-			expectedThinking: nil,
-		},
-		{
-			name:             "string_high",
-			input:            `{ "think": "high" }`,
-			expectedThinking: &ThinkValue{Value: "high"},
-		},
-		{
-			name:             "string_medium",
-			input:            `{ "think": "medium" }`,
-			expectedThinking: &ThinkValue{Value: "medium"},
-		},
-		{
-			name:             "string_low",
-			input:            `{ "think": "low" }`,
-			expectedThinking: &ThinkValue{Value: "low"},
-		},
-		{
-			name:             "invalid_string",
-			input:            `{ "think": "invalid" }`,
-			expectedThinking: nil,
-			expectedError:    true,
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			var req GenerateRequest
-			err := json.Unmarshal([]byte(test.input), &req)
-			if test.expectedError {
-				require.Error(t, err)
-			} else {
-				require.NoError(t, err)
-				if test.expectedThinking == nil {
-					assert.Nil(t, req.Think)
-				} else {
-					require.NotNil(t, req.Think)
-					assert.Equal(t, test.expectedThinking.Value, req.Think.Value)
-				}
-			}
-		})
-	}
-}
--- a/api/types_typescript_test.go
+++ b/api/types_typescript_test.go
@@ -1,142 +0,0 @@
-package api
-
-import (
-	"testing"
-)
-
-func TestToolParameterToTypeScriptType(t *testing.T) {
-	tests := []struct {
-		name     string
-		param    ToolProperty
-		expected string
-	}{
-		{
-			name: "single string type",
-			param: ToolProperty{
-				Type: PropertyType{"string"},
-			},
-			expected: "string",
-		},
-		{
-			name: "single number type",
-			param: ToolProperty{
-				Type: PropertyType{"number"},
-			},
-			expected: "number",
-		},
-		{
-			name: "integer maps to number",
-			param: ToolProperty{
-				Type: PropertyType{"integer"},
-			},
-			expected: "number",
-		},
-		{
-			name: "boolean type",
-			param: ToolProperty{
-				Type: PropertyType{"boolean"},
-			},
-			expected: "boolean",
-		},
-		{
-			name: "array type",
-			param: ToolProperty{
-				Type: PropertyType{"array"},
-			},
-			expected: "any[]",
-		},
-		{
-			name: "object type",
-			param: ToolProperty{
-				Type: PropertyType{"object"},
-			},
-			expected: "Record<string, any>",
-		},
-		{
-			name: "null type",
-			param: ToolProperty{
-				Type: PropertyType{"null"},
-			},
-			expected: "null",
-		},
-		{
-			name: "multiple types as union",
-			param: ToolProperty{
-				Type: PropertyType{"string", "number"},
-			},
-			expected: "string | number",
-		},
-		{
-			name: "string or null union",
-			param: ToolProperty{
-				Type: PropertyType{"string", "null"},
-			},
-			expected: "string | null",
-		},
-		{
-			name: "anyOf with single types",
-			param: ToolProperty{
-				AnyOf: []ToolProperty{
-					{Type: PropertyType{"string"}},
-					{Type: PropertyType{"number"}},
-				},
-			},
-			expected: "string | number",
-		},
-		{
-			name: "anyOf with multiple types in each branch",
-			param: ToolProperty{
-				AnyOf: []ToolProperty{
-					{Type: PropertyType{"string", "null"}},
-					{Type: PropertyType{"number"}},
-				},
-			},
-			expected: "string | null | number",
-		},
-		{
-			name: "nested anyOf",
-			param: ToolProperty{
-				AnyOf: []ToolProperty{
-					{Type: PropertyType{"boolean"}},
-					{
-						AnyOf: []ToolProperty{
-							{Type: PropertyType{"string"}},
-							{Type: PropertyType{"number"}},
-						},
-					},
-				},
-			},
-			expected: "boolean | string | number",
-		},
-		{
-			name: "empty type returns any",
-			param: ToolProperty{
-				Type: PropertyType{},
-			},
-			expected: "any",
-		},
-		{
-			name: "unknown type maps to any",
-			param: ToolProperty{
-				Type: PropertyType{"unknown_type"},
-			},
-			expected: "any",
-		},
-		{
-			name: "multiple types including array",
-			param: ToolProperty{
-				Type: PropertyType{"string", "array", "null"},
-			},
-			expected: "string | any[] | null",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			result := tt.param.ToTypeScriptType()
-			if result != tt.expected {
-				t.Errorf("ToTypeScriptType() = %q, want %q", result, tt.expected)
-			}
-		})
-	}
-}
--- a/app/lifecycle/logging.go
+++ b/app/lifecycle/logging.go
@@ -4,14 +4,20 @@ import (
 	"fmt"
 	"log/slog"
 	"os"
+	"path/filepath"
 	"strconv"
 	"strings"

 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/logutil"
 )

 func InitLogging() {
+	level := slog.LevelInfo
+
+	if envconfig.Debug() {
+		level = slog.LevelDebug
+	}
+
 	var logFile *os.File
 	var err error
 	// Detect if we're a GUI app on windows, and if not, send logs to console
@@ -27,8 +33,20 @@ func InitLogging() {
 			return
 		}
 	}
+	handler := slog.NewTextHandler(logFile, &slog.HandlerOptions{
+		Level:     level,
+		AddSource: true,
+		ReplaceAttr: func(_ []string, attr slog.Attr) slog.Attr {
+			if attr.Key == slog.SourceKey {
+				source := attr.Value.Any().(*slog.Source)
+				source.File = filepath.Base(source.File)
+			}
+			return attr
+		},
+	})
+
+	slog.SetDefault(slog.New(handler))

-	slog.SetDefault(logutil.NewLogger(logFile, envconfig.LogLevel()))
 	slog.Info("ollama app started")
 }

--- a/benchmark/server_benchmark_test.go
+++ b/benchmark/server_benchmark_test.go
@@ -0,0 +1,178 @@
+package benchmark
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Command line flags
+var modelFlag string
+
+func init() {
+	flag.StringVar(&modelFlag, "m", "", "Name of the model to benchmark")
+	flag.Lookup("m").DefValue = "model"
+}
+
+// modelName returns the model name from flags, failing the test if not set
+func modelName(b *testing.B) string {
+	if modelFlag == "" {
+		b.Fatal("Error: -m flag is required for benchmark tests")
+	}
+	return modelFlag
+}
+
+type TestCase struct {
+	name      string
+	prompt    string
+	maxTokens int
+}
+
+// runGenerateBenchmark contains the common generate and metrics logic
+func runGenerateBenchmark(b *testing.B, ctx context.Context, client *api.Client, req *api.GenerateRequest) {
+	start := time.Now()
+	var ttft time.Duration
+	var metrics api.Metrics
+
+	err := client.Generate(ctx, req, func(resp api.GenerateResponse) error {
+		if ttft == 0 && resp.Response != "" {
+			ttft = time.Since(start)
+		}
+		if resp.Done {
+			metrics = resp.Metrics
+		}
+		return nil
+	})
+
+	// Report custom metrics as part of the benchmark results
+	b.ReportMetric(float64(ttft.Milliseconds()), "ttft_ms")
+	b.ReportMetric(float64(metrics.LoadDuration.Milliseconds()), "load_ms")
+
+	// Token throughput metrics
+	promptThroughput := float64(metrics.PromptEvalCount) / metrics.PromptEvalDuration.Seconds()
+	genThroughput := float64(metrics.EvalCount) / metrics.EvalDuration.Seconds()
+	b.ReportMetric(promptThroughput, "prompt_tok/s")
+	b.ReportMetric(genThroughput, "gen_tok/s")
+
+	// Token counts
+	b.ReportMetric(float64(metrics.PromptEvalCount), "prompt_tokens")
+	b.ReportMetric(float64(metrics.EvalCount), "gen_tokens")
+	if err != nil {
+		b.Fatal(err)
+	}
+}
+
+// BenchmarkColdStart runs benchmarks with model loading from cold state
+func BenchmarkColdStart(b *testing.B) {
+	client := setup(b)
+	tests := []TestCase{
+		{"short_prompt", "Write a long story", 100},
+		{"medium_prompt", "Write a detailed economic analysis", 500},
+		{"long_prompt", "Write a comprehensive AI research paper", 1000},
+	}
+	m := modelName(b)
+
+	for _, tt := range tests {
+		b.Run(fmt.Sprintf("%s/cold/%s", m, tt.name), func(b *testing.B) {
+			ctx := context.Background()
+
+			// Set number of tokens as our throughput metric
+			b.SetBytes(int64(tt.maxTokens))
+
+			for b.Loop() {
+				b.StopTimer()
+				// Ensure model is unloaded before each iteration
+				unload(client, m, b)
+				b.StartTimer()
+
+				req := &api.GenerateRequest{
+					Model:   m,
+					Prompt:  tt.prompt,
+					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
+				}
+
+				runGenerateBenchmark(b, ctx, client, req)
+			}
+		})
+	}
+}
+
+// BenchmarkWarmStart runs benchmarks with pre-loaded model
+func BenchmarkWarmStart(b *testing.B) {
+	client := setup(b)
+	tests := []TestCase{
+		{"short_prompt", "Write a long story", 100},
+		{"medium_prompt", "Write a detailed economic analysis", 500},
+		{"long_prompt", "Write a comprehensive AI research paper", 1000},
+	}
+	m := modelName(b)
+
+	for _, tt := range tests {
+		b.Run(fmt.Sprintf("%s/warm/%s", m, tt.name), func(b *testing.B) {
+			ctx := context.Background()
+
+			// Pre-warm the model
+			warmup(client, m, tt.prompt, b)
+
+			// Set number of tokens as our throughput metric
+			b.SetBytes(int64(tt.maxTokens))
+
+			for b.Loop() {
+				req := &api.GenerateRequest{
+					Model:   m,
+					Prompt:  tt.prompt,
+					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
+				}
+
+				runGenerateBenchmark(b, ctx, client, req)
+			}
+		})
+	}
+}
+
+// setup verifies server and model availability
+func setup(b *testing.B) *api.Client {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		b.Fatal(err)
+	}
+	if _, err := client.Show(context.Background(), &api.ShowRequest{Model: modelName(b)}); err != nil {
+		b.Fatalf("Model unavailable: %v", err)
+	}
+
+	return client
+}
+
+// warmup ensures the model is loaded and warmed up
+func warmup(client *api.Client, model string, prompt string, b *testing.B) {
+	for range 3 {
+		err := client.Generate(
+			context.Background(),
+			&api.GenerateRequest{
+				Model:   model,
+				Prompt:  prompt,
+				Options: map[string]any{"num_predict": 50, "temperature": 0.1},
+			},
+			func(api.GenerateResponse) error { return nil },
+		)
+		if err != nil {
+			b.Logf("Error during model warm-up: %v", err)
+		}
+	}
+}
+
+// unload forces model unloading using KeepAlive: 0 parameter
+func unload(client *api.Client, model string, b *testing.B) {
+	req := &api.GenerateRequest{
+		Model:     model,
+		KeepAlive: &api.Duration{Duration: 0},
+	}
+	if err := client.Generate(context.Background(), req, func(api.GenerateResponse) error { return nil }); err != nil {
+		b.Logf("Unload error: %v", err)
+	}
+	time.Sleep(1 * time.Second)
+}
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -39,7 +39,6 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
-	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/runner"
 	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/model"
@@ -47,23 +46,6 @@ import (
 	"github.com/ollama/ollama/version"
 )

-// ensureThinkingSupport emits a warning if the model does not advertise thinking support
-func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
-	if name == "" {
-		return
-	}
-	resp, err := client.Show(ctx, &api.ShowRequest{Model: name})
-	if err != nil {
-		return
-	}
-	for _, cap := range resp.Capabilities {
-		if cap == model.CapabilityThinking {
-			return
-		}
-	}
-	fmt.Fprintf(os.Stderr, "warning: model %q does not support thinking output\n", name)
-}
-
 var errModelfileNotFound = errors.New("specified Modelfile wasn't found")

 func getModelfileName(cmd *cobra.Command) (string, error) {
@@ -283,9 +265,6 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 	req := &api.GenerateRequest{
 		Model:     opts.Model,
 		KeepAlive: opts.KeepAlive,
-
-		// pass Think here so we fail before getting to the chat prompt if the model doesn't support it
-		Think: opts.Think,
 	}

 	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
@@ -320,34 +299,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.Format = format

-	thinkFlag := cmd.Flags().Lookup("think")
-	if thinkFlag.Changed {
-		thinkStr, err := cmd.Flags().GetString("think")
-		if err != nil {
-			return err
-		}
-
-		// Handle different values for --think
-		switch thinkStr {
-		case "", "true":
-			// --think or --think=true
-			opts.Think = &api.ThinkValue{Value: true}
-		case "false":
-			opts.Think = &api.ThinkValue{Value: false}
-		case "high", "medium", "low":
-			opts.Think = &api.ThinkValue{Value: thinkStr}
-		default:
-			return fmt.Errorf("invalid value for --think: %q (must be true, false, high, medium, or low)", thinkStr)
-		}
-	} else {
-		opts.Think = nil
-	}
-	hidethinking, err := cmd.Flags().GetBool("hidethinking")
-	if err != nil {
-		return err
-	}
-	opts.HideThinking = hidethinking
-
 	keepAlive, err := cmd.Flags().GetString("keepalive")
 	if err != nil {
 		return err
@@ -411,11 +362,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}

-	opts.Think, err = inferThinkingOption(&info.Capabilities, &opts, thinkFlag.Changed)
-	if err != nil {
-		return err
-	}
-
 	opts.MultiModal = slices.Contains(info.Capabilities, model.CapabilityVision)

 	// TODO: remove the projector info and vision info checks below,
@@ -595,13 +541,12 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 			} else {
 				until = format.HumanTime(m.ExpiresAt, "Never")
 			}
-			ctxStr := strconv.Itoa(m.ContextLength)
-			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, ctxStr, until})
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, until})
 		}
 	}

 	table := tablewriter.NewWriter(os.Stdout)
-	table.SetHeader([]string{"NAME", "ID", "SIZE", "PROCESSOR", "CONTEXT", "UNTIL"})
+	table.SetHeader([]string{"NAME", "ID", "SIZE", "PROCESSOR", "UNTIL"})
 	table.SetHeaderAlignment(tablewriter.ALIGN_LEFT)
 	table.SetAlignment(tablewriter.ALIGN_LEFT)
 	table.SetHeaderLine(false)
@@ -802,38 +747,11 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 				case float64:
 					v = fmt.Sprintf("%g", vData)
 				case []any:
-					targetWidth := 10 // Small width where we are displaying the data in a column
-
-					var itemsToShow int
-					totalWidth := 1 // Start with 1 for opening bracket
-
-					// Find how many we can fit
-					for i := range vData {
-						itemStr := fmt.Sprintf("%v", vData[i])
-						width := runewidth.StringWidth(itemStr)
-
-						// Add separator width (", ") for all items except the first
-						if i > 0 {
-							width += 2
-						}
-
-						// Check if adding this item would exceed our width limit
-						if totalWidth+width > targetWidth && i > 0 {
-							break
-						}
-
-						totalWidth += width
-						itemsToShow++
-					}
-
-					// Format the output
-					if itemsToShow < len(vData) {
-						v = fmt.Sprintf("%v", vData[:itemsToShow])
-						v = strings.TrimSuffix(v, "]")
-						v += fmt.Sprintf(" ...+%d more]", len(vData)-itemsToShow)
-					} else {
-						v = fmt.Sprintf("%v", vData)
+					n := 3
+					if len(vData) < n {
+						n = len(vData)
 					}
+					v = fmt.Sprintf("%v", vData[:n])
 				default:
 					v = fmt.Sprintf("%T", vData)
 				}
@@ -854,19 +772,10 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {

 	head := func(s string, n int) (rows [][]string) {
 		scanner := bufio.NewScanner(strings.NewReader(s))
-		count := 0
-		for scanner.Scan() {
-			text := strings.TrimSpace(scanner.Text())
-			if text == "" {
-				continue
+		for scanner.Scan() && (len(rows) < n || n < 0) {
+			if text := scanner.Text(); text != "" {
+				rows = append(rows, []string{"", strings.TrimSpace(text)})
 			}
-			count++
-			if n < 0 || count <= n {
-				rows = append(rows, []string{"", text})
-			}
-		}
-		if n >= 0 && count > n {
-			rows = append(rows, []string{"", "..."})
 		}
 		return
 	}
@@ -978,19 +887,17 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 type generateContextKey string

 type runOptions struct {
-	Model        string
-	ParentModel  string
-	Prompt       string
-	Messages     []api.Message
-	WordWrap     bool
-	Format       string
-	System       string
-	Images       []api.ImageData
-	Options      map[string]any
-	MultiModal   bool
-	KeepAlive    *api.Duration
-	Think        *api.ThinkValue
-	HideThinking bool
+	Model       string
+	ParentModel string
+	Prompt      string
+	Messages    []api.Message
+	WordWrap    bool
+	Format      string
+	System      string
+	Images      []api.ImageData
+	Options     map[string]any
+	MultiModal  bool
+	KeepAlive   *api.Duration
 }

 type displayResponseState struct {
@@ -1029,11 +936,10 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 				}

 				switch ch {
-				case ' ', '\t':
+				case ' ':
 					state.wordBuffer = ""
-				case '\n', '\r':
+				case '\n':
 					state.lineLength = 0
-					state.wordBuffer = ""
 				default:
 					state.wordBuffer += string(ch)
 				}
@@ -1047,26 +953,6 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 	}
 }

-func thinkingOutputOpeningText(plainText bool) string {
-	text := "Thinking...\n"
-
-	if plainText {
-		return text
-	}
-
-	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault + readline.ColorGrey
-}
-
-func thinkingOutputClosingText(plainText bool) string {
-	text := "...done thinking.\n\n"
-
-	if plainText {
-		return text
-	}
-
-	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault
-}
-
 func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -1091,55 +977,19 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	}()

 	var state *displayResponseState = &displayResponseState{}
-	var thinkingContent strings.Builder
 	var latest api.ChatResponse
 	var fullResponse strings.Builder
-	var thinkTagOpened bool = false
-	var thinkTagClosed bool = false
-
-	role := "assistant"
+	var role string

 	fn := func(response api.ChatResponse) error {
-		if response.Message.Content != "" || !opts.HideThinking {
-			p.StopAndClear()
-		}
+		p.StopAndClear()

 		latest = response

 		role = response.Message.Role
-		if response.Message.Thinking != "" && !opts.HideThinking {
-			if !thinkTagOpened {
-				fmt.Print(thinkingOutputOpeningText(false))
-				thinkTagOpened = true
-				thinkTagClosed = false
-			}
-			thinkingContent.WriteString(response.Message.Thinking)
-			displayResponse(response.Message.Thinking, opts.WordWrap, state)
-		}
-
 		content := response.Message.Content
-		if thinkTagOpened && !thinkTagClosed && (content != "" || len(response.Message.ToolCalls) > 0) {
-			if !strings.HasSuffix(thinkingContent.String(), "\n") {
-				fmt.Println()
-			}
-			fmt.Print(thinkingOutputClosingText(false))
-			thinkTagOpened = false
-			thinkTagClosed = true
-			state = &displayResponseState{}
-		}
-		// purposefully not putting thinking blocks in the response, which would
-		// only be needed if we later added tool calling to the cli (they get
-		// filtered out anyway since current models don't expect them unless you're
-		// about to finish some tool calls)
 		fullResponse.WriteString(content)

-		if response.Message.ToolCalls != nil {
-			toolCalls := response.Message.ToolCalls
-			if len(toolCalls) > 0 {
-				fmt.Print(renderToolCalls(toolCalls, false))
-			}
-		}
-
 		displayResponse(content, opts.WordWrap, state)

 		return nil
@@ -1154,7 +1004,6 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		Messages: opts.Messages,
 		Format:   json.RawMessage(opts.Format),
 		Options:  opts.Options,
-		Think:    opts.Think,
 	}

 	if opts.KeepAlive != nil {
@@ -1165,14 +1014,6 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		if errors.Is(err, context.Canceled) {
 			return nil, nil
 		}
-
-		// this error should ideally be wrapped properly by the client
-		if strings.Contains(err.Error(), "upstream error") {
-			p.StopAndClear()
-			fmt.Println("An error occurred while processing your message. Please try again.")
-			fmt.Println()
-			return nil, nil
-		}
 		return nil, err
 	}

@@ -1224,49 +1065,15 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	}()

 	var state *displayResponseState = &displayResponseState{}
-	var thinkingContent strings.Builder
-	var thinkTagOpened bool = false
-	var thinkTagClosed bool = false
-
-	plainText := !term.IsTerminal(int(os.Stdout.Fd()))

 	fn := func(response api.GenerateResponse) error {
+		p.StopAndClear()
+
 		latest = response
 		content := response.Response

-		if response.Response != "" || !opts.HideThinking {
-			p.StopAndClear()
-		}
-
-		if response.Thinking != "" && !opts.HideThinking {
-			if !thinkTagOpened {
-				fmt.Print(thinkingOutputOpeningText(plainText))
-				thinkTagOpened = true
-				thinkTagClosed = false
-			}
-			thinkingContent.WriteString(response.Thinking)
-			displayResponse(response.Thinking, opts.WordWrap, state)
-		}
-
-		if thinkTagOpened && !thinkTagClosed && (content != "" || len(response.ToolCalls) > 0) {
-			if !strings.HasSuffix(thinkingContent.String(), "\n") {
-				fmt.Println()
-			}
-			fmt.Print(thinkingOutputClosingText(plainText))
-			thinkTagOpened = false
-			thinkTagClosed = true
-			state = &displayResponseState{}
-		}
-
 		displayResponse(content, opts.WordWrap, state)

-		if response.ToolCalls != nil {
-			toolCalls := response.ToolCalls
-			if len(toolCalls) > 0 {
-				fmt.Print(renderToolCalls(toolCalls, plainText))
-			}
-		}
-
 		return nil
 	}

@@ -1290,7 +1097,6 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		System:    opts.System,
 		Options:   opts.Options,
 		KeepAlive: opts.KeepAlive,
-		Think:     opts.Think,
 	}

 	if err := client.Generate(ctx, &request, fn); err != nil {
@@ -1394,11 +1200,11 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 		return err
 	}
 	if err := client.Heartbeat(cmd.Context()); err != nil {
-		if !(strings.Contains(err.Error(), " refused") || strings.Contains(err.Error(), "could not connect")) {
+		if !strings.Contains(err.Error(), " refused") {
 			return err
 		}
 		if err := startApp(cmd.Context(), client); err != nil {
-			return fmt.Errorf("ollama server not responding - %w", err)
+			return errors.New("could not connect to ollama app, is it running?")
 		}
 	}
 	return nil
@@ -1469,14 +1275,14 @@ func NewCLI() *cobra.Command {

 	createCmd := &cobra.Command{
 		Use:     "create MODEL",
-		Short:   "Create a model",
+		Short:   "Create a model from a Modelfile",
 		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    CreateHandler,
 	}

-	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\")")
-	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
+	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\"")
+	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_0)")

 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1506,9 +1312,6 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 	runCmd.Flags().String("format", "", "Response format (e.g. json)")
-	runCmd.Flags().String("think", "", "Enable thinking mode: true/false or high/medium/low for supported models")
-	runCmd.Flags().Lookup("think").NoOptDefVal = "true"
-	runCmd.Flags().Bool("hidethinking", false, "Hide thinking output (if provided)")

 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",
@@ -1560,6 +1363,7 @@ func NewCLI() *cobra.Command {
 		PreRunE: checkServerHeartbeat,
 		RunE:    ListRunningHandler,
 	}
+
 	copyCmd := &cobra.Command{
 		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
@@ -1612,7 +1416,6 @@ func NewCLI() *cobra.Command {
 			appendEnvDocs(cmd, []envconfig.EnvVar{
 				envVars["OLLAMA_DEBUG"],
 				envVars["OLLAMA_HOST"],
-				envVars["OLLAMA_CONTEXT_LENGTH"],
 				envVars["OLLAMA_KEEP_ALIVE"],
 				envVars["OLLAMA_MAX_LOADED_MODELS"],
 				envVars["OLLAMA_MAX_QUEUE"],
@@ -1649,70 +1452,3 @@ func NewCLI() *cobra.Command {

 	return rootCmd
 }
-
-// If the user has explicitly set thinking options, either through the CLI or
-// through the `/set think` or `set nothink` interactive options, then we
-// respect them. Otherwise, we check model capabilities to see if the model
-// supports thinking. If the model does support thinking, we enable it.
-// Otherwise, we unset the thinking option (which is different than setting it
-// to false).
-//
-// If capabilities are not provided, we fetch them from the server.
-func inferThinkingOption(caps *[]model.Capability, runOpts *runOptions, explicitlySetByUser bool) (*api.ThinkValue, error) {
-	if explicitlySetByUser {
-		return runOpts.Think, nil
-	}
-
-	if caps == nil {
-		client, err := api.ClientFromEnvironment()
-		if err != nil {
-			return nil, err
-		}
-		ret, err := client.Show(context.Background(), &api.ShowRequest{
-			Model: runOpts.Model,
-		})
-		if err != nil {
-			return nil, err
-		}
-		caps = &ret.Capabilities
-	}
-
-	thinkingSupported := false
-	for _, cap := range *caps {
-		if cap == model.CapabilityThinking {
-			thinkingSupported = true
-		}
-	}
-
-	if thinkingSupported {
-		return &api.ThinkValue{Value: true}, nil
-	}
-
-	return nil, nil
-}
-
-func renderToolCalls(toolCalls []api.ToolCall, plainText bool) string {
-	out := ""
-	formatExplanation := ""
-	formatValues := ""
-	if !plainText {
-		formatExplanation = readline.ColorGrey + readline.ColorBold
-		formatValues = readline.ColorDefault
-		out += formatExplanation
-	}
-	for i, toolCall := range toolCalls {
-		argsAsJSON, err := json.Marshal(toolCall.Function.Arguments)
-		if err != nil {
-			return ""
-		}
-		if i > 0 {
-			out += "\n"
-		}
-		// all tool calls are unexpected since we don't currently support registering any in the CLI
-		out += fmt.Sprintf("  Model called a non-existent function '%s()' with arguments: %s", formatValues+toolCall.Function.Name+formatExplanation, formatValues+string(argsAsJSON)+formatExplanation)
-	}
-	if !plainText {
-		out += readline.ColorDefault
-	}
-	return out
-}
--- a/cmd/cmd_test.go
+++ b/cmd/cmd_test.go
@@ -2,6 +2,7 @@ package cmd

 import (
 	"bytes"
+	"context"
 	"encoding/json"
 	"io"
 	"net/http"
@@ -225,7 +226,6 @@ Weigh anchor!
  System
    You are a pirate!    
    Ahoy, matey!         
-    ...                  

 `
 		if diff := cmp.Diff(expect, b.String()); diff != "" {
@@ -337,7 +337,7 @@ func TestDeleteHandler(t *testing.T) {
 	t.Cleanup(mockServer.Close)

 	cmd := &cobra.Command{}
-	cmd.SetContext(t.Context())
+	cmd.SetContext(context.TODO())
 	if err := DeleteHandler(cmd, []string{"test-model"}); err != nil {
 		t.Fatalf("DeleteHandler failed: %v", err)
 	}
@@ -399,6 +399,11 @@ func TestGetModelfileName(t *testing.T) {
 			var expectedFilename string

 			if tt.fileExists {
+				tempDir, err := os.MkdirTemp("", "modelfiledir")
+				defer os.RemoveAll(tempDir)
+				if err != nil {
+					t.Fatalf("temp modelfile dir creation failed: %v", err)
+				}
 				var fn string
 				if tt.modelfileName != "" {
 					fn = tt.modelfileName
@@ -406,7 +411,7 @@ func TestGetModelfileName(t *testing.T) {
 					fn = "Modelfile"
 				}

-				tempFile, err := os.CreateTemp(t.TempDir(), fn)
+				tempFile, err := os.CreateTemp(tempDir, fn)
 				if err != nil {
 					t.Fatalf("temp modelfile creation failed: %v", err)
 				}
@@ -525,7 +530,7 @@ func TestPushHandler(t *testing.T) {

 			cmd := &cobra.Command{}
 			cmd.Flags().Bool("insecure", false, "")
-			cmd.SetContext(t.Context())
+			cmd.SetContext(context.TODO())

 			// Redirect stderr to capture progress output
 			oldStderr := os.Stderr
@@ -630,7 +635,7 @@ func TestListHandler(t *testing.T) {
 			t.Setenv("OLLAMA_HOST", mockServer.URL)

 			cmd := &cobra.Command{}
-			cmd.SetContext(t.Context())
+			cmd.SetContext(context.TODO())

 			// Capture stdout
 			oldStdout := os.Stdout
@@ -725,7 +730,7 @@ func TestCreateHandler(t *testing.T) {
 			}))
 			t.Setenv("OLLAMA_HOST", mockServer.URL)
 			t.Cleanup(mockServer.Close)
-			tempFile, err := os.CreateTemp(t.TempDir(), "modelfile")
+			tempFile, err := os.CreateTemp("", "modelfile")
 			if err != nil {
 				t.Fatal(err)
 			}
@@ -745,7 +750,7 @@ func TestCreateHandler(t *testing.T) {
 			}

 			cmd.Flags().Bool("insecure", false, "")
-			cmd.SetContext(t.Context())
+			cmd.SetContext(context.TODO())

 			// Redirect stderr to capture progress output
 			oldStderr := os.Stderr
--- a/cmd/interactive.go
+++ b/cmd/interactive.go
@@ -44,7 +44,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "Use \"\"\" to begin a multi-line message.")

 		if opts.MultiModal {
-			fmt.Fprintf(os.Stderr, "Use %s to include .jpg, .png, or .webp images.\n", filepath.FromSlash("/path/to/file"))
+			fmt.Fprintf(os.Stderr, "Use %s to include .jpg or .png images.\n", filepath.FromSlash("/path/to/file"))
 		}

 		fmt.Fprintln(os.Stderr, "")
@@ -62,8 +62,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /set noformat          Disable formatting")
 		fmt.Fprintln(os.Stderr, "  /set verbose           Show LLM stats")
 		fmt.Fprintln(os.Stderr, "  /set quiet             Disable LLM stats")
-		fmt.Fprintln(os.Stderr, "  /set think             Enable thinking")
-		fmt.Fprintln(os.Stderr, "  /set nothink           Disable thinking")
 		fmt.Fprintln(os.Stderr, "")
 	}

@@ -130,7 +128,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {

 	var sb strings.Builder
 	var multiline MultilineState
-	var thinkExplicitlySet bool = opts.Think != nil

 	for {
 		line, err := scanner.Readline()
@@ -198,19 +195,11 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Model = args[1]
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
-			opts.Think, err = inferThinkingOption(nil, &opts, thinkExplicitlySet)
-			if err != nil {
-				return err
-			}
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				if strings.Contains(err.Error(), "not found") {
 					fmt.Printf("error: %v\n", err)
 					continue
 				}
-				if strings.Contains(err.Error(), "does not support thinking") {
-					fmt.Printf("error: %v\n", err)
-					continue
-				}
 				return err
 			}
 			continue
@@ -271,35 +260,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						return err
 					}
 					fmt.Println("Set 'quiet' mode.")
-				case "think":
-					thinkValue := api.ThinkValue{Value: true}
-					var maybeLevel string
-					if len(args) > 2 {
-						maybeLevel = args[2]
-					}
-					if maybeLevel != "" {
-						// TODO(drifkin): validate the level, could be model dependent
-						// though... It will also be validated on the server once a call is
-						// made.
-						thinkValue.Value = maybeLevel
-					}
-					opts.Think = &thinkValue
-					thinkExplicitlySet = true
-					if client, err := api.ClientFromEnvironment(); err == nil {
-						ensureThinkingSupport(cmd.Context(), client, opts.Model)
-					}
-					if maybeLevel != "" {
-						fmt.Printf("Set 'think' mode to '%s'.\n", maybeLevel)
-					} else {
-						fmt.Println("Set 'think' mode.")
-					}
-				case "nothink":
-					opts.Think = &api.ThinkValue{Value: false}
-					thinkExplicitlySet = true
-					if client, err := api.ClientFromEnvironment(); err == nil {
-						ensureThinkingSupport(cmd.Context(), client, opts.Model)
-					}
-					fmt.Println("Set 'nothink' mode.")
 				case "format":
 					if len(args) < 3 || args[2] != "json" {
 						fmt.Println("Invalid or missing format. For 'json' mode use '/set format json'")
@@ -398,21 +358,18 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				case "modelfile":
 					fmt.Println(resp.Modelfile)
 				case "parameters":
-					fmt.Println("Model defined parameters:")
 					if resp.Parameters == "" {
-						fmt.Println("  No additional parameters were specified for this model.")
+						fmt.Println("No parameters were specified for this model.")
 					} else {
-						for _, l := range strings.Split(resp.Parameters, "\n") {
-							fmt.Printf("  %s\n", l)
+						if len(opts.Options) > 0 {
+							fmt.Println("User defined parameters:")
+							for k, v := range opts.Options {
+								fmt.Printf("%-*s %v\n", 30, k, v)
+							}
+							fmt.Println()
 						}
-					}
-					fmt.Println()
-					if len(opts.Options) > 0 {
-						fmt.Println("User defined parameters:")
-						for k, v := range opts.Options {
-							fmt.Printf("  %-*s %v\n", 30, k, v)
-						}
-						fmt.Println()
+						fmt.Println("Model defined parameters:")
+						fmt.Println(resp.Parameters)
 					}
 				case "system":
 					switch {
@@ -491,12 +448,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {

 			assistant, err := chat(cmd, opts)
 			if err != nil {
-				if strings.Contains(err.Error(), "does not support thinking") ||
-					strings.Contains(err.Error(), "invalid think value") {
-					fmt.Printf("error: %v\n", err)
-					sb.Reset()
-					continue
-				}
 				return err
 			}
 			if assistant != nil {
@@ -560,7 +511,7 @@ func extractFileNames(input string) []string {
 	// Regex to match file paths starting with optional drive letter, / ./ \ or .\ and include escaped or unescaped spaces (\ or %20)
 	// and followed by more characters and a file extension
 	// This will capture non filename strings, but we'll check for file existence to remove mismatches
-	regexPattern := `(?:[a-zA-Z]:)?(?:\./|/|\\)[\S\\ ]+?\.(?i:jpg|jpeg|png|webp)\b`
+	regexPattern := `(?:[a-zA-Z]:)?(?:\./|/|\\)[\S\\ ]+?\.(?i:jpg|jpeg|png)\b`
 	re := regexp.MustCompile(regexPattern)

 	return re.FindAllString(input, -1)
@@ -580,8 +531,6 @@ func extractFileData(input string) (string, []api.ImageData, error) {
 			return "", imgs, err
 		}
 		fmt.Fprintf(os.Stderr, "Added image '%s'\n", nfp)
-		input = strings.ReplaceAll(input, "'"+nfp+"'", "")
-		input = strings.ReplaceAll(input, "'"+fp+"'", "")
 		input = strings.ReplaceAll(input, fp, "")
 		imgs = append(imgs, data)
 	}
@@ -602,7 +551,7 @@ func getImageData(filePath string) ([]byte, error) {
 	}

 	contentType := http.DetectContentType(buf)
-	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png", "image/webp"}
+	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png"}
 	if !slices.Contains(allowedTypes, contentType) {
 		return nil, fmt.Errorf("invalid image type: %s", contentType)
 	}
--- a/cmd/interactive_test.go
+++ b/cmd/interactive_test.go
@@ -1,8 +1,6 @@
 package cmd

 import (
-	"os"
-	"path/filepath"
 	"testing"

 	"github.com/stretchr/testify/assert"
@@ -12,17 +10,14 @@ func TestExtractFilenames(t *testing.T) {
 	// Unix style paths
 	input := ` some preamble 
 ./relative\ path/one.png inbetween1 ./not a valid two.jpg inbetween2 ./1.svg
-/unescaped space /three.jpeg inbetween3 /valid\ path/dir/four.png "./quoted with spaces/five.JPG
-/unescaped space /six.webp inbetween6 /valid\ path/dir/seven.WEBP`
+/unescaped space /three.jpeg inbetween3 /valid\ path/dir/four.png "./quoted with spaces/five.JPG`
 	res := extractFileNames(input)
-	assert.Len(t, res, 7)
+	assert.Len(t, res, 5)
 	assert.Contains(t, res[0], "one.png")
 	assert.Contains(t, res[1], "two.jpg")
 	assert.Contains(t, res[2], "three.jpeg")
 	assert.Contains(t, res[3], "four.png")
 	assert.Contains(t, res[4], "five.JPG")
-	assert.Contains(t, res[5], "six.webp")
-	assert.Contains(t, res[6], "seven.WEBP")
 	assert.NotContains(t, res[4], '"')
 	assert.NotContains(t, res, "inbetween1")
 	assert.NotContains(t, res, "./1.svg")
@@ -33,12 +28,10 @@ func TestExtractFilenames(t *testing.T) {
 /absolute/nospace/three.jpeg inbetween3 /absolute/with space/four.png inbetween4
 ./relative\ path/five.JPG inbetween5 "./relative with/spaces/six.png inbetween6
 d:\path with\spaces\seven.JPEG inbetween7 c:\users\jdoe\eight.png inbetween8 
- d:\program files\someplace\nine.png inbetween9 "E:\program files\someplace\ten.PNG
-c:/users/jdoe/eleven.webp inbetween11 c:/program files/someplace/twelve.WebP inbetween12
-d:\path with\spaces\thirteen.WEBP some ending
+ d:\program files\someplace\nine.png inbetween9 "E:\program files\someplace\ten.PNG some ending
 `
 	res = extractFileNames(input)
-	assert.Len(t, res, 13)
+	assert.Len(t, res, 10)
 	assert.NotContains(t, res, "inbetween2")
 	assert.Contains(t, res[0], "one.png")
 	assert.Contains(t, res[0], "c:")
@@ -56,31 +49,4 @@ d:\path with\spaces\thirteen.WEBP some ending
 	assert.Contains(t, res[8], "d:")
 	assert.Contains(t, res[9], "ten.PNG")
 	assert.Contains(t, res[9], "E:")
-	assert.Contains(t, res[10], "eleven.webp")
-	assert.Contains(t, res[10], "c:")
-	assert.Contains(t, res[11], "twelve.WebP")
-	assert.Contains(t, res[11], "c:")
-	assert.Contains(t, res[12], "thirteen.WEBP")
-	assert.Contains(t, res[12], "d:")
-}
-
-// Ensure that file paths wrapped in single quotes are removed with the quotes.
-func TestExtractFileDataRemovesQuotedFilepath(t *testing.T) {
-	dir := t.TempDir()
-	fp := filepath.Join(dir, "img.jpg")
-	data := make([]byte, 600)
-	copy(data, []byte{
-		0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10, 'J', 'F', 'I', 'F',
-		0x00, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-		0xff, 0xd9,
-	})
-	if err := os.WriteFile(fp, data, 0o600); err != nil {
-		t.Fatalf("failed to write test image: %v", err)
-	}
-
-	input := "before '" + fp + "' after"
-	cleaned, imgs, err := extractFileData(input)
-	assert.NoError(t, err)
-	assert.Len(t, imgs, 1)
-	assert.Equal(t, cleaned, "before  after")
 }
--- a/cmd/start_darwin.go
+++ b/cmd/start_darwin.go
@@ -5,7 +5,7 @@ import (
 	"errors"
 	"os"
 	"os/exec"
-	"regexp"
+	"strings"

 	"github.com/ollama/ollama/api"
 )
@@ -19,12 +19,11 @@ func startApp(ctx context.Context, client *api.Client) error {
 	if err != nil {
 		return err
 	}
-	r := regexp.MustCompile(`^.*/Ollama\s?\d*.app`)
-	m := r.FindStringSubmatch(link)
-	if len(m) != 1 {
+	if !strings.Contains(link, "Ollama.app") {
 		return errors.New("could not find ollama app")
 	}
-	if err := exec.Command("/usr/bin/open", "-j", "-a", m[0], "--args", "--fast-startup").Run(); err != nil {
+	path := strings.Split(link, "Ollama.app")
+	if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil {
 		return err
 	}
 	return waitForServer(ctx, client)
--- a/cmd/start_windows.go
+++ b/cmd/start_windows.go
@@ -4,27 +4,17 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"log/slog"
 	"os"
 	"os/exec"
-	"path"
 	"path/filepath"
 	"strings"
 	"syscall"
-	"unsafe"

 	"github.com/ollama/ollama/api"
-	"golang.org/x/sys/windows"
-)
-
-const (
-	Installer = "OllamaSetup.exe"
 )

 func startApp(ctx context.Context, client *api.Client) error {
-	if len(isProcRunning(Installer)) > 0 {
-		return fmt.Errorf("upgrade in progress...")
-	}
+	// log.Printf("XXX Attempting to find and start ollama app")
 	AppName := "ollama app.exe"
 	exe, err := os.Executable()
 	if err != nil {
@@ -45,11 +35,14 @@ func startApp(ctx context.Context, client *api.Client) error {
 			}
 		}
 	}
+	// log.Printf("XXX attempting to start app %s", appExe)

 	cmd_path := "c:\\Windows\\system32\\cmd.exe"
-	cmd := exec.Command(cmd_path, "/c", appExe, "--hide", "--fast-startup")
+	cmd := exec.Command(cmd_path, "/c", appExe)
+	// TODO - these hide flags aren't working - still pops up a command window for some reason
 	cmd.SysProcAttr = &syscall.SysProcAttr{CreationFlags: 0x08000000, HideWindow: true}

+	// TODO this didn't help either...
 	cmd.Stdin = strings.NewReader("")
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
@@ -63,50 +56,3 @@ func startApp(ctx context.Context, client *api.Client) error {
 	}
 	return waitForServer(ctx, client)
 }
-
-func isProcRunning(procName string) []uint32 {
-	pids := make([]uint32, 2048)
-	var ret uint32
-	if err := windows.EnumProcesses(pids, &ret); err != nil || ret == 0 {
-		slog.Debug("failed to check for running installers", "error", err)
-		return nil
-	}
-	if ret > uint32(len(pids)) {
-		pids = make([]uint32, ret+10)
-		if err := windows.EnumProcesses(pids, &ret); err != nil || ret == 0 {
-			slog.Debug("failed to check for running installers", "error", err)
-			return nil
-		}
-	}
-	if ret < uint32(len(pids)) {
-		pids = pids[:ret]
-	}
-	var matches []uint32
-	for _, pid := range pids {
-		if pid == 0 {
-			continue
-		}
-		hProcess, err := windows.OpenProcess(windows.PROCESS_QUERY_INFORMATION|windows.PROCESS_VM_READ, false, pid)
-		if err != nil {
-			continue
-		}
-		defer windows.CloseHandle(hProcess)
-		var module windows.Handle
-		var cbNeeded uint32
-		cb := (uint32)(unsafe.Sizeof(module))
-		if err := windows.EnumProcessModules(hProcess, &module, cb, &cbNeeded); err != nil {
-			continue
-		}
-		var sz uint32 = 1024 * 8
-		moduleName := make([]uint16, sz)
-		cb = uint32(len(moduleName)) * (uint32)(unsafe.Sizeof(uint16(0)))
-		if err := windows.GetModuleBaseName(hProcess, module, &moduleName[0], cb); err != nil && err != syscall.ERROR_INSUFFICIENT_BUFFER {
-			continue
-		}
-		exeFile := path.Base(strings.ToLower(syscall.UTF16ToString(moduleName)))
-		if strings.EqualFold(exeFile, procName) {
-			matches = append(matches, pid)
-		}
-	}
-	return matches
-}
--- a/cmd/warn_thinking_test.go
+++ b/cmd/warn_thinking_test.go
@@ -1,63 +0,0 @@
-package cmd
-
-import (
-	"encoding/json"
-	"io"
-	"net/http"
-	"net/http/httptest"
-	"os"
-	"strings"
-	"testing"
-
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/types/model"
-)
-
-// Test that a warning is printed when thinking is requested but not supported.
-func TestWarnMissingThinking(t *testing.T) {
-	cases := []struct {
-		capabilities []model.Capability
-		expectWarn   bool
-	}{
-		{capabilities: []model.Capability{model.CapabilityThinking}, expectWarn: false},
-		{capabilities: []model.Capability{}, expectWarn: true},
-	}
-
-	for _, tc := range cases {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path != "/api/show" || r.Method != http.MethodPost {
-				t.Fatalf("unexpected request to %s %s", r.URL.Path, r.Method)
-			}
-			var req api.ShowRequest
-			if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-				t.Fatalf("decode request: %v", err)
-			}
-			resp := api.ShowResponse{Capabilities: tc.capabilities}
-			if err := json.NewEncoder(w).Encode(resp); err != nil {
-				t.Fatalf("encode response: %v", err)
-			}
-		}))
-		defer srv.Close()
-
-		t.Setenv("OLLAMA_HOST", srv.URL)
-		client, err := api.ClientFromEnvironment()
-		if err != nil {
-			t.Fatal(err)
-		}
-		oldStderr := os.Stderr
-		r, w, _ := os.Pipe()
-		os.Stderr = w
-		ensureThinkingSupport(t.Context(), client, "m")
-		w.Close()
-		os.Stderr = oldStderr
-		out, _ := io.ReadAll(r)
-
-		warned := strings.Contains(string(out), "warning:")
-		if tc.expectWarn && !warned {
-			t.Errorf("expected warning, got none")
-		}
-		if !tc.expectWarn && warned {
-			t.Errorf("did not expect warning, got: %s", string(out))
-		}
-	}
-}
--- a/convert/convert.go
+++ b/convert/convert.go
@@ -1,7 +1,6 @@
 package convert

 import (
-	"cmp"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -15,12 +14,13 @@ import (
 )

 type ModelParameters struct {
-	Architectures []string `json:"architectures"`
-	VocabSize     uint32   `json:"vocab_size"`
+	Architectures []string       `json:"architectures"`
+	VocabSize     uint32         `json:"vocab_size"`
+	TextModel     TextParameters `json:"text_config"`
+}

-	TextModel struct {
-		VocabSize uint32 `json:"vocab_size"`
-	} `json:"text_config"`
+type TextParameters struct {
+	VocabSize uint32 `json:"vocab_size"`
 }

 type AdapterParameters struct {
@@ -53,11 +53,8 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	}

 	for _, sv := range t.SpecialVocabulary {
-		kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken
 		kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID)
-		if len(sv.IDs) > 0 {
-			kv[fmt.Sprintf("tokenizer.ggml.%s_token_ids", sv.Key())] = sv.IDs
-		}
+		kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken
 	}

 	return kv
@@ -176,8 +173,6 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	switch p.Architectures[0] {
 	case "LlamaForCausalLM":
 		conv = &llamaModel{}
-	case "MllamaForConditionalGeneration":
-		conv = &mllamaModel{}
 	case "Llama4ForConditionalGeneration":
 		conv = &llama4Model{}
 	case "Mistral3ForConditionalGeneration":
@@ -190,20 +185,14 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &gemma2Model{}
 	case "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration":
 		conv = &gemma3Model{Architecture: p.Architectures[0]}
-	case "Gemma3nForConditionalGeneration":
-		conv = &gemma3nModel{}
 	case "Phi3ForCausalLM":
 		conv = &phi3Model{}
 	case "Qwen2ForCausalLM":
 		conv = &qwen2Model{}
-	case "Qwen2_5_VLForConditionalGeneration":
-		conv = &qwen25VLModel{}
 	case "BertModel":
 		conv = &bertModel{}
 	case "CohereForCausalLM":
 		conv = &commandrModel{}
-	case "GptOssForCausalLM":
-		conv = &gptossModel{}
 	default:
 		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}
@@ -223,22 +212,24 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		return err
 	}

-	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
+	vocabSize := int(p.VocabSize)
+	if vocabSize == 0 {
+		tVocabSize := int(p.TextModel.VocabSize)
+		vocabSize = tVocabSize
+	}

 	switch {
 	case vocabSize == 0:
-		slog.Debug("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
+		slog.Warn("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
 	case vocabSize > len(t.Vocabulary.Tokens):
-		slog.Debug("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
+		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
 		for i := range vocabSize - len(t.Vocabulary.Tokens) {
 			t.Vocabulary.Tokens = append(t.Vocabulary.Tokens, fmt.Sprintf("[PAD%d]", i))
 			t.Vocabulary.Scores = append(t.Vocabulary.Scores, -1)
 			t.Vocabulary.Types = append(t.Vocabulary.Types, tokenTypeUserDefined)
 		}
 	case vocabSize < len(t.Vocabulary.Tokens):
-		slog.Debug("vocabulary is larger than expected", "want", vocabSize, "got", len(t.Vocabulary.Tokens))
-		p.VocabSize = uint32(len(t.Vocabulary.Tokens))
-		p.TextModel.VocabSize = uint32(len(t.Vocabulary.Tokens))
+		return fmt.Errorf("vocabulary is larger than expected '%d' instead of '%d'", len(t.Vocabulary.Tokens), vocabSize)
 	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
--- a/convert/convert_gemma3n.go
+++ b/convert/convert_gemma3n.go
@@ -1,165 +0,0 @@
-package convert
-
-import (
-	"slices"
-	"strings"
-
-	"github.com/ollama/ollama/fs/ggml"
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-	"gonum.org/v1/gonum/stat/distuv"
-)
-
-type gemma3nModel struct {
-	ModelParameters
-
-	TextModel struct {
-		ActivationSparsityPattern []float32 `json:"activation_sparsity_pattern"`
-		AltupActiveIdx            uint32    `json:"altup_active_idx"`
-		AltupCoefClip             float32   `json:"altup_coef_clip"`
-		AltupCorrectScale         bool      `json:"altup_correct_scale"`
-		AltupLRMultiplier         float32   `json:"altup_lr_multiplier"`
-		AltupNumInputs            uint32    `json:"altup_num_inputs"`
-		HeadDim                   uint32    `json:"head_dim"`
-		HiddenSize                uint32    `json:"hidden_size"`
-		HiddenSizePerLayerInput   uint32    `json:"hidden_size_per_layer_input"`
-		IntermediateSize          uint32    `json:"intermediate_size"`
-		MaxPositionEmbeddings     uint32    `json:"max_position_embeddings"`
-		NumAttentionHeads         uint32    `json:"num_attention_heads"`
-		NumHiddenLayers           uint32    `json:"num_hidden_layers"`
-		NumKeyValueHeads          uint32    `json:"num_key_value_heads"`
-		NumKVSharedLayers         uint32    `json:"num_kv_shared_layers"`
-		RMSNormEPS                float32   `json:"rms_norm_eps"`
-		RopeLocalBaseFreq         float32   `json:"rope_local_base_freq"`
-		RopeTheta                 float32   `json:"rope_theta"`
-		SlidingWindow             uint32    `json:"sliding_window"`
-		LayerTypes                []string  `json:"layer_types"`
-	} `json:"text_config"`
-	VisionModel struct{} `json:"vision_config"`
-}
-
-func (m *gemma3nModel) KV(t *Tokenizer) ggml.KV {
-	kv := m.ModelParameters.KV(t)
-	kv["general.architecture"] = "gemma3n"
-	kv["gemma3n.activation_sparsity_scale"] = slices.Collect(func(yield func(float32) bool) {
-		norm := distuv.Normal{Mu: 0, Sigma: 1}
-		for _, v := range m.TextModel.ActivationSparsityPattern {
-			if !yield(float32(norm.Quantile(float64(v)))) {
-				break
-			}
-		}
-	})
-	kv["gemma3n.altup.active_idx"] = m.TextModel.AltupActiveIdx
-	kv["gemma3n.altup.correct_scale"] = m.TextModel.AltupCorrectScale
-	kv["gemma3n.altup.lr_multiplier"] = m.TextModel.AltupLRMultiplier
-	kv["gemma3n.altup.num_inputs"] = m.TextModel.AltupNumInputs
-	kv["gemma3n.attention.head_count_kv"] = m.TextModel.NumKeyValueHeads
-	kv["gemma3n.attention.head_count"] = m.TextModel.NumAttentionHeads
-	kv["gemma3n.attention.layer_norm_rms_epsilon"] = m.TextModel.RMSNormEPS
-	kv["gemma3n.attention.sliding_window"] = m.TextModel.SlidingWindow
-	kv["gemma3n.attention.sliding_window_pattern"] = slices.Collect(func(yield func(bool) bool) {
-		for _, t := range m.TextModel.LayerTypes {
-			if !yield(t == "sliding_attention") {
-				break
-			}
-		}
-	})
-	kv["gemma3n.attention.shared_kv_layers"] = m.TextModel.NumKVSharedLayers
-	kv["gemma3n.block_count"] = m.TextModel.NumHiddenLayers
-	kv["gemma3n.context_length"] = m.TextModel.MaxPositionEmbeddings
-	kv["gemma3n.embedding_length_per_layer_input"] = m.TextModel.HiddenSizePerLayerInput
-	kv["gemma3n.embedding_length"] = m.TextModel.HiddenSize
-	kv["gemma3n.feed_forward_length"] = m.TextModel.IntermediateSize
-	kv["gemma3n.head_dim"] = m.TextModel.HeadDim
-	kv["gemma3n.rope.freq_base_local"] = m.TextModel.RopeLocalBaseFreq
-	kv["gemma3n.rope.freq_base"] = m.TextModel.RopeTheta
-	return kv
-}
-
-func (m *gemma3nModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	out, ts := mergeTensors(ts,
-		merge{"altup_proj.*.weight", "altup_proj.weight"},
-		merge{"altup_unembd_proj.*.weight", "altup_unembd_proj.weight"},
-	)
-
-	for _, t := range ts {
-		switch {
-		case strings.Contains(t.Name(), "audio_tower"),
-			strings.Contains(t.Name(), "embed_audio"),
-			strings.Contains(t.Name(), "vision_tower"),
-			strings.Contains(t.Name(), "embed_vision"):
-			// TODO: handle audio and vision towers
-			continue
-		case strings.Contains(t.Name(), "altup_predict_coef"),
-			strings.Contains(t.Name(), "altup_correct_coef"):
-			if m.TextModel.AltupCoefClip > 0 {
-				t.SetRepacker(func(name string, data []float32, shape []uint64) (_ []float32, err error) {
-					dims := make([]int, len(shape))
-					for i := range shape {
-						dims[i] = int(shape[i])
-					}
-
-					var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-
-					t, err = tensor.Clamp(t, -m.TextModel.AltupCoefClip, m.TextModel.AltupCoefClip)
-					if err != nil {
-						return nil, err
-					}
-
-					if err := t.Reshape(t.Shape().TotalSize()); err != nil {
-						return nil, err
-					}
-
-					return native.VectorF32(t.(*tensor.Dense))
-				})
-			}
-		}
-
-		out = append(out, &ggml.Tensor{
-			Name:     t.Name(),
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (m *gemma3nModel) Replacements() []string {
-	return []string{
-		"model.language_model.embed_tokens_per_layer", "per_layer_token_embd",
-		"model.language_model.embed_tokens", "token_embd",
-		"model.language_model.per_layer_model_projection", "per_layer_model_proj",
-		"model.language_model.per_layer_projection_norm", "per_layer_proj_norm", "model.language_model.altup_projections", "altup_proj",
-		"model.language_model.altup_unembed_projections", "altup_unembd_proj",
-		"model.language_model.norm", "output_norm",
-		"model.language_model.layers", "blk",
-
-		"input_layernorm", "attn_norm",
-		"self_attn.q_proj", "attn_q",
-		"self_attn.q_norm", "attn_q_norm",
-		"self_attn.k_proj", "attn_k",
-		"self_attn.k_norm", "attn_k_norm",
-		"self_attn.v_proj", "attn_v",
-		"self_attn.o_proj", "attn_output",
-		"post_attention_layernorm", "post_attention_norm",
-		"pre_feedforward_layernorm", "ffn_norm",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.up_proj", "ffn_up",
-		"mlp.down_proj", "ffn_down",
-		"post_feedforward_layernorm", "post_ffw_norm",
-		"per_layer_input_gate", "inp_gate",
-		"per_layer_projection", "proj",
-		"post_per_layer_input_norm", "post_norm",
-		"altup.", "altup_",
-		"modality_router", "router",
-		"prediction_coefs", "predict_coef",
-		"correction_coefs", "correct_coef",
-		"correct_output_scale", "correct_scale.weight",
-		"laurel.", "laurel_",
-		"linear_left", "l",
-		"linear_right", "r",
-		"post_laurel_norm", "post_norm",
-	}
-}
--- a/convert/convert_gptoss.go
+++ b/convert/convert_gptoss.go
@@ -1,178 +0,0 @@
-package convert
-
-import (
-	"bytes"
-	"cmp"
-	"encoding/binary"
-	"io"
-	"slices"
-	"strings"
-
-	"github.com/ollama/ollama/fs/ggml"
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-)
-
-type gptossModel struct {
-	ModelParameters
-	HiddenLayers         uint32  `json:"num_hidden_layers"`
-	HiddenSize           uint32  `json:"hidden_size"`
-	IntermediateSize     uint32  `json:"intermediate_size"`
-	AttentionHeads       uint32  `json:"num_attention_heads"`
-	KeyValueHeads        uint32  `json:"num_key_value_heads"`
-	HeadDim              uint32  `json:"head_dim"`
-	Experts              uint32  `json:"num_experts"`
-	ExpertsPerToken      uint32  `json:"experts_per_token"`
-	RMSNormEpsilon       float32 `json:"rms_norm_eps"`
-	InitialContextLength uint32  `json:"initial_context_length"`
-	RopeTheta            float32 `json:"rope_theta"`
-	RopeScalingFactor    float32 `json:"rope_scaling_factor"`
-	SlidingWindow        uint32  `json:"sliding_window"`
-}
-
-var _ ModelConverter = (*gptossModel)(nil)
-
-func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
-	kv := m.ModelParameters.KV(t)
-	kv["general.architecture"] = "gptoss"
-	kv["general.file_type"] = uint32(4)
-	kv["gptoss.context_length"] = uint32(m.RopeScalingFactor * float32(m.InitialContextLength))
-	kv["gptoss.block_count"] = m.HiddenLayers
-	kv["gptoss.embedding_length"] = m.HiddenSize
-	kv["gptoss.feed_forward_length"] = m.IntermediateSize
-	kv["gptoss.expert_count"] = m.Experts
-	kv["gptoss.expert_used_count"] = m.ExpertsPerToken
-	kv["gptoss.attention.head_count"] = m.AttentionHeads
-	kv["gptoss.attention.head_count_kv"] = m.KeyValueHeads
-	kv["gptoss.attention.key_length"] = m.HeadDim
-	kv["gptoss.attention.value_length"] = m.HeadDim
-	kv["gptoss.attention.layer_norm_rms_epsilon"] = cmp.Or(m.RMSNormEpsilon, 1e-5)
-	kv["gptoss.attention.sliding_window"] = m.SlidingWindow
-	kv["gptoss.rope.freq_base"] = m.RopeTheta
-	kv["gptoss.rope.scaling.factor"] = m.RopeScalingFactor
-	kv["gptoss.rope.scaling.original_context_length"] = m.InitialContextLength
-	kv["tokenizer.ggml.bos_token_id"] = uint32(199998) // <|startoftext|>
-	kv["tokenizer.ggml.add_bos_token"] = false
-	kv["tokenizer.ggml.eos_token_id"] = uint32(199999) // <|endoftext|>
-	kv["tokenizer.ggml.eos_token_ids"] = []int32{
-		199999, /* <|endoftext|> */
-		200002, /* <|return|> */
-		200012, /* <|call|> */
-	}
-	kv["tokenizer.ggml.add_eos_token"] = false
-	return kv
-}
-
-func (m *gptossModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	var out []*ggml.Tensor
-	mxfp4s := make(map[string]*mxfp4)
-	for _, t := range ts {
-		if strings.HasSuffix(t.Name(), ".blocks") || strings.HasSuffix(t.Name(), ".scales") {
-			dot := strings.LastIndex(t.Name(), ".")
-			name, suffix := t.Name()[:dot], t.Name()[dot+1:]
-			if _, ok := mxfp4s[name]; !ok {
-				mxfp4s[name] = &mxfp4{}
-			}
-
-			switch suffix {
-			case "blocks":
-				mxfp4s[name].blocks = t
-			case "scales":
-				mxfp4s[name].scales = t
-			}
-		} else {
-			out = append(out, &ggml.Tensor{
-				Name:     t.Name(),
-				Kind:     t.Kind(),
-				Shape:    t.Shape(),
-				WriterTo: t,
-			})
-		}
-	}
-
-	for name, mxfp4 := range mxfp4s {
-		dims := mxfp4.blocks.Shape()
-		out = append(out, &ggml.Tensor{
-			Name:     name,
-			Kind:     uint32(ggml.TensorTypeMXFP4),
-			Shape:    []uint64{dims[0], dims[1], dims[2] * dims[3] * 2},
-			WriterTo: mxfp4,
-		})
-	}
-
-	return out
-}
-
-func (m *gptossModel) Replacements() []string {
-	return []string{
-		// noop replacements so other replacements will not be applied
-		".blocks", ".blocks",
-		".scales", ".scales",
-		// real replacements
-		"block", "blk",
-		"attn.norm", "attn_norm",
-		"attn.qkv", "attn_qkv",
-		"attn.sinks", "attn_sinks",
-		"attn.out", "attn_out",
-		"mlp.norm", "ffn_norm",
-		"mlp.gate", "ffn_gate_inp",
-		"mlp.mlp1_", "ffn_gate_up_exps.",
-		"mlp.mlp2_", "ffn_down_exps.",
-		"embedding", "token_embd",
-		"norm", "output_norm",
-		"unembedding", "output",
-		"scale", "weight",
-	}
-}
-
-type mxfp4 struct {
-	blocks, scales Tensor
-}
-
-func (m *mxfp4) WriteTo(w io.Writer) (int64, error) {
-	var b bytes.Buffer
-	if _, err := m.blocks.WriteTo(&b); err != nil {
-		return 0, err
-	}
-
-	blocksDims := make([]int, len(m.blocks.Shape()))
-	for i, d := range m.blocks.Shape() {
-		blocksDims[i] = int(d)
-	}
-
-	var blocks tensor.Tensor = tensor.New(tensor.WithShape(blocksDims...), tensor.WithBacking(b.Bytes()))
-
-	var s bytes.Buffer
-	if _, err := m.scales.WriteTo(&s); err != nil {
-		return 0, err
-	}
-
-	scalesDims := slices.Repeat([]int{1}, len(m.blocks.Shape()))
-	for i, d := range m.scales.Shape() {
-		scalesDims[i] = int(d)
-	}
-
-	var scales tensor.Tensor = tensor.New(tensor.WithShape(scalesDims...), tensor.WithBacking(s.Bytes()))
-
-	out, err := tensor.Concat(3, scales, blocks)
-	if err != nil {
-		return 0, err
-	}
-
-	out = tensor.Materialize(out)
-
-	if err := out.Reshape(out.Shape().TotalSize()); err != nil {
-		return 0, err
-	}
-
-	u8s, err := native.VectorU8(out.(*tensor.Dense))
-	if err != nil {
-		return 0, err
-	}
-
-	if err := binary.Write(w, binary.LittleEndian, u8s); err != nil {
-		return 0, err
-	}
-
-	return 0, nil
-}
--- a/convert/convert_llama.go
+++ b/convert/convert_llama.go
@@ -139,8 +139,7 @@ func (p *llamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	}

 	for _, t := range ts {
-		if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") ||
-			strings.HasSuffix(t.Name(), "attn_q_proj.weight") || strings.HasSuffix(t.Name(), "attn_k_proj.weight") {
+		if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") {
 			if !p.skipRepack {
 				t.SetRepacker(p.repack)
 			}
@@ -182,9 +181,9 @@ func (p *llamaModel) repack(name string, data []float32, shape []uint64) ([]floa
 	}

 	var heads uint32
-	if strings.HasSuffix(name, "attn_q.weight") || strings.HasSuffix(name, "attn_q_proj.weight") {
+	if strings.HasSuffix(name, "attn_q.weight") {
 		heads = p.NumAttentionHeads
-	} else if strings.HasSuffix(name, "attn_k.weight") || strings.HasSuffix(name, "attn_k_proj.weight") {
+	} else if strings.HasSuffix(name, "attn_k.weight") {
 		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
 	} else {
 		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
--- a/convert/convert_mixtral.go
+++ b/convert/convert_mixtral.go
@@ -2,6 +2,9 @@ package convert

 import (
 	"fmt"
+	"io"
+	"slices"
+	"strings"

 	"github.com/ollama/ollama/fs/ggml"
 )
@@ -27,38 +30,65 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 }

 func (p *mixtralModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	merges := make([]merge, 0, p.NumHiddenLayers*6)
-	for i := range p.NumHiddenLayers {
-		merges = append(merges, merge{
-			fmt.Sprintf("blk.%d.*.w1.weight", i),
-			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w1.bias", i),
-			fmt.Sprintf("blk.%d.ffn_gate_exps.bias", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w2.weight", i),
-			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w2.bias", i),
-			fmt.Sprintf("blk.%d.ffn_up_exps.bias", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w3.weight", i),
-			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w3.bias", i),
-			fmt.Sprintf("blk.%d.ffn_down_exps.bias", i),
+	oldnew := []string{
+		"model.layers", "blk",
+		"w1", "ffn_gate_exps",
+		"w2", "ffn_down_exps",
+		"w3", "ffn_up_exps",
+	}
+
+	for i := range p.NumLocalExperts {
+		oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".")
+	}
+
+	// group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor
+	namer := strings.NewReplacer(oldnew...)
+	experts := make(map[string]experts)
+
+	// merge experts into a single tensor while removing them from ts
+	ts = slices.DeleteFunc(ts, func(t Tensor) bool {
+		if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") {
+			return false
+		}
+
+		name := namer.Replace(t.Name())
+		experts[name] = append(experts[name], t)
+		return true
+	})
+
+	var out []*ggml.Tensor
+	for n, e := range experts {
+		// TODO(mxyng): sanity check experts
+		out = append(out, &ggml.Tensor{
+			Name:     n,
+			Kind:     e[0].Kind(),
+			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),
+			WriterTo: e,
 		})
 	}

-	out, ts := mergeTensors(ts, merges...)
 	return append(out, p.llamaModel.Tensors(ts)...)
 }

 func (p *mixtralModel) Replacements() []string {
 	return append(
 		p.llamaModel.Replacements(),
-		"model.layers", "blk",
 		"block_sparse_moe.gate", "ffn_gate_inp",
-		"block_sparse_moe.experts.", ".",
 	)
 }
+
+type experts []Tensor
+
+func (e experts) WriteTo(w io.Writer) (int64, error) {
+	// TODO(mxyng): experts _should_ be numerically sorted by expert but this should check
+	for _, t := range e {
+		// the canonical merged experts tensor stacks all experts along a new, 0 axis,
+		// e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers
+		// this accomplishes the same thing by writing each expert tensor in sequence
+		if _, err := t.WriteTo(w); err != nil {
+			return 0, err
+		}
+	}
+
+	return 0, nil
+}
--- a/convert/convert_mllama.go
+++ b/convert/convert_mllama.go
@@ -1,179 +0,0 @@
-package convert
-
-import (
-	"strings"
-
-	"github.com/ollama/ollama/fs/ggml"
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-)
-
-type mllamaModel struct {
-	ModelParameters
-	TextModel struct {
-		llamaModel
-
-		CrossAttentionLayers []int32 `json:"cross_attention_layers"`
-	} `json:"text_config"`
-	VisionModel struct {
-		NumHiddenLayers           uint32  `json:"num_hidden_layers"`
-		NumGlobalLayers           uint32  `json:"num_global_layers"`
-		IntermediateLayersIndices []int32 `json:"intermediate_layers_indices"`
-
-		HiddenSize       uint32 `json:"hidden_size"`
-		IntermediateSize uint32 `json:"intermediate_size"`
-
-		AttentionHeads uint32 `json:"attention_heads"`
-
-		ImageSize   uint32  `json:"image_size"`
-		PatchSize   uint32  `json:"patch_size"`
-		NumChannels uint32  `json:"num_channels"`
-		MaxNumTiles uint32  `json:"max_num_tiles"`
-		NormEpsilon float32 `json:"norm_eps"`
-		RopeTheta   float32 `json:"rope.freq_base"`
-	} `json:"vision_config"`
-}
-
-func (m *mllamaModel) KV(t *Tokenizer) ggml.KV {
-	kv := m.ModelParameters.KV(t)
-	kv["general.architecture"] = "mllama"
-
-	for k, v := range m.TextModel.KV(t) {
-		if strings.HasPrefix(k, "llama.") {
-			kv[strings.ReplaceAll(k, "llama.", "mllama.")] = v
-		}
-	}
-
-	kv["mllama.attention.cross_attention_layers"] = m.TextModel.CrossAttentionLayers
-
-	kv["mllama.vision.block_count"] = m.VisionModel.NumHiddenLayers
-	kv["mllama.vision.global.block_count"] = m.VisionModel.NumGlobalLayers
-	kv["mllama.vision.intermediate_layers_indices"] = m.VisionModel.IntermediateLayersIndices
-
-	kv["mllama.vision.embedding_length"] = m.VisionModel.HiddenSize
-	kv["mllama.vision.feed_forward_length"] = m.VisionModel.IntermediateSize
-
-	kv["mllama.vision.attention.head_count"] = m.VisionModel.AttentionHeads
-	kv["mllama.vision.attention.layer_norm_epsilon"] = m.VisionModel.NormEpsilon
-
-	kv["mllama.vision.image_size"] = m.VisionModel.ImageSize
-	kv["mllama.vision.patch_size"] = m.VisionModel.PatchSize
-	kv["mllama.vision.max_num_tiles"] = m.VisionModel.MaxNumTiles
-	kv["mllama.vision.num_channels"] = m.VisionModel.NumChannels
-
-	return kv
-}
-
-func (m *mllamaModel) Replacements() []string {
-	return append(
-		m.TextModel.Replacements(),
-		"language_model.", "",
-		"gate_attn", "attn_gate",
-		"gate_ffn", "ffn_gate",
-		"cross_attn.", "cross_attn_",
-		"vision_model", "v",
-		"class_embedding", "class_embd",
-		"patch_embedding", "patch_embd",
-		"gated_positional_embedding.tile_embedding", "tile_position_embd",
-		"gated_positional_embedding.embedding", "position_embd.weight",
-		"gated_positional_embedding", "position_embd",
-		"embedding.weight", "weight",
-		"pre_tile_positional_embedding", "pre_tile_position_embd",
-		"post_tile_positional_embedding", "post_tile_position_embd",
-		"layernorm_pre", "pre_ln",
-		"layernorm_post", "post_ln",
-		"global_transformer.layers", "global.blk",
-		"transformer.layers", "blk",
-		"mlp.fc1", "ffn_up",
-		"mlp.fc2", "ffn_down",
-		"multi_modal_projector", "mm.0",
-	)
-}
-
-func (m *mllamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	var out []*ggml.Tensor
-	var text []Tensor
-	for _, t := range ts {
-		if !strings.HasPrefix(t.Name(), "v.") && !strings.HasPrefix(t.Name(), "mm.") {
-			text = append(text, t)
-		} else if t.Name() == "v.position_embd.gate" {
-			for _, name := range []string{"v.position_embd.gate", "v.tile_position_embd.gate"} {
-				tt := t.Clone()
-				tt.SetRepacker(m.repack(name))
-				out = append(out, &ggml.Tensor{
-					Name:     name,
-					Kind:     t.Kind(),
-					Shape:    t.Shape(),
-					WriterTo: tt,
-				})
-			}
-		} else {
-			if t.Name() == "v.pre_tile_position_embd.gate" || t.Name() == "v.post_tile_position_embd.gate" {
-				t.SetRepacker(m.repack(t.Name()))
-			} else if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") {
-				t.SetRepacker(m.repack(t.Name()))
-			} else if strings.HasSuffix(t.Name(), "attn_gate") || strings.HasSuffix(t.Name(), "ffn_gate") {
-				t.SetRepacker(m.repack(t.Name()))
-			}
-
-			out = append(out, &ggml.Tensor{
-				Name:     t.Name(),
-				Kind:     t.Kind(),
-				Shape:    t.Shape(),
-				WriterTo: t,
-			})
-		}
-	}
-
-	return append(out, m.TextModel.Tensors(text)...)
-}
-
-func (m *mllamaModel) repack(name string) Repacker {
-	return func(_ string, data []float32, shape []uint64) (_ []float32, err error) {
-		dims := make([]int, len(shape))
-		for i, dim := range shape {
-			dims[i] = int(dim)
-		}
-
-		var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-
-		if strings.HasSuffix(name, "attn_q.weight") || strings.HasSuffix(name, "attn_k.weight") {
-			heads := m.VisionModel.AttentionHeads
-			if err := t.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
-				return nil, err
-			}
-
-			if err := t.T(0, 2, 1, 3); err != nil {
-				return nil, err
-			}
-
-			if err := t.Reshape(dims...); err != nil {
-				return nil, err
-			}
-
-			if err := t.Transpose(); err != nil {
-				return nil, err
-			}
-		} else {
-			t, err = tensor.Tanh(t)
-			if err != nil {
-				return nil, err
-			}
-
-			if name == "v.position_embd.gate" {
-				t, err = tensor.Sub(float32(1), t)
-				if err != nil {
-					return nil, err
-				}
-			}
-		}
-
-		t = tensor.Materialize(t)
-		// flatten tensor so it can be return as a vector
-		if err := t.Reshape(t.Shape().TotalSize()); err != nil {
-			return nil, err
-		}
-
-		return native.VectorF32(t.(*tensor.Dense))
-	}
-}
--- a/convert/convert_qwen2.go
+++ b/convert/convert_qwen2.go
@@ -15,7 +15,6 @@ type qwen2Model struct {
 		Type                          string     `json:"type"`
 		Factor                        ropeFactor `json:"factor"`
 		OriginalMaxPositionEmbeddings uint32     `json:"original_max_position_embeddings"`
-		MropeSection                  []int32    `json:"mrope_section"`
 	} `json:"rope_scaling"`
 	RMSNormEPS float32 `json:"rms_norm_eps"`
 }
@@ -40,8 +39,6 @@ func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
 	case "yarn":
 		kv["qwen2.rope.scaling.type"] = q.RopeScaling.Type
 		kv["qwen2.rope.scaling.factor"] = q.RopeScaling.Factor
-	case "mrope", "default":
-		kv["qwen2.rope.mrope_section"] = q.RopeScaling.MropeSection
 	default:
 		panic("unknown rope scaling type")
 	}
--- a/convert/convert_qwen25vl.go
+++ b/convert/convert_qwen25vl.go
@@ -1,102 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"slices"
-	"strings"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type qwen25VLModel struct {
-	qwen2Model
-
-	VisionModel struct {
-		Depth               uint32  `json:"depth"`
-		HiddenSize          uint32  `json:"hidden_size"`
-		NumHeads            uint32  `json:"num_heads"`
-		InChannels          uint32  `json:"in_chans"`
-		PatchSize           uint32  `json:"patch_size"`
-		SpatialMergeSize    uint32  `json:"spatial_merge_size"`
-		SpatialPatchSize    uint32  `json:"spatial_patch_size"`
-		WindowSize          uint32  `json:"window_size"`
-		RMSNormEps          float32 `json:"layer_norm_epsilon"`
-		RopeTheta           float32 `json:"rope_theta"`
-		FullAttentionBlocks []int32 `json:"fullatt_block_indexes"`
-		TemporalPatchSize   uint32  `json:"temporal_patch_size"`
-	} `json:"vision_config"`
-}
-
-var _ ModelConverter = (*qwen25VLModel)(nil)
-
-func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
-	kv := q.ModelParameters.KV(t)
-	kv["general.architecture"] = "qwen25vl"
-
-	for k, v := range q.qwen2Model.KV(t) {
-		if strings.HasPrefix(k, "qwen2.") {
-			kv[strings.Replace(k, "qwen2.", "qwen25vl.", 1)] = v
-		}
-	}
-
-	if q.VisionModel.FullAttentionBlocks == nil {
-		kv["qwen25vl.vision.fullatt_block_indexes"] = []int32{7, 15, 23, 31}
-	}
-
-	kv["qwen25vl.vision.block_count"] = cmp.Or(q.VisionModel.Depth, 32)
-	kv["qwen25vl.vision.embedding_length"] = q.VisionModel.HiddenSize
-	kv["qwen25vl.vision.attention.head_count"] = cmp.Or(q.VisionModel.NumHeads, 16)
-	kv["qwen25vl.vision.num_channels"] = q.VisionModel.InChannels
-	kv["qwen25vl.vision.patch_size"] = cmp.Or(q.VisionModel.PatchSize, 14)
-	kv["qwen25vl.vision.spatial_merge_size"] = cmp.Or(q.VisionModel.SpatialMergeSize, 2)
-	kv["qwen25vl.vision.spatial_patch_size"] = q.VisionModel.SpatialPatchSize
-	kv["qwen25vl.vision.window_size"] = cmp.Or(q.VisionModel.WindowSize, 112)
-	kv["qwen25vl.vision.attention.layer_norm_epsilon"] = cmp.Or(q.VisionModel.RMSNormEps, 1e-6)
-	kv["qwen25vl.vision.rope.freq_base"] = cmp.Or(q.VisionModel.RopeTheta, 1e4)
-	kv["qwen25vl.vision.fullatt_block_indexes"] = q.VisionModel.FullAttentionBlocks
-	kv["qwen25vl.vision.temporal_patch_size"] = cmp.Or(q.VisionModel.TemporalPatchSize, 2)
-
-	return kv
-}
-
-func (q *qwen25VLModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	var out []*ggml.Tensor
-
-	for _, t := range ts {
-		if strings.Contains(t.Name(), "patch_embed.proj") {
-			for t := range splitDim(t, 2,
-				split{Replacer: strings.NewReplacer("patch_embed.proj", "patch_embd_0")},
-				split{Replacer: strings.NewReplacer("patch_embed.proj", "patch_embd_1")},
-			) {
-				t.Shape = slices.DeleteFunc(t.Shape, func(i uint64) bool { return i == 1 })
-				out = append(out, t)
-			}
-		} else if strings.Contains(t.Name(), "attn.qkv") {
-			out = append(out, slices.Collect(splitDim(t, 0,
-				split{Replacer: strings.NewReplacer("attn.qkv", "attn_q")},
-				split{Replacer: strings.NewReplacer("attn.qkv", "attn_k")},
-				split{Replacer: strings.NewReplacer("attn.qkv", "attn_v")},
-			))...)
-		} else {
-			out = append(out, &ggml.Tensor{
-				Name:     t.Name(),
-				Kind:     t.Kind(),
-				Shape:    t.Shape(),
-				WriterTo: t,
-			})
-		}
-	}
-
-	return out
-}
-
-func (p *qwen25VLModel) Replacements() []string {
-	return append(
-		p.qwen2Model.Replacements(),
-		"visual", "v",
-		"blocks", "blk",
-		"attn.proj", "attn_out",
-		"norm1", "ln1",
-		"norm2", "ln2",
-	)
-}
--- a/convert/convert_test.go
+++ b/convert/convert_test.go
@@ -11,13 +11,14 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
-	"maps"
 	"os"
 	"path/filepath"
 	"slices"
 	"strings"
 	"testing"

+	"golang.org/x/exp/maps"
+
 	"github.com/ollama/ollama/fs/ggml"
 )

@@ -46,7 +47,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	}
 	t.Cleanup(func() { r.Close() })

-	m, err := ggml.Decode(r, -1)
+	m, _, err := ggml.Decode(r, -1)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -136,7 +137,9 @@ func TestConvertModel(t *testing.T) {
 				t.Fatal(err)
 			}

-			for _, k := range slices.Sorted(maps.Keys(expect)) {
+			keys := maps.Keys(expect)
+			slices.Sort(keys)
+			for _, k := range keys {
 				if v, ok := actual[k]; !ok {
 					t.Errorf("missing %s", k)
 				} else if v != expect[k] {
@@ -329,7 +332,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()

-			m, err := ggml.Decode(r, -1)
+			m, _, err := ggml.Decode(r, -1)
 			if err != nil {
 				t.Fatal(err)
 			}
@@ -340,7 +343,9 @@ func TestConvertAdapter(t *testing.T) {

 			actual := generateResultsJSON(t, r, m.KV(), m.Tensors())

-			for _, k := range slices.Sorted(maps.Keys(c.Expected)) {
+			keys := maps.Keys(c.Expected)
+			slices.Sort(keys)
+			for _, k := range keys {
 				if v, ok := actual[k]; !ok {
 					t.Errorf("missing %s", k)
 				} else if v != c.Expected[k] {
--- a/convert/fs.go
+++ b/convert/fs.go
@@ -0,0 +1,58 @@
+package convert
+
+import (
+	"archive/zip"
+	"errors"
+	"io"
+	"io/fs"
+	"os"
+	"path/filepath"
+)
+
+type ZipReader struct {
+	r *zip.Reader
+	p string
+
+	// limit is the maximum size of a file that can be read directly
+	// from the zip archive. Files larger than this size will be extracted
+	limit int64
+}
+
+func NewZipReader(r *zip.Reader, p string, limit int64) fs.FS {
+	return &ZipReader{r, p, limit}
+}
+
+func (z *ZipReader) Open(name string) (fs.File, error) {
+	r, err := z.r.Open(name)
+	if err != nil {
+		return nil, err
+	}
+	defer r.Close()
+
+	if fi, err := r.Stat(); err != nil {
+		return nil, err
+	} else if fi.Size() < z.limit {
+		return r, nil
+	}
+
+	if !filepath.IsLocal(name) {
+		return nil, zip.ErrInsecurePath
+	}
+
+	n := filepath.Join(z.p, name)
+	if _, err := os.Stat(n); errors.Is(err, os.ErrNotExist) {
+		w, err := os.Create(n)
+		if err != nil {
+			return nil, err
+		}
+		defer w.Close()
+
+		if _, err := io.Copy(w, r); err != nil {
+			return nil, err
+		}
+	} else if err != nil {
+		return nil, err
+	}
+
+	return os.Open(n)
+}
--- a/convert/reader.go
+++ b/convert/reader.go
@@ -31,31 +31,25 @@ func (t tensorBase) Shape() []uint64 {
 }

 const (
-	tensorKindFP32 uint32 = iota
-	tensorKindFP16
-	tensorKindMXFP4 = 4
-	tensorKindBF16  = 30
+	tensorKindF32 uint32 = iota
+	tensorKindF16
 )

 func (t tensorBase) Kind() uint32 {
 	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
-		strings.HasSuffix(t.name, ".bias") ||
 		t.name == "token_types.weight" ||
-		t.name == "v.positional_embedding_vlm" ||
-		t.name == "v.tile_position_embd.weight" ||
-		t.name == "v.pre_tile_position_embd.weight" ||
-		t.name == "v.post_tile_position_embd.weight" {
+		t.name == "v.positional_embedding_vlm" {
 		// these tensors are always F32
-		return tensorKindFP32
+		return 0
 	}

 	switch len(t.shape) {
 	case 0:
 		panic("invalid tensor shape")
 	case 1:
-		return tensorKindFP32
+		return tensorKindF32
 	default:
-		return tensorKindFP16
+		return tensorKindF16
 	}
 }

--- a/convert/reader_safetensors.go
+++ b/convert/reader_safetensors.go
@@ -1,7 +1,6 @@
 package convert

 import (
-	"bufio"
 	"bytes"
 	"encoding/binary"
 	"encoding/json"
@@ -9,12 +8,12 @@ import (
 	"fmt"
 	"io"
 	"io/fs"
-	"maps"
 	"slices"
 	"strings"

 	"github.com/d4l3k/go-bfloat16"
 	"github.com/x448/float16"
+	"golang.org/x/exp/maps"
 )

 type safetensorMetadata struct {
@@ -47,7 +46,8 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
 			return nil, err
 		}

-		keys := slices.Sorted(maps.Keys(headers))
+		keys := maps.Keys(headers)
+		slices.Sort(keys)

 		names := make(map[string]struct{}, len(keys))

@@ -94,15 +94,6 @@ type safetensor struct {
 	*tensorBase
 }

-func (st safetensor) Kind() uint32 {
-	kind := st.tensorBase.Kind()
-	if st.dtype == "BF16" && kind != tensorKindFP32 {
-		kind = tensorKindBF16
-	}
-
-	return kind
-}
-
 func (st safetensor) Clone() Tensor {
 	return &safetensor{
 		fs:     st.fs,
@@ -125,41 +116,26 @@ func (st safetensor) WriteTo(w io.Writer) (int64, error) {
 	}
 	defer f.Close()

-	r, err := func() (io.Reader, error) {
-		if readerAt, ok := f.(io.ReaderAt); ok {
-			return io.NewSectionReader(readerAt, st.offset, st.size), nil
-		} else if seeker, ok := f.(io.Seeker); ok {
-			_, err := seeker.Seek(st.offset, io.SeekStart)
-			return f, err
-		} else {
-			_, err := io.CopyN(io.Discard, f, st.offset)
-			return f, err
+	if seeker, ok := f.(io.Seeker); ok {
+		if _, err := seeker.Seek(st.offset, io.SeekStart); err != nil {
+			return 0, err
+		}
+	} else {
+		if _, err := io.CopyN(io.Discard, f, st.offset); err != nil {
+			return 0, err
 		}
-	}()
-	if err != nil {
-		return 0, err
-	}
-
-	br := bufio.NewReaderSize(r, min(32<<10, int(st.size)))
-	// special case when input and output are same type and the
-	// tensor doesn't need repacking
-	if (st.repacker == nil) &&
-		((st.dtype == "F32" && st.Kind() == tensorKindFP32) ||
-			(st.dtype == "F16" && st.Kind() == tensorKindFP16) ||
-			(st.dtype == "U8")) {
-		return io.CopyN(w, br, st.size)
 	}

 	var f32s []float32
 	switch st.dtype {
 	case "F32":
 		f32s = make([]float32, st.size/4)
-		if err = binary.Read(br, binary.LittleEndian, f32s); err != nil {
+		if err = binary.Read(f, binary.LittleEndian, f32s); err != nil {
 			return 0, err
 		}
 	case "F16":
 		u16s := make([]uint16, st.size/2)
-		if err = binary.Read(br, binary.LittleEndian, u16s); err != nil {
+		if err = binary.Read(f, binary.LittleEndian, u16s); err != nil {
 			return 0, err
 		}

@@ -170,7 +146,7 @@ func (st safetensor) WriteTo(w io.Writer) (int64, error) {

 	case "BF16":
 		u8s := make([]uint8, st.size)
-		if err = binary.Read(br, binary.LittleEndian, u8s); err != nil {
+		if err = binary.Read(f, binary.LittleEndian, u8s); err != nil {
 			return 0, err
 		}

@@ -187,18 +163,15 @@ func (st safetensor) WriteTo(w io.Writer) (int64, error) {
 	}

 	switch st.Kind() {
-	case tensorKindFP32:
+	case tensorKindF32:
 		return 0, binary.Write(w, binary.LittleEndian, f32s)
-	case tensorKindFP16:
+	case tensorKindF16:
 		f16s := make([]uint16, len(f32s))
 		for i := range f32s {
 			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
 		}

 		return 0, binary.Write(w, binary.LittleEndian, f16s)
-	case tensorKindBF16:
-		u8s := bfloat16.EncodeFloat32(f32s)
-		return 0, binary.Write(w, binary.LittleEndian, u8s)
 	default:
 		return 0, fmt.Errorf("unknown storage type: %d", st.Kind())
 	}
--- a/convert/reader_test.go
+++ b/convert/reader_test.go
@@ -1,232 +0,0 @@
-package convert
-
-import (
-	"bytes"
-	"encoding/binary"
-	"os"
-	"path/filepath"
-	"testing"
-
-	"github.com/d4l3k/go-bfloat16"
-	"github.com/google/go-cmp/cmp"
-	"github.com/x448/float16"
-)
-
-func TestSafetensors(t *testing.T) {
-	t.Parallel()
-
-	root, err := os.OpenRoot(t.TempDir())
-	if err != nil {
-		t.Fatal(err)
-	}
-	defer root.Close()
-
-	cases := []struct {
-		name,
-		dtype string
-		offset,
-		size int64
-		shape []uint64
-		setup func(*testing.T, *os.File)
-		want  []byte
-	}{
-		{
-			name:  "fp32-fp32",
-			dtype: "F32",
-			size:  32 * 4, // 32 floats, each 4 bytes
-			shape: []uint64{32},
-			setup: func(t *testing.T, f *os.File) {
-				f32s := make([]float32, 32)
-				for i := range f32s {
-					f32s[i] = float32(i)
-				}
-
-				if err := binary.Write(f, binary.LittleEndian, f32s); err != nil {
-					t.Fatal(err)
-				}
-			},
-			want: []byte{
-				0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x40, 0x40,
-				0x00, 0x00, 0x80, 0x40, 0x00, 0x00, 0xa0, 0x40, 0x00, 0x00, 0xc0, 0x40, 0x00, 0x00, 0xe0, 0x40,
-				0x00, 0x00, 0x00, 0x41, 0x00, 0x00, 0x10, 0x41, 0x00, 0x00, 0x20, 0x41, 0x00, 0x00, 0x30, 0x41,
-				0x00, 0x00, 0x40, 0x41, 0x00, 0x00, 0x50, 0x41, 0x00, 0x00, 0x60, 0x41, 0x00, 0x00, 0x70, 0x41,
-				0x00, 0x00, 0x80, 0x41, 0x00, 0x00, 0x88, 0x41, 0x00, 0x00, 0x90, 0x41, 0x00, 0x00, 0x98, 0x41,
-				0x00, 0x00, 0xa0, 0x41, 0x00, 0x00, 0xa8, 0x41, 0x00, 0x00, 0xb0, 0x41, 0x00, 0x00, 0xb8, 0x41,
-				0x00, 0x00, 0xc0, 0x41, 0x00, 0x00, 0xc8, 0x41, 0x00, 0x00, 0xd0, 0x41, 0x00, 0x00, 0xd8, 0x41,
-				0x00, 0x00, 0xe0, 0x41, 0x00, 0x00, 0xe8, 0x41, 0x00, 0x00, 0xf0, 0x41, 0x00, 0x00, 0xf8, 0x41,
-			},
-		},
-		{
-			name:  "fp32-fp16",
-			dtype: "F32",
-			size:  32 * 4, // 32 floats, each 4 bytes
-			shape: []uint64{16, 2},
-			setup: func(t *testing.T, f *os.File) {
-				f32s := make([]float32, 32)
-				for i := range f32s {
-					f32s[i] = float32(i)
-				}
-
-				if err := binary.Write(f, binary.LittleEndian, f32s); err != nil {
-					t.Fatal(err)
-				}
-			},
-			want: []byte{
-				0x00, 0x00, 0x00, 0x3c, 0x00, 0x40, 0x00, 0x42, 0x00, 0x44, 0x00, 0x45, 0x00, 0x46, 0x00, 0x47,
-				0x00, 0x48, 0x80, 0x48, 0x00, 0x49, 0x80, 0x49, 0x00, 0x4a, 0x80, 0x4a, 0x00, 0x4b, 0x80, 0x4b,
-				0x00, 0x4c, 0x40, 0x4c, 0x80, 0x4c, 0xc0, 0x4c, 0x00, 0x4d, 0x40, 0x4d, 0x80, 0x4d, 0xc0, 0x4d,
-				0x00, 0x4e, 0x40, 0x4e, 0x80, 0x4e, 0xc0, 0x4e, 0x00, 0x4f, 0x40, 0x4f, 0x80, 0x4f, 0xc0, 0x4f,
-			},
-		},
-		{
-			name:  "fp16-fp16",
-			dtype: "F16",
-			size:  32 * 2, // 32 floats, each 2 bytes
-			shape: []uint64{16, 2},
-			setup: func(t *testing.T, f *os.File) {
-				u16s := make([]uint16, 32)
-				for i := range u16s {
-					u16s[i] = float16.Fromfloat32(float32(i)).Bits()
-				}
-
-				if err := binary.Write(f, binary.LittleEndian, u16s); err != nil {
-					t.Fatal(err)
-				}
-			},
-			want: []byte{
-				0x00, 0x00, 0x00, 0x3c, 0x00, 0x40, 0x00, 0x42, 0x00, 0x44, 0x00, 0x45, 0x00, 0x46, 0x00, 0x47,
-				0x00, 0x48, 0x80, 0x48, 0x00, 0x49, 0x80, 0x49, 0x00, 0x4a, 0x80, 0x4a, 0x00, 0x4b, 0x80, 0x4b,
-				0x00, 0x4c, 0x40, 0x4c, 0x80, 0x4c, 0xc0, 0x4c, 0x00, 0x4d, 0x40, 0x4d, 0x80, 0x4d, 0xc0, 0x4d,
-				0x00, 0x4e, 0x40, 0x4e, 0x80, 0x4e, 0xc0, 0x4e, 0x00, 0x4f, 0x40, 0x4f, 0x80, 0x4f, 0xc0, 0x4f,
-			},
-		},
-		{
-			name:  "fp16-fp32",
-			dtype: "F16",
-			size:  32 * 2, // 32 floats, each 2 bytes
-			shape: []uint64{32},
-			setup: func(t *testing.T, f *os.File) {
-				u16s := make([]uint16, 32)
-				for i := range u16s {
-					u16s[i] = float16.Fromfloat32(float32(i)).Bits()
-				}
-
-				if err := binary.Write(f, binary.LittleEndian, u16s); err != nil {
-					t.Fatal(err)
-				}
-			},
-			want: []byte{
-				0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x40, 0x40,
-				0x00, 0x00, 0x80, 0x40, 0x00, 0x00, 0xa0, 0x40, 0x00, 0x00, 0xc0, 0x40, 0x00, 0x00, 0xe0, 0x40,
-				0x00, 0x00, 0x00, 0x41, 0x00, 0x00, 0x10, 0x41, 0x00, 0x00, 0x20, 0x41, 0x00, 0x00, 0x30, 0x41,
-				0x00, 0x00, 0x40, 0x41, 0x00, 0x00, 0x50, 0x41, 0x00, 0x00, 0x60, 0x41, 0x00, 0x00, 0x70, 0x41,
-				0x00, 0x00, 0x80, 0x41, 0x00, 0x00, 0x88, 0x41, 0x00, 0x00, 0x90, 0x41, 0x00, 0x00, 0x98, 0x41,
-				0x00, 0x00, 0xa0, 0x41, 0x00, 0x00, 0xa8, 0x41, 0x00, 0x00, 0xb0, 0x41, 0x00, 0x00, 0xb8, 0x41,
-				0x00, 0x00, 0xc0, 0x41, 0x00, 0x00, 0xc8, 0x41, 0x00, 0x00, 0xd0, 0x41, 0x00, 0x00, 0xd8, 0x41,
-				0x00, 0x00, 0xe0, 0x41, 0x00, 0x00, 0xe8, 0x41, 0x00, 0x00, 0xf0, 0x41, 0x00, 0x00, 0xf8, 0x41,
-			},
-		},
-		{
-			name:  "bf16-bf16",
-			dtype: "BF16",
-			size:  32 * 2, // 32 brain floats, each 2 bytes
-			shape: []uint64{16, 2},
-			setup: func(t *testing.T, f *os.File) {
-				f32s := make([]float32, 32)
-				for i := range f32s {
-					f32s[i] = float32(i)
-				}
-
-				if err := binary.Write(f, binary.LittleEndian, bfloat16.EncodeFloat32(f32s)); err != nil {
-					t.Fatal(err)
-				}
-			},
-			want: []byte{
-				0x00, 0x00, 0x80, 0x3f, 0x00, 0x40, 0x40, 0x40, 0x80, 0x40, 0xa0, 0x40, 0xc0, 0x40, 0xe0, 0x40,
-				0x00, 0x41, 0x10, 0x41, 0x20, 0x41, 0x30, 0x41, 0x40, 0x41, 0x50, 0x41, 0x60, 0x41, 0x70, 0x41,
-				0x80, 0x41, 0x88, 0x41, 0x90, 0x41, 0x98, 0x41, 0xa0, 0x41, 0xa8, 0x41, 0xb0, 0x41, 0xb8, 0x41,
-				0xc0, 0x41, 0xc8, 0x41, 0xd0, 0x41, 0xd8, 0x41, 0xe0, 0x41, 0xe8, 0x41, 0xf0, 0x41, 0xf8, 0x41,
-			},
-		},
-		{
-			name:  "bf16-fp32",
-			dtype: "BF16",
-			size:  32 * 2, // 32 brain floats, each 2 bytes
-			shape: []uint64{32},
-			setup: func(t *testing.T, f *os.File) {
-				f32s := make([]float32, 32)
-				for i := range f32s {
-					f32s[i] = float32(i)
-				}
-
-				if err := binary.Write(f, binary.LittleEndian, bfloat16.EncodeFloat32(f32s)); err != nil {
-					t.Fatal(err)
-				}
-			},
-			want: []byte{
-				0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x40, 0x40,
-				0x00, 0x00, 0x80, 0x40, 0x00, 0x00, 0xa0, 0x40, 0x00, 0x00, 0xc0, 0x40, 0x00, 0x00, 0xe0, 0x40,
-				0x00, 0x00, 0x00, 0x41, 0x00, 0x00, 0x10, 0x41, 0x00, 0x00, 0x20, 0x41, 0x00, 0x00, 0x30, 0x41,
-				0x00, 0x00, 0x40, 0x41, 0x00, 0x00, 0x50, 0x41, 0x00, 0x00, 0x60, 0x41, 0x00, 0x00, 0x70, 0x41,
-				0x00, 0x00, 0x80, 0x41, 0x00, 0x00, 0x88, 0x41, 0x00, 0x00, 0x90, 0x41, 0x00, 0x00, 0x98, 0x41,
-				0x00, 0x00, 0xa0, 0x41, 0x00, 0x00, 0xa8, 0x41, 0x00, 0x00, 0xb0, 0x41, 0x00, 0x00, 0xb8, 0x41,
-				0x00, 0x00, 0xc0, 0x41, 0x00, 0x00, 0xc8, 0x41, 0x00, 0x00, 0xd0, 0x41, 0x00, 0x00, 0xd8, 0x41,
-				0x00, 0x00, 0xe0, 0x41, 0x00, 0x00, 0xe8, 0x41, 0x00, 0x00, 0xf0, 0x41, 0x00, 0x00, 0xf8, 0x41,
-			},
-		},
-		{
-			name:  "u8-u8",
-			dtype: "U8",
-			size:  32, // 32 brain floats, each 1 bytes
-			shape: []uint64{32},
-			setup: func(t *testing.T, f *os.File) {
-				u8s := make([]uint8, 32)
-				for i := range u8s {
-					u8s[i] = uint8(i)
-				}
-
-				if err := binary.Write(f, binary.LittleEndian, u8s); err != nil {
-					t.Fatal(err)
-				}
-			},
-			want: []byte{
-				0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f,
-				0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f,
-			},
-		},
-	}
-
-	for _, tt := range cases {
-		t.Run(tt.name, func(t *testing.T) {
-			path := filepath.Base(t.Name())
-			st := safetensor{
-				fs:     root.FS(),
-				path:   path,
-				dtype:  tt.dtype,
-				offset: tt.offset,
-				size:   tt.size,
-				tensorBase: &tensorBase{
-					name:  tt.name,
-					shape: tt.shape,
-				},
-			}
-
-			f, err := root.Create(path)
-			if err != nil {
-				t.Fatal(err)
-			}
-			defer f.Close()
-
-			tt.setup(t, f)
-
-			var b bytes.Buffer
-			if _, err := st.WriteTo(&b); err != nil {
-				t.Fatal(err)
-			}
-
-			if diff := cmp.Diff(tt.want, b.Bytes()); diff != "" {
-				t.Errorf("safetensor.WriteTo() mismatch (-want +got):\n%s", diff)
-			}
-		})
-	}
-}
--- a/convert/tensor.go
+++ b/convert/tensor.go
@@ -1,129 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"io"
-	"iter"
-	"path"
-	"slices"
-	"strings"
-
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type split struct {
-	*strings.Replacer
-	dim int
-
-	// fn is an optional function to apply to the tensor after slicing
-	fn func(tensor.Tensor) (tensor.Tensor, error)
-}
-
-// splitDim splits a tensor along a specified dimension into multiple tensors. The dimension
-// is split evenly based on the number of replacers provided unless a specific count is given.
-func splitDim(t Tensor, dim int, splits ...split) iter.Seq[*ggml.Tensor] {
-	return func(yield func(*ggml.Tensor) bool) {
-		var offset int
-		for _, split := range splits {
-			t := t.Clone()
-			shape := slices.Clone(t.Shape())
-			shape[dim] = cmp.Or(uint64(split.dim), shape[dim]/uint64(len(splits)))
-
-			slice := slices.Repeat([]tensor.Slice{nil}, len(shape))
-			slice[dim] = tensor.S(offset, offset+int(shape[dim]))
-			offset += int(shape[dim])
-
-			t.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
-				dims := make([]int, len(shape))
-				for i := range shape {
-					dims[i] = int(shape[i])
-				}
-
-				var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-				tt, err := tt.Slice(slice...)
-				if err != nil {
-					return nil, err
-				}
-
-				tt = tensor.Materialize(tt)
-
-				if split.fn != nil {
-					tt, err = split.fn(tt)
-					if err != nil {
-						return nil, err
-					}
-				}
-
-				// flatten tensor so it can be written as a vector
-				if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
-					return nil, err
-				}
-
-				return native.VectorF32(tt.(*tensor.Dense))
-			})
-
-			if !yield(&ggml.Tensor{
-				Name:     split.Replace(t.Name()),
-				Kind:     t.Kind(),
-				Shape:    shape,
-				WriterTo: t,
-			}) {
-				break
-			}
-		}
-	}
-}
-
-type merge struct {
-	pattern, name string
-}
-
-// mergeTensors merges tensors that match a given pattern into a single tensor.
-func mergeTensors(unmatched []Tensor, merges ...merge) (out []*ggml.Tensor, _ []Tensor) {
-	var matched []Tensor
-	for i := range merges {
-		matched, unmatched = slicesSplitFunc(unmatched, func(t Tensor) bool {
-			matched, _ := path.Match(merges[i].pattern, t.Name())
-			return matched
-		})
-
-		if len(matched) > 0 {
-			out = append(out, &ggml.Tensor{
-				Name:     merges[i].name,
-				Kind:     matched[0].Kind(),
-				Shape:    append([]uint64{uint64(len(matched))}, matched[0].Shape()...),
-				WriterTo: mergeGroup(matched),
-			})
-		}
-	}
-
-	return out, unmatched
-}
-
-// slicesSplitFunc splits a slice into two slices based on a predicate function.
-func slicesSplitFunc[S ~[]E, E comparable](s S, fn func(e E) bool) (matched, unmatched S) {
-	for _, e := range s {
-		if fn(e) {
-			matched = append(matched, e)
-		} else {
-			unmatched = append(unmatched, e)
-		}
-	}
-
-	return matched, unmatched
-}
-
-type mergeGroup []Tensor
-
-func (g mergeGroup) WriteTo(w io.Writer) (int64, error) {
-	for _, t := range g {
-		if _, err := t.WriteTo(w); err != nil {
-			return 0, err
-		}
-	}
-
-	return 0, nil
-}
--- a/convert/tensor_test.go
+++ b/convert/tensor_test.go
@@ -1,953 +0,0 @@
-package convert
-
-import (
-	"bytes"
-	"encoding/binary"
-	"io"
-	"iter"
-	"slices"
-	"strings"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-	"github.com/ollama/ollama/fs/ggml"
-	"github.com/pdevine/tensor"
-)
-
-type fakeTensor struct {
-	name  string
-	shape []uint64
-	data  []float32
-
-	repacker Repacker
-}
-
-func (f fakeTensor) Name() string {
-	return f.name
-}
-
-func (f fakeTensor) Shape() []uint64 {
-	return f.shape
-}
-
-func (f fakeTensor) Kind() uint32 {
-	return 0
-}
-
-func (f *fakeTensor) SetRepacker(fn Repacker) {
-	f.repacker = fn
-}
-
-func (f fakeTensor) Clone() Tensor {
-	return &fakeTensor{
-		name:     f.name,
-		shape:    slices.Clone(f.shape),
-		data:     slices.Clone(f.data),
-		repacker: f.repacker,
-	}
-}
-
-func (f fakeTensor) WriteTo(w io.Writer) (n int64, err error) {
-	data := f.data
-	if f.repacker != nil {
-		data, err = f.repacker(f.name, data, f.shape)
-		if err != nil {
-			return 0, err
-		}
-	}
-
-	if err := binary.Write(w, binary.LittleEndian, data); err != nil {
-		return 0, err
-	}
-
-	return int64(len(data) * 4), nil
-}
-
-func mul(shape []uint64) int {
-	n := 1
-	for _, dim := range shape {
-		n *= int(dim)
-	}
-	return n
-}
-
-func TestSplitDim(t *testing.T) {
-	t.Run("2d", func(t *testing.T) {
-		r := fakeTensor{
-			name:  "a.b",
-			shape: []uint64{3, 4},
-			data:  []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11},
-		}
-
-		t.Run("no split", func(t *testing.T) {
-			for tt := range splitDim(&r, 0, split{Replacer: strings.NewReplacer("a", "x")}) {
-				if tt.Name != "x.b" {
-					t.Fatalf("expected name 'x', got '%s'", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 4}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("even split", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 1,
-				split{Replacer: strings.NewReplacer("a", "x")},
-				split{Replacer: strings.NewReplacer("b", "y")},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 4, 5, 8, 9}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'a.y', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{2, 3, 6, 7, 10, 11}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("uneven split", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 0,
-				split{Replacer: strings.NewReplacer("a", "x"), dim: 2},
-				split{Replacer: strings.NewReplacer("b", "y"), dim: 1},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{2, 4}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'a.y', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{1, 4}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{8, 9, 10, 11}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("three way split", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 0,
-				split{Replacer: strings.NewReplacer("a", "x"), dim: 1},
-				split{Replacer: strings.NewReplacer("b", "y"), dim: 1},
-				split{Replacer: strings.NewReplacer("b", "z"), dim: 1},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{1, 4}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 2, 3}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{1, 4}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{4, 5, 6, 7}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.z" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{1, 4}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{8, 9, 10, 11}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("uneven three way split", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 1,
-				split{Replacer: strings.NewReplacer("a", "x"), dim: 2},
-				split{Replacer: strings.NewReplacer("b", "y"), dim: 1},
-				split{Replacer: strings.NewReplacer("b", "z"), dim: 1},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 4, 5, 8, 9}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 1}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{2, 6, 10}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.z" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 1}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{3, 7, 11}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("split with transpose", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 1,
-				split{Replacer: strings.NewReplacer("a", "x")},
-				split{Replacer: strings.NewReplacer("b", "y"), fn: func(tt tensor.Tensor) (tensor.Tensor, error) {
-					return tensor.Transpose(tt, 1, 0)
-				}},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 4, 5, 8, 9}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'a.y', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{2, 6, 10, 3, 7, 11}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-	})
-	t.Run("3d", func(t *testing.T) {
-		r := fakeTensor{
-			name:  "a.b",
-			shape: []uint64{3, 4, 2},
-			data:  []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23},
-		}
-
-		t.Run("no split", func(t *testing.T) {
-			for tt := range splitDim(&r, 0, split{Replacer: strings.NewReplacer("a", "x")}) {
-				if tt.Name != "x.b" {
-					t.Fatalf("expected name 'x', got '%s'", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 4, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("even split", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 1,
-				split{Replacer: strings.NewReplacer("a", "x")},
-				split{Replacer: strings.NewReplacer("b", "y")},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 2, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'a.y', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 2, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("uneven split", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 0,
-				split{Replacer: strings.NewReplacer("a", "x"), dim: 2},
-				split{Replacer: strings.NewReplacer("b", "y"), dim: 1},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{2, 4, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'a.y', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{1, 4, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{16, 17, 18, 19, 20, 21, 22, 23}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("three way split", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 0,
-				split{Replacer: strings.NewReplacer("a", "x"), dim: 1},
-				split{Replacer: strings.NewReplacer("b", "y"), dim: 1},
-				split{Replacer: strings.NewReplacer("b", "z"), dim: 1},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{1, 4, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{1, 4, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{8, 9, 10, 11, 12, 13, 14, 15}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.z" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{1, 4, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{16, 17, 18, 19, 20, 21, 22, 23}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-
-		t.Run("uneven three way split", func(t *testing.T) {
-			next, stop := iter.Pull(splitDim(&r, 1,
-				split{Replacer: strings.NewReplacer("a", "x"), dim: 2},
-				split{Replacer: strings.NewReplacer("b", "y"), dim: 1},
-				split{Replacer: strings.NewReplacer("b", "z"), dim: 1},
-			))
-			defer stop()
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "x.b" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 2, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.y" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 1, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{4, 5, 12, 13, 20, 21}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-
-			{
-				tt, ok := next()
-				if !ok {
-					t.Fatal("expected at least one split")
-				}
-
-				if tt.Name != "a.z" {
-					t.Fatal("expected name 'x.b', got", tt.Name)
-				}
-
-				if diff := cmp.Diff(tt.Shape, []uint64{3, 1, 2}); diff != "" {
-					t.Errorf("unexpected shape (-want +got):\n%s", diff)
-				}
-
-				var b bytes.Buffer
-				if _, err := tt.WriteTo(&b); err != nil {
-					t.Fatal(err)
-				}
-
-				f32s := make([]float32, mul(tt.Shape))
-				if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(f32s, []float32{6, 7, 14, 15, 22, 23}); diff != "" {
-					t.Errorf("unexpected data (-want +got):\n%s", diff)
-				}
-			}
-		})
-	})
-}
-
-func TestMerge(t *testing.T) {
-	unmatched := []Tensor{
-		&fakeTensor{
-			name:  "a.0.b",
-			shape: []uint64{5, 2},
-			data:  []float32{10, 11, 12, 13, 14, 15, 16, 17, 18, 19},
-		},
-		&fakeTensor{
-			name:  "a.1.b",
-			shape: []uint64{5, 2},
-			data:  []float32{20, 21, 22, 23, 24, 25, 26, 27, 28, 29},
-		},
-		&fakeTensor{
-			name:  "c.0.d",
-			shape: []uint64{5, 2},
-			data:  []float32{30, 31, 32, 33, 34, 35, 36, 37, 38, 39},
-		},
-		&fakeTensor{
-			name:  "c.1.d",
-			shape: []uint64{5, 2},
-			data:  []float32{40, 41, 42, 43, 44, 45, 46, 47, 48, 49},
-		},
-		&fakeTensor{
-			name:  "e.0.f",
-			shape: []uint64{5, 2},
-			data:  []float32{50, 51, 52, 53, 54, 55, 56, 57, 58, 59},
-		},
-	}
-
-	checkMatched := func(t *testing.T, n int, matched []*ggml.Tensor) {
-		for i := range n {
-			got := matched[i]
-			if diff := cmp.Diff([]uint64{2, 5, 2}, got.Shape); diff != "" {
-				t.Errorf("unexpected (-want +got):\n%s", diff)
-			}
-
-			var b bytes.Buffer
-			if _, err := got.WriteTo(&b); err != nil {
-				t.Fatal(err)
-			}
-
-			f32s := make([]float32, 20)
-			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-				t.Fatal(err)
-			}
-
-			offset := 10 + (i * 20)
-			want := make([]float32, 20)
-			for j := range 20 {
-				want[j] = float32(offset + j)
-			}
-
-			if diff := cmp.Diff(want, f32s); diff != "" {
-				t.Errorf("unexpected data (-want +got):\n%s", diff)
-			}
-		}
-	}
-
-	t.Run("single merge", func(t *testing.T) {
-		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"})
-		if len(unmatched) != 3 {
-			t.Error("expected 3 remaining tensors, got", len(unmatched))
-		}
-
-		if len(matched) != 1 {
-			t.Error("expected 1 merged tensor, got", len(matched))
-		}
-
-		checkMatched(t, 1, matched)
-	})
-
-	t.Run("multiple merges", func(t *testing.T) {
-		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"}, merge{"c.*.d", "c.d"})
-		if len(unmatched) != 1 {
-			t.Error("expected 1 remaining tensors, got", len(unmatched))
-		}
-
-		if len(matched) != 2 {
-			t.Error("expected 2 merged tensor, got", len(matched))
-		}
-
-		checkMatched(t, 2, matched)
-	})
-
-	t.Run("no match", func(t *testing.T) {
-		matched, unmatched := mergeTensors(unmatched, merge{"x.*.y", "x.y"})
-		if len(unmatched) != 5 {
-			t.Error("expected 5 remaining tensors, got", len(unmatched))
-		}
-
-		if len(matched) != 0 {
-			t.Error("expected no merged tensors, got", len(matched))
-		}
-	})
-}
--- a/convert/tokenizer.go
+++ b/convert/tokenizer.go
@@ -8,10 +8,11 @@ import (
 	"fmt"
 	"io/fs"
 	"log/slog"
-	"maps"
 	"os"
 	"slices"
 	"strings"
+
+	"golang.org/x/exp/maps"
 )

 const (
@@ -109,7 +110,6 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 	}

 	if f, err := fsys.Open("tokenizer_config.json"); errors.Is(err, os.ErrNotExist) {
-		// noop
 	} else if err != nil {
 		return nil, err
 	} else {
@@ -171,34 +171,6 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 		}
 	}

-	if f, err := fsys.Open("generation_config.json"); errors.Is(err, os.ErrNotExist) {
-	} else if err != nil {
-		return nil, err
-	} else {
-		defer f.Close()
-
-		var p map[string]json.RawMessage
-		if err := json.NewDecoder(f).Decode(&p); err != nil {
-			return nil, err
-		}
-
-		for _, st := range specialTokenTypes {
-			if bts, ok := p[fmt.Sprintf("%s_token_id", st)]; ok {
-				var ids []int32
-				if err := json.Unmarshal(bts, &ids); err != nil {
-					// value is not a list so the existing ID is used
-					continue
-				}
-
-				if i := slices.IndexFunc(t.SpecialVocabulary, func(sv *SpecialVocabulary) bool {
-					return sv.Type == st
-				}); i >= 0 {
-					t.SpecialVocabulary[i].IDs = ids
-				}
-			}
-		}
-	}
-
 	return t, nil
 }

@@ -259,8 +231,11 @@ func parseVocabularyFromTokenizer(fsys fs.FS) (*Vocabulary, error) {
 		tokens[token.ID] = token
 	}

+	keys := maps.Keys(tokens)
+	slices.Sort(keys)
+
 	v := Vocabulary{Model: "gpt2"}
-	for _, k := range slices.Sorted(maps.Keys(tokens)) {
+	for _, k := range keys {
 		token := tokens[k]
 		v.Tokens = append(v.Tokens, token.Content)
 		v.Scores = append(v.Scores, float32(token.ID))
@@ -305,9 +280,6 @@ type SpecialVocabulary struct {
 	ID       int
 	Content  string
 	AddToken bool
-
-	// IDs is populated by generation_config.json
-	IDs []int32
 }

 func (sv SpecialVocabulary) Key() string {
--- a/convert/tokenizer_test.go
+++ b/convert/tokenizer_test.go
@@ -247,67 +247,6 @@ func TestParseTokenizer(t *testing.T) {
 				Pre: "default",
 			},
 		},
-		{
-			name: "generation config eos token ids",
-			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
-				"tokenizer.json": strings.NewReader(`{
-					"added_tokens": [
-						{
-							"id": 0,
-							"content": "<bos>",
-							"special": true
-						},
-						{
-							"id": 1,
-							"content": "<eos>",
-							"special": true
-						},
-						{
-							"id": 2,
-							"content": "<eot>",
-							"special": true
-						},
-						{
-							"id": 3,
-							"content": "<eom>",
-							"special": true
-						}
-					],
-					"model": {
-						"vocab": {
-							"<bos>": 0,
-							"<eos>": 1,
-							"<eot>": 2,
-							"<eom>": 3
-						}
-					}
-				}`),
-				"tokenizer_config.json": strings.NewReader(`{
-					"add_bos_token": true,
-					"add_eos_token": false,
-					"bos_token": "<bos>",
-					"eos_token": "<eos>"
-				}`),
-				"generation_config.json": strings.NewReader(`{
-					"bos_token_id": 0,
-					"eos_token_id": [1, 2, 3]
-				}`),
-			}),
-			specialTokenTypes: []string{"pad", "eos", "bos", "unk"},
-			want: &Tokenizer{
-				Vocabulary: &Vocabulary{
-					Model:  "gpt2",
-					Tokens: []string{"<bos>", "<eos>", "<eot>", "<eom>"},
-					Scores: []float32{0, 1, 2, 3},
-					Types:  []int32{3, 3, 3, 3},
-				},
-				SpecialVocabulary: []*SpecialVocabulary{
-					{Type: "eos", Content: "<eos>", ID: 1, IDs: []int32{1, 2, 3}, AddToken: false},
-					{Type: "bos", Content: "<bos>", ID: 0, AddToken: true},
-				},
-				Pre: "default",
-			},
-		},
 	}

 	for _, tt := range cases {
--- a/discover/amd_linux.go
+++ b/discover/amd_linux.go
@@ -58,7 +58,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 	driverMajor, driverMinor, err := AMDDriverVersion()
 	if err != nil {
 		// TODO - if we see users crash and burn with the upstreamed kernel this can be adjusted to hard-fail rocm support and fallback to CPU
-		slog.Warn("ollama recommends running the https://www.amd.com/en/support/download/linux-drivers.html", "error", err)
+		slog.Warn("ollama recommends running the https://www.amd.com/en/support/linux-drivers", "error", err)
 	}

 	// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
@@ -97,7 +97,6 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		return a < b
 	})
 	gpuCount := 0
-	gpuOrdinalID := 0
 	for _, match := range matches {
 		slog.Debug("evaluating amdgpu node " + match)
 		fp, err := os.Open(match)
@@ -188,6 +187,10 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			continue
 		}

+		// Keep track of numeric IDs based on valid GPUs
+		gpuID := gpuCount
+		gpuCount += 1
+
 		// Look up the memory for the current node
 		totalMemory := uint64(0)
 		usedMemory := uint64(0)
@@ -266,7 +269,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		if uniqueID != 0 {
 			ID = fmt.Sprintf("GPU-%016x", uniqueID)
 		} else {
-			ID = strconv.Itoa(gpuOrdinalID)
+			ID = strconv.Itoa(gpuID)
 		}

 		gpuInfo := RocmGPUInfo{
@@ -284,40 +287,13 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				DriverMinor:   driverMinor,
 			},
 			usedFilepath: usedFile,
-			index:        gpuCount,
+			index:        gpuID,
 		}

-		// Keep track of numeric IDs based on valid GPUs
-		gpuCount += 1
-
-		// If the user wants to filter to a subset of devices, filter out if we aren't a match
-		if len(visibleDevices) > 0 {
-			include := false
-			for _, visible := range visibleDevices {
-				if (uniqueID != 0 && visible == gpuInfo.ID) || visible == strconv.Itoa(gpuInfo.index) {
-					include = true
-					break
-				}
-			}
-			if !include {
-				reason := "filtering out device per user request"
-				slog.Info(reason, "id", gpuInfo.ID, "index", gpuInfo.index, "visible_devices", visibleDevices)
-				unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-					GpuInfo: gpuInfo.GpuInfo,
-					Reason:  reason,
-				})
-
-				continue
-			}
-		}
-
-		// Ordinal IDs are based on the visible GPUs
-		gpuOrdinalID += 1
-
 		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
 		if totalMemory < IGPUMemLimit {
 			reason := "unsupported Radeon iGPU detected skipping"
-			slog.Info(reason, "id", gpuInfo.ID, "total", format.HumanBytes2(totalMemory))
+			slog.Info(reason, "id", gpuID, "total", format.HumanBytes2(totalMemory))
 			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
 				GpuInfo: gpuInfo.GpuInfo,
 				Reason:  reason,
@@ -330,7 +306,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		}
 		if int(major) < minVer {
 			reason := fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch)
-			slog.Warn(reason, "gpu", gpuInfo.ID)
+			slog.Warn(reason, "gpu", gpuID)
 			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
 				GpuInfo: gpuInfo.GpuInfo,
 				Reason:  reason,
@@ -339,8 +315,29 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			continue
 		}

-		slog.Debug("amdgpu memory", "gpu", gpuInfo.ID, "total", format.HumanBytes2(totalMemory))
-		slog.Debug("amdgpu memory", "gpu", gpuInfo.ID, "available", format.HumanBytes2(totalMemory-usedMemory))
+		slog.Debug("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
+		slog.Debug("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
+
+		// If the user wants to filter to a subset of devices, filter out if we aren't a match
+		if len(visibleDevices) > 0 {
+			include := false
+			for _, visible := range visibleDevices {
+				if visible == gpuInfo.ID || visible == strconv.Itoa(gpuInfo.index) {
+					include = true
+					break
+				}
+			}
+			if !include {
+				reason := "filtering out device per user request"
+				slog.Info(reason, "id", gpuInfo.ID, "visible_devices", visibleDevices)
+				unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
+					GpuInfo: gpuInfo.GpuInfo,
+					Reason:  reason,
+				})
+
+				continue
+			}
+		}

 		// Final validation is gfx compatibility - load the library if we haven't already loaded it
 		// even if the user overrides, we still need to validate the library
--- a/discover/cuda_common.go
+++ b/discover/cuda_common.go
@@ -56,7 +56,6 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 				}
 			}
 		}
-		return "sbsa"
 	}

 	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
--- a/discover/gpu.go
+++ b/discover/gpu.go
@@ -263,8 +263,6 @@ func GetGPUInfo() GpuInfoList {
 				var driverMinor int
 				if cHandles.cudart != nil {
 					C.cudart_bootstrap(*cHandles.cudart, C.int(i), &memInfo)
-					driverMajor = int(cHandles.cudart.driver_major)
-					driverMinor = int(cHandles.cudart.driver_minor)
 				} else {
 					C.nvcuda_bootstrap(*cHandles.nvcuda, C.int(i), &memInfo)
 					driverMajor = int(cHandles.nvcuda.driver_major)
@@ -672,7 +670,7 @@ func loadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string, e
 }

 func getVerboseState() C.uint16_t {
-	if envconfig.LogLevel() < slog.LevelInfo {
+	if envconfig.Debug() {
 		return C.uint16_t(1)
 	}
 	return C.uint16_t(0)
--- a/discover/gpu_info_cudart.c
+++ b/discover/gpu_info_cudart.c
@@ -69,15 +69,18 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
  }

  int version = 0;
+  cudartDriverVersion_t driverVersion;
+  driverVersion.major = 0;
+  driverVersion.minor = 0;

  // Report driver version if we're in verbose mode, ignore errors
  ret = (*resp->ch.cudaDriverGetVersion)(&version);
  if (ret != CUDART_SUCCESS) {
    LOG(resp->ch.verbose, "cudaDriverGetVersion failed: %d\n", ret);
  } else {
-    resp->ch.driver_major = version / 1000;
-    resp->ch.driver_minor = (version - (resp->ch.driver_major * 1000)) / 10;
-    LOG(resp->ch.verbose, "CUDA driver version: %d-%d\n", resp->ch.driver_major, resp->ch.driver_minor);
+    driverVersion.major = version / 1000;
+    driverVersion.minor = (version - (driverVersion.major * 1000)) / 10;
+    LOG(resp->ch.verbose, "CUDA driver version: %d-%d\n", driverVersion.major, driverVersion.minor);
  }

  ret = (*resp->ch.cudaGetDeviceCount)(&resp->num_devices);
--- a/discover/gpu_info_cudart.h
+++ b/discover/gpu_info_cudart.h
@@ -29,6 +29,11 @@ typedef struct cudartMemory_st {
  size_t used;
 } cudartMemory_t;

+typedef struct cudartDriverVersion {
+  int major;
+  int minor;
+} cudartDriverVersion_t;
+
 typedef struct cudaUUID {
    unsigned char bytes[16];
 } cudaUUID_t;
@@ -118,8 +123,6 @@ typedef struct cudaDeviceProp {
 typedef struct cudart_handle {
  void *handle;
  uint16_t verbose;
-  int driver_major;
-  int driver_minor;
  cudartReturn_t (*cudaSetDevice)(int device);
  cudartReturn_t (*cudaDeviceSynchronize)(void);
  cudartReturn_t (*cudaDeviceReset)(void);
--- a/discover/types.go
+++ b/discover/types.go
@@ -171,8 +171,7 @@ func (si SystemInfo) GetOptimalThreadCount() int {
 // For each GPU, check if it does NOT support flash attention
 func (l GpuInfoList) FlashAttentionSupported() bool {
 	for _, gpu := range l {
-		supportsFA := gpu.Library == "cpu" ||
-			gpu.Library == "metal" ||
+		supportsFA := gpu.Library == "metal" ||
 			(gpu.Library == "cuda" && gpu.DriverMajor >= 7) ||
 			gpu.Library == "rocm"

--- a/docs/README.md
+++ b/docs/README.md
@@ -4,7 +4,6 @@
 * [Quickstart](../README.md#quickstart)
 * [Examples](./examples.md)
 * [Importing models](./import.md)
-* [MacOS Documentation](./macos.md)
 * [Linux Documentation](./linux.md)
 * [Windows Documentation](./windows.md)
 * [Docker Documentation](./docker.md)
--- a/docs/api.md
+++ b/docs/api.md
@@ -19,7 +19,7 @@

 ### Model names

-Model names follow a `model:tag` format, where `model` can have an optional namespace such as `example/model`. Some examples are `orca-mini:3b-q8_0` and `llama3:70b`. The tag is optional and, if not provided, will default to `latest`. The tag is used to identify a specific version.
+Model names follow a `model:tag` format, where `model` can have an optional namespace such as `example/model`. Some examples are `orca-mini:3b-q4_1` and `llama3:70b`. The tag is optional and, if not provided, will default to `latest`. The tag is used to identify a specific version.

 ### Durations

@@ -43,7 +43,6 @@ Generate a response for a given prompt with a provided model. This is a streamin
 - `prompt`: the prompt to generate a response for
 - `suffix`: the text after the model response
 - `images`: (optional) a list of base64-encoded images (for multimodal models such as `llava`)
- `think`: (for thinking models) should the model think before responding?

 Advanced parameters (optional):

@@ -491,39 +490,28 @@ 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
- `think`: (for thinking models) should the model think before responding?

 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
- `thinking`: (for thinking models) the model's thinking process
 - `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_name` (optional): add the name of the tool that was executed to inform the model of the result

 Advanced parameters (optional):

- `format`: the format to return a response in. Format can be `json` or a JSON schema.
+- `format`: the format to return a response in. Format can be `json` or a JSON schema. 
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)

-### Tool calling
-
-Tool calling is supported by providing a list of tools in the `tools` parameter. The model will generate a response that includes a list of tool calls. See the [Chat request (Streaming with tools)](#chat-request-streaming-with-tools) example below.
-
-Models can also explain the result of the tool call in the response. See the [Chat request (With history, with tools)](#chat-request-with-history-with-tools) example below.
-
-[See models with tool calling capabilities](https://ollama.com/search?c=tool).
-
 ### Structured outputs

 Structured outputs are supported by providing a JSON schema in the `format` parameter. The model will generate a response that matches the schema. See the [Chat request (Structured outputs)](#chat-request-structured-outputs) example below.

 ### Examples

-#### Chat request (Streaming)
+#### Chat Request (Streaming)

 ##### Request

@@ -578,88 +566,6 @@ Final response:
 }
 ```

-#### Chat request (Streaming with tools)
-
-##### Request
-
-```shell
-curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.2",
-  "messages": [
-    {
-      "role": "user",
-      "content": "what is the weather in tokyo?"
-    }
-  ],
-  "tools": [
-    {
-      "type": "function",
-      "function": {
-        "name": "get_weather",
-        "description": "Get the weather in a given city",
-        "parameters": {
-          "type": "object",
-          "properties": {
-            "city": {
-              "type": "string",
-              "description": "The city to get the weather for"
-            }
-          },
-          "required": ["city"]
-        }
-      }
-    }
-  ],
-  "stream": true
-}'
-```
-
-##### Response
-
-A stream of JSON objects is returned:
-```json
-{
-    "model": "llama3.2",
-    "created_at": "2025-07-07T20:22:19.184789Z",
-    "message": {
-        "role": "assistant",
-        "content": "",
-        "tool_calls": [
-            {
-                "function": {
-                    "name": "get_weather",
-                    "arguments": {
-                        "city": "Tokyo"
-                    }
-                },
-            }
-        ]
-    },
-    "done": false
-}
-```
-
-Final response:
-
-```json
-{
-  "model":"llama3.2",
-  "created_at":"2025-07-07T20:22:19.19314Z",
-  "message": {
-    "role": "assistant",
-    "content": ""
-  },
-  "done_reason": "stop",
-  "done": true,
-  "total_duration": 182242375,
-  "load_duration": 41295167,
-  "prompt_eval_count": 169,
-  "prompt_eval_duration": 24573166,
-  "eval_count": 15,
-  "eval_duration": 115959084
-}
-```
-
 #### Chat request (No streaming)

 ##### Request
@@ -697,74 +603,6 @@ curl http://localhost:11434/api/chat -d '{
 }
 ```

-#### Chat request (No streaming, with tools)
-
-##### Request
-
-
-```shell
-curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.2",
-  "messages": [
-    {
-      "role": "user",
-      "content": "what is the weather in tokyo?"
-    }
-  ],
-  "tools": [
-    {
-      "type": "function",
-      "function": {
-        "name": "get_weather",
-        "description": "Get the weather in a given city",
-        "parameters": {
-          "type": "object",
-          "properties": {
-            "city": {
-              "type": "string",
-              "description": "The city to get the weather for"
-            }
-          },
-          "required": ["city"]
-        }
-      }
-    }
-  ],
-  "stream": false 
-}'
-```
-
-##### Response
-
-```json
-{
-  "model": "llama3.2",
-  "created_at": "2025-07-07T20:32:53.844124Z",
-  "message": {
-    "role": "assistant",
-    "content": "",
-    "tool_calls": [
-      {
-        "function": {
-          "name": "get_weather",
-          "arguments": {
-            "city": "Tokyo"
-          }
-        },
-      }
-    ]
-  },
-  "done_reason": "stop",
-  "done": true,
-  "total_duration": 3244883583,
-  "load_duration": 2969184542,
-  "prompt_eval_count": 169,
-  "prompt_eval_duration": 141656333,
-  "eval_count": 18,
-  "eval_duration": 133293625
-}
-```
-
 #### Chat request (Structured outputs)

 ##### Request
@@ -871,87 +709,6 @@ Final response:
 }
 ```

-
-#### Chat request (With history, with tools)
-
-##### Request
-
-```shell
-curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.2",
-  "messages": [
-    {
-      "role": "user",
-      "content": "what is the weather in Toronto?"
-    },
-    // the message from the model appended to history
-    {
-      "role": "assistant",
-      "content": "",
-      "tool_calls": [
-        {
-          "function": {
-            "name": "get_temperature",
-            "arguments": {
-              "city": "Toronto"
-            }
-          },
-        }
-      ]
-    },
-    // the tool call result appended to history
-    {
-      "role": "tool",
-      "content": "11 degrees celsius",
-      "tool_name": "get_temperature",
-    }
-  ],
-  "stream": false,
-  "tools": [
-    {
-      "type": "function",
-      "function": {
-        "name": "get_weather",
-        "description": "Get the weather in a given city",
-        "parameters": {
-          "type": "object",
-          "properties": {
-            "city": {
-              "type": "string",
-              "description": "The city to get the weather for"
-            }
-          },
-          "required": ["city"]
-        }
-      }
-    }
-  ]
-}'
-```
-
-##### Response
-
-```json
-{
-  "model": "llama3.2",
-  "created_at": "2025-07-07T20:43:37.688511Z",
-  "message": {
-    "role": "assistant",
-    "content": "The current temperature in Toronto is 11°C."
-  },
-  "done_reason": "stop",
-  "done": true,
-  "total_duration": 890771750,
-  "load_duration": 707634750,
-  "prompt_eval_count": 94,
-  "prompt_eval_duration": 91703208,
-  "eval_count": 11,
-  "eval_duration": 90282125
-}
-
-```
-
-
 #### Chat request (with images)

 ##### Request
@@ -1195,8 +952,19 @@ If you are creating a model from a safetensors directory or from a GGUF file, yo

 | Type | Recommended |
 | --- | :-: |
+| q2_K | |
+| q3_K_L | |
+| q3_K_M | |
+| q3_K_S | |
+| q4_0 | |
+| q4_1 | |
 | q4_K_M | * |
 | q4_K_S | |
+| q5_0 | |
+| q5_1 | |
+| q5_K_M | |
+| q5_K_S | |
+| q6_K | |
 | q8_0 | * |

 ### Examples
@@ -1241,8 +1009,8 @@ Quantize a non-quantized model.

 ```shell
 curl http://localhost:11434/api/create -d '{
-  "model": "llama3.2:quantized",
-  "from": "llama3.2:3b-instruct-fp16",
+  "model": "llama3.1:quantized",
+  "from": "llama3.1:8b-instruct-fp16",
  "quantize": "q4_K_M"
 }'
 ```
@@ -1252,14 +1020,12 @@ curl http://localhost:11434/api/create -d '{
 A stream of JSON objects is returned:

 ```json
-{"status":"quantizing F16 model to Q4_K_M","digest":"0","total":6433687776,"completed":12302}
-{"status":"quantizing F16 model to Q4_K_M","digest":"0","total":6433687776,"completed":6433687552}
-{"status":"verifying conversion"}
-{"status":"creating new layer sha256:fb7f4f211b89c6c4928ff4ddb73db9f9c0cfca3e000c3e40d6cf27ddc6ca72eb"}
-{"status":"using existing layer sha256:966de95ca8a62200913e3f8bfbf84c8494536f1b94b49166851e76644e966396"}
-{"status":"using existing layer sha256:fcc5a6bec9daf9b561a68827b67ab6088e1dba9d1fa2a50d7bbcc8384e0a265d"}
-{"status":"using existing layer sha256:a70ff7e570d97baaf4e62ac6e6ad9975e04caa6d900d3742d37698494479e0cd"}
+{"status":"quantizing F16 model to Q4_K_M"}
+{"status":"creating new layer sha256:667b0c1932bc6ffc593ed1d03f895bf2dc8dc6df21db3042284a6f4416b06a29"}
+{"status":"using existing layer sha256:11ce4ee3e170f6adebac9a991c22e22ab3f8530e154ee669954c4bc73061c258"}
+{"status":"using existing layer sha256:0ba8f0e314b4264dfd19df045cde9d4c394a52474bf92ed6a3de22a4ca31a177"}
 {"status":"using existing layer sha256:56bb8bd477a519ffa694fc449c2413c6f0e1d3b1c88fa7e3c9d88d3ae49d4dcb"}
+{"status":"creating new layer sha256:455f34728c9b5dd3376378bfb809ee166c145b0b4c1f1a6feca069055066ef9a"}
 {"status":"writing manifest"}
 {"status":"success"}
 ```
@@ -1397,37 +1163,29 @@ A single JSON object will be returned.
 {
  "models": [
    {
-      "name": "deepseek-r1:latest",
-      "model": "deepseek-r1:latest",
-      "modified_at": "2025-05-10T08:06:48.639712648-07:00",
-      "size": 4683075271,
-      "digest": "0a8c266910232fd3291e71e5ba1e058cc5af9d411192cf88b6d30e92b6e73163",
+      "name": "codellama:13b",
+      "modified_at": "2023-11-04T14:56:49.277302595-07:00",
+      "size": 7365960935,
+      "digest": "9f438cb9cd581fc025612d27f7c1a6669ff83a8bb0ed86c94fcf4c5440555697",
      "details": {
-        "parent_model": "",
        "format": "gguf",
-        "family": "qwen2",
-        "families": [
-          "qwen2"
-        ],
-        "parameter_size": "7.6B",
-        "quantization_level": "Q4_K_M"
+        "family": "llama",
+        "families": null,
+        "parameter_size": "13B",
+        "quantization_level": "Q4_0"
      }
    },
    {
-      "name": "llama3.2:latest",
-      "model": "llama3.2:latest",
-      "modified_at": "2025-05-04T17:37:44.706015396-07:00",
-      "size": 2019393189,
-      "digest": "a80c4f17acd55265feec403c7aef86be0c25983ab279d83f3bcd3abbcb5b8b72",
+      "name": "llama3:latest",
+      "modified_at": "2023-12-07T09:32:18.757212583-08:00",
+      "size": 3825819519,
+      "digest": "fe938a131f40e6f6d40083c9f0f430a515233eb2edaa6d72eb85c50d64f2300e",
      "details": {
-        "parent_model": "",
        "format": "gguf",
        "family": "llama",
-        "families": [
-          "llama"
-        ],
-        "parameter_size": "3.2B",
-        "quantization_level": "Q4_K_M"
+        "families": null,
+        "parameter_size": "7B",
+        "quantization_level": "Q4_0"
      }
    }
  ]
@@ -1593,7 +1351,7 @@ Then there is a series of downloading responses. Until any of the download is co

 ```json
 {
-  "status": "pulling digestname",
+  "status": "downloading digestname",
  "digest": "digestname",
  "total": 2142590208,
  "completed": 241970
--- a/docs/benchmark.md
+++ b/docs/benchmark.md
@@ -0,0 +1,59 @@
+# Benchmark
+
+Go benchmark tests that measure end-to-end performance of a running Ollama server. Run these tests to evaluate model inference performance on your hardware and measure the impact of code changes.
+
+## When to use
+
+Run these benchmarks when:
+- Making changes to the model inference engine
+- Modifying model loading/unloading logic
+- Changing prompt processing or token generation code
+- Implementing a new model architecture
+- Testing performance across different hardware setups
+
+## Prerequisites
+- Ollama server running locally with `ollama serve` on `127.0.0.1:11434`
+## Usage and Examples
+
+>[!NOTE]
+>All commands must be run from the root directory of the Ollama project.
+
+Basic syntax:
+```bash
+go test -bench=. ./benchmark/... -m $MODEL_NAME
+```
+
+Required flags:
+- `-bench=.`: Run all benchmarks
+- `-m`: Model name to benchmark
+
+Optional flags:
+- `-count N`: Number of times to run the benchmark (useful for statistical analysis)
+- `-timeout T`: Maximum time for the benchmark to run (e.g. "10m" for 10 minutes)
+
+Common usage patterns:
+
+Single benchmark run with a model specified:
+```bash
+go test -bench=. ./benchmark/... -m llama3.3
+```
+
+## Output metrics
+
+The benchmark reports several key metrics:
+
+- `gen_tok/s`: Generated tokens per second
+- `prompt_tok/s`: Prompt processing tokens per second
+- `ttft_ms`: Time to first token in milliseconds
+- `load_ms`: Model load time in milliseconds
+- `gen_tokens`: Total tokens generated
+- `prompt_tokens`: Total prompt tokens processed
+
+Each benchmark runs two scenarios:
+- Cold start: Model is loaded from disk for each test
+- Warm start: Model is pre-loaded in memory
+
+Three prompt lengths are tested for each scenario:
+- Short prompt (100 tokens)
+- Medium prompt (500 tokens)
+- Long prompt (1000 tokens)
--- a/docs/development.md
+++ b/docs/development.md
@@ -118,7 +118,7 @@ To run tests, use `go test`:
 go test ./...
 ```

-> NOTE: In rare circumstances, you may need to change a package using the new
+> NOTE: In rare cirumstances, you may nedd to change a package using the new
 > "synctest" package in go1.24.
 >
 > If you do not have the "synctest" package enabled, you will not see build or
--- a/docs/docs.json
+++ b/docs/docs.json
@@ -1,75 +0,0 @@
-{
-  "$schema": "https://mintlify.com/docs.json",
-  "theme": "mint",
-  "background": {
-    "color": {
-      "light": "#ffffff",
-      "dark": "#000000"
-    }
-  },
-  "appearance": {
-    "default": "light"
-  },
-  "styling": {
-    "codeblocks": "system"
-  },
-  "contextual": {
-    "options": ["copy", "chatgpt", "claude", "view"]
-  },
-  "fonts": {
-    "heading": {
-      "family": "Inter"
-    },
-    "body": {
-      "family": "Inter"
-    }
-  },
-  "name": "Ollama",
-  "colors": {
-    "primary": "#000",
-    "light": "#b5b5b5",
-    "dark": "#fff"
-  },
-  "favicon": "/ollama.png",
-  "logo": {
-    "light": "/ollama.png",
-    "dark": "/favicon.svg"
-  },
-  "navigation": {
-    "tabs": [
-      {
-        "tab": "Documentation",
-        "groups": [
-          {
-            "group": "Home",
-            "pages": ["index", "quickstart", "faq", "troubleshooting"]
-          },
-          {
-            "group": "Platforms",
-            "pages": ["linux", "windows", "docker"]
-          },
-          {
-            "group": "Features",
-            "pages": [
-              "modelfile",
-              "apis",
-              "openai",
-              "import",
-              "gpu",
-              "benchmark"
-            ]
-          }
-        ]
-      },
-      {
-        "tab": "Development",
-        "groups": [
-          {
-            "group": " ",
-            "pages": ["development", "examples", "template"]
-          }
-        ]
-      }
-    ]
-  }
-}
--- a/docs/faq.md
+++ b/docs/faq.md
@@ -20,9 +20,9 @@ Please refer to the [GPU docs](./gpu.md).

 ## How can I specify the context window size?

-By default, Ollama uses a context window size of 4096 tokens for most models. The `gpt-oss` model has a default context window size of 8192 tokens.
+By default, Ollama uses a context window size of 4096 tokens. 

-This can be overridden in Settings in the Windows and macOS App, or with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use:
+This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use: 

 ```shell
 OLLAMA_CONTEXT_LENGTH=8192 ollama serve
@@ -46,8 +46,6 @@ curl http://localhost:11434/api/generate -d '{
 }'
 ```

-Setting the context length higher may cause the model to not be able to fit onto the GPU which make the model run more slowly.
-
 ## How can I tell if my model was loaded onto the GPU?

 Use the `ollama ps` command to see what models are currently loaded into memory.
@@ -59,8 +57,8 @@ ollama ps
 > **Output**:
 >
 > ```
-> NAME           ID              SIZE     PROCESSOR    CONTEXT    UNTIL
-> gpt-oss:20b    05afbac4bad6    16 GB    100% GPU     8192       4 minutes from now
+> 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:
@@ -150,11 +148,9 @@ docker build -t ollama-with-ca .
 docker run -d -e HTTPS_PROXY=https://my.proxy.example.com -p 11434:11434 ollama-with-ca
 ```

-## Does Ollama send my prompts and responses back to ollama.com?
+## Does Ollama send my prompts and answers back to ollama.com?

-If you're running a model locally, your prompts and responses will always stay on your machine. Ollama Turbo in the App allows you to run your queries on Ollama's servers if you don't have a powerful enough GPU. Web search lets a model query the web, giving you more accurate and up-to-date information. Both Turbo and web search require sending your prompts and responses to Ollama.com. This data is neither logged nor stored.
-
-If you don't want to see the Turbo and web search options in the app, you can disable them in Settings by turning on Airplane mode. In Airplane mode, all models will run locally, and your prompts and responses will stay on your machine.
+No. Ollama runs locally, and conversation data does not leave your machine.

 ## How can I expose Ollama on my network?

@@ -296,7 +292,7 @@ If too many requests are sent to the server, it will respond with a 503 error in

 ## How does Ollama handle concurrent requests?

-Ollama supports two levels of concurrent processing.  If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time.  For a given model, if there is sufficient available memory when the model is loaded, it can be configured to allow parallel request processing.
+Ollama supports two levels of concurrent processing.  If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time.  For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing.

 If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded.  As prior models become idle, one or more will be unloaded to make room for the new model.  Queued requests will be processed in order.  When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads.

@@ -305,7 +301,7 @@ Parallel request processing for a given model results in increasing the context
 The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:

 - `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory.  The default is 3 * the number of GPUs or 3 for CPU inference.
- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default is 1, and will handle 1 request per model at a time.
+- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default will auto-select either 4 or 1 based on available memory.
 - `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512

 Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6.2 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.
@@ -337,16 +333,3 @@ The currently available K/V cache quantization types are:
 How much the cache quantization impacts the model's response quality will depend on the model and the task.  Models that have a high GQA count (e.g. Qwen2) may see a larger impact on precision from quantization than models with a low GQA count.

 You may need to experiment with different quantization types to find the best balance between memory usage and quality.
-
-## How can I stop Ollama from starting when I login to my computer
-
-Ollama for Windows and macOS register as a login item during installation.  You can disable this if you prefer not to have Ollama automatically start.  Ollama will respect this setting across upgrades, unless you uninstall the application.
-
-**Windows**
- Remove `%APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup\Ollama.lnk`
-
-**MacOS Monterey (v12)**
- Open `Settings` -> `Users & Groups` -> `Login Items` and find the `Ollama` entry, then click the `-` (minus) to remove
-
-**MacOS Ventura (v13) and later**
- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.
--- a/docs/gpu.md
+++ b/docs/gpu.md
@@ -7,8 +7,6 @@ Check your compute compatibility to see if your card is supported:

 | Compute Capability | Family              | Cards                                                                                                       |
 | ------------------ | ------------------- | ----------------------------------------------------------------------------------------------------------- |
-| 12.0               | GeForce RTX 50xx    | `RTX 5060` `RTX 5060 Ti` `RTX 5070` `RTX 5070 Ti` `RTX 5080` `RTX 5090`                                     |
-|                    | NVIDIA Professioal  | `RTX PRO 4000 Blackwell` `RTX PRO 4500 Blackwell` `RTX PRO 5000 Blackwell` `RTX PRO 6000 Blackwell`         |
 | 9.0                | NVIDIA              | `H200` `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`                                                                                       |
--- a/docs/import.md
+++ b/docs/import.md
@@ -53,8 +53,6 @@ FROM /path/to/safetensors/directory

 If you create the Modelfile in the same directory as the weights, you can use the command `FROM .`.

-If you do not create the Modelfile, ollama will act as if there was a Modelfile with the command `FROM .`.
-
 Now run the `ollama create` command from the directory where you created the `Modelfile`:

 ```shell
@@ -134,12 +132,22 @@ success

 ### Supported Quantizations

+- `q4_0`
+- `q4_1`
+- `q5_0`
+- `q5_1`
 - `q8_0`

 #### K-means Quantizations

+- `q3_K_S`
+- `q3_K_M`
+- `q3_K_L`
 - `q4_K_S`
 - `q4_K_M`
+- `q5_K_S`
+- `q5_K_M`
+- `q6_K`


 ## Sharing your model on ollama.com
--- a/docs/linux.md
+++ b/docs/linux.md
@@ -16,7 +16,7 @@ curl -fsSL https://ollama.com/install.sh | sh
 Download and extract the package:

 ```shell
-curl -LO https://ollama.com/download/ollama-linux-amd64.tgz
+curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
 sudo tar -C /usr -xzf ollama-linux-amd64.tgz
 ```

@@ -34,11 +34,7 @@ ollama -v

 ### AMD GPU install

-If you have an AMD GPU, **also** download and extract the additional ROCm package:
-
-> [!IMPORTANT]
-> The ROCm tgz contains only AMD dependent libraries.  You must extract **both** `ollama-linux-amd64.tgz` and `ollama-linux-amd64-rocm.tgz` into the same location.
-
+If you have an AMD GPU, also download and extract the additional ROCm package:

 ```shell
 curl -L https://ollama.com/download/ollama-linux-amd64-rocm.tgz -o ollama-linux-amd64-rocm.tgz
@@ -116,8 +112,8 @@ sudo systemctl status ollama
 > While AMD has contributed the `amdgpu` driver upstream to the official linux
 > kernel source, the version is older and may not support all ROCm features. We
 > recommend you install the latest driver from
-> [AMD](https://www.amd.com/en/support/download/linux-drivers.html) for best support
-> of your Radeon GPU.
+> https://www.amd.com/en/support/linux-drivers for best support of your Radeon
+> GPU.

 ## Customizing

--- a/docs/macos.md
+++ b/docs/macos.md
@@ -1,42 +0,0 @@
-# Ollama for macOS
-
-## System Requirements
-
-* MacOS Monterey (v12) or newer
-* Apple M series (CPU and GPU support) or x86 (CPU only)
-
-
-## Filesystem Requirements
-
-The preferred method of installation is to mount the `ollama.dmg` and drag-and-drop the Ollama application to the system-wide `Applications` folder.  Upon startup, the Ollama app will verify the `ollama` CLI is present in your PATH, and if not detected, will prompt for permission to create a link in `/usr/local/bin`
-
-Once you've installed Ollama, you'll need additional space for storing the Large Language models, which can be tens to hundreds of GB in size.  If your home directory doesn't have enough space, you can change where the binaries are installed, and where the models are stored.
-
-### Changing Install Location
-
-To install the Ollama application somewhere other than `Applications`, place the Ollama application in the desired location, and ensure the CLI `Ollama.app/Contents/Resources/ollama` or a sym-link to the CLI can be found in your path.  Upon first start decline the "Move to Applications?" request.
-
-
-## Troubleshooting
-
-Ollama on MacOS stores files in a few different locations.
- `~/.ollama` contains models and configuration
- `~/.ollama/logs` contains logs
-    - *app.log* contains most recent logs from the GUI application
-    - *server.log* contains the most recent server logs
- `<install location>/Ollama.app/Contents/Resources/ollama` the CLI binary
-
-## Uninstall
-
-To fully remove Ollama from your system, remove the following files and folders:
-
-```
-sudo rm -rf /Applications/Ollama.app
-sudo rm /usr/local/bin/ollama
-rm -rf "~/Library/Application Support/Ollama"
-rm -rf "~/Library/Saved Application State/com.electron.ollama.savedState"
-rm -rf ~/Library/Caches/com.electron.ollama/
-rm -rf ~/Library/Caches/ollama
-rm -rf ~/Library/WebKit/com.electron.ollama
-rm -rf ~/.ollama
-```
--- a/docs/modelfile.md
+++ b/docs/modelfile.md
@@ -150,7 +150,7 @@ PARAMETER <parameter> <parametervalue>

 | Parameter      | Description                                                                                                                                                                                                                                             | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
-| num_ctx        | Sets the size of the context window used to generate the next token. (Default: 4096)                                                                                                                                                                    | int        | num_ctx 4096         |
+| num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
--- a/docs/openai.md
+++ b/docs/openai.md
@@ -72,7 +72,7 @@ client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
 # Define the schema for the response
 class FriendInfo(BaseModel):
    name: str
-    age: int
+    age: int 
    is_available: bool

 class FriendList(BaseModel):
--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@@ -9,7 +9,7 @@ cat ~/.ollama/logs/server.log
 On **Linux** systems with systemd, the logs can be found with this command:

 ```shell
-journalctl -u ollama --no-pager --follow --pager-end
+journalctl -u ollama --no-pager --follow --pager-end 
 ```

 When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
@@ -23,7 +23,7 @@ docker logs <container-name>
 If manually running `ollama serve` in a terminal, the logs will be on that terminal.

 When you run Ollama on **Windows**, there are a few different locations. You can view them in the explorer window by hitting `<cmd>+R` and type in:
- `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log`
+- `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log` 
 - `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored

@@ -38,7 +38,7 @@ Join the [Discord](https://discord.gg/ollama) for help interpreting the logs.

 ## LLM libraries

-Ollama includes multiple LLM libraries compiled for different GPUs and CPU vector features. Ollama tries to pick the best one based on the capabilities of your system. If this autodetection has problems, or you run into other problems (e.g. crashes in your GPU) you can workaround this by forcing a specific LLM library. `cpu_avx2` will perform the best, followed by `cpu_avx` and the slowest but most compatible is `cpu`. Rosetta emulation under MacOS will work with the `cpu` library.
+Ollama includes multiple LLM libraries compiled for different GPUs and CPU vector features. Ollama tries to pick the best one based on the capabilities of your system. If this autodetection has problems, or you run into other problems (e.g. crashes in your GPU) you can workaround this by forcing a specific LLM library. `cpu_avx2` will perform the best, followed by `cpu_avx` an the slowest but most compatible is `cpu`. Rosetta emulation under MacOS will work with the `cpu` library. 

 In the server log, you will see a message that looks something like this (varies from release to release):

@@ -97,7 +97,7 @@ If none of those resolve the problem, gather additional information and file an

 On linux, AMD GPU access typically requires `video` and/or `render` group membership to access the `/dev/kfd` device.  If permissions are not set up correctly, Ollama will detect this and report an error in the server log.

-When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -lnd /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the **numeric** group IDs on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.   For example, in the following output `crw-rw---- 1 0  44 226,   0 Sep 16 16:55 /dev/dri/card0` the group ID column is `44`
+When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -lnd /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the **numeric** group IDs on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.   For example, in the following output `crw-rw---- 1 0  44 226,   0 Sep 16 16:55 /dev/dri/card0` the group ID column is `44` 

 If you are experiencing problems getting Ollama to correctly discover or use your GPU for inference, the following may help isolate the failure.
 - `AMD_LOG_LEVEL=3` Enable info log levels in the AMD HIP/ROCm libraries.  This can help show more detailed error codes that can help troubleshoot problems
--- a/docs/turbo.md
+++ b/docs/turbo.md
@@ -1,107 +0,0 @@
-# Turbo
-
-> ⚠️ Turbo is preview
-
-Ollama’s [Turbo](https://ollama.com/turbo) is a new way to run open-source models with acceleration from datacenter-grade hardware.
-
-Currently, the following models are available in Turbo:
-
- `gpt-oss:20b`
- `gpt-oss:120b`
-
-## Get started
-
-### Ollama for macOS & Windows
-
-Download Ollama
-
- Select a model such as `gpt-oss:20b` or `gpt-oss:120b`
- Click on **Turbo**. You’ll be prompted to create an account or sign in
-
-### Ollama’s CLI
-
- [Sign up](https://ollama.com/signup) for an Ollama account
- Add your Ollama key [to ollama.com](https://ollama.com/settings/keys).
-
-  On macOS and Linux:
-
-  ```shell
-  cat ~/.ollama/id_ed25519.pub
-  ```
-
-  On Windows:
-
-  ```
-  type "%USERPROFILE%\.ollama\id_ed25519.pub"
-  ```
-
- Then run a model setting `OLLAMA_HOST` to `ollama.com`:
-  ```shell
-  OLLAMA_HOST=ollama.com ollama run gpt-oss:120b
-  ```
-
-### Ollama’s Python library
-
- Download Ollama's [Python library](https://github.com/ollama/ollama-python)
- [Sign up](https://ollama.com/signup) for an Ollama account
- Create an API key by visiting https://ollama.com/settings/keys
-
-```python
-from ollama import Client
-
-client = Client(
-    host="https://ollama.com",
-    headers={'Authorization': '<api key>'}
-)
-
-messages = [
-  {
-    'role': 'user',
-    'content': 'Why is the sky blue?',
-  },
-]
-
-for part in client.chat('gpt-oss:120b', messages=messages, stream=True):
-  print(part['message']['content'], end='', flush=True)
-```
-
-### Ollama’s JavaScript library
-
- Download Ollama's [JavaScript library](https://github.com/ollama/ollama-js)
- [Sign up](https://ollama.com/signup) for an Ollama account
- Create an API key by visiting https://ollama.com/settings/keys
-
-```typescript
-import { Ollama } from 'ollama';
-
-const ollama = new Ollama({
-  host: 'https://ollama.com',
-  headers: {
-	  Authorization: "Bearer <api key>"
-  }
-});
-
-const response = await ollama.chat({
-  model: 'gpt-oss:120b',
-  messages: [{ role: 'user', content: 'Explain quantum computing' }],
-  stream: true
-});
-
-for await (const part of response) {
-    process.stdout.write(part.message.content)
-}
-```
-
-### Community integrations
-
-Turbo mode is also compatible with several community integrations.
-
-#### Open WebUI
-
- Go to **settings** → **Admin settings** → **Connections**
- Under **Ollama API,** click **+**
- For the **URL** put `https://ollama.com`
- For the **API key,** create an API key on https://ollama.com/settings/keys and add it.
- Click **Save**
-
-Now, if you navigate to the model selector, Turbo models should be available under **External**.
--- a/docs/windows.md
+++ b/docs/windows.md
@@ -30,6 +30,20 @@ To install the Ollama application in a location different than your home directo
 OllamaSetup.exe /DIR="d:\some\location"
 ```

+### Changing Model Location
+
+To change where Ollama stores the downloaded models instead of using your home directory, set the environment variable `OLLAMA_MODELS` in your user account.
+
+1. Start the Settings (Windows 11) or Control Panel (Windows 10) application and search for _environment variables_.
+
+2. Click on _Edit environment variables for your account_.
+
+3. Edit or create a new variable for your user account for `OLLAMA_MODELS` where you want the models stored
+
+4. Click OK/Apply to save.
+
+If Ollama is already running, Quit the tray application and relaunch it from the Start menu, or a new terminal started after you saved the environment variables.
+
 ## API Access

 Here's a quick example showing API access from `powershell`
@@ -68,9 +82,9 @@ If you'd like to install or integrate Ollama as a service, a standalone
 `ollama-windows-amd64.zip` zip file is available containing only the Ollama CLI
 and GPU library dependencies for Nvidia.  If you have an AMD GPU, also download
 and extract the additional ROCm package `ollama-windows-amd64-rocm.zip` into the
-same directory.  Both zip files are necessary for a complete AMD installation.
-This allows for embedding Ollama in existing applications, or running it as a
-system service via `ollama serve` with tools such as [NSSM](https://nssm.cc/). 
+same directory.  This allows for embedding Ollama in existing applications, or
+running it as a system service via `ollama serve` with tools such as
+[NSSM](https://nssm.cc/). 

 > [!NOTE]  
 > If you are upgrading from a prior version, you should remove the old directories first.
--- a/envconfig/config.go
+++ b/envconfig/config.go
@@ -149,22 +149,9 @@ func Bool(k string) func() bool {
 	}
 }

-// LogLevel returns the log level for the application.
-// Values are 0 or false INFO (Default), 1 or true DEBUG, 2 TRACE
-func LogLevel() slog.Level {
-	level := slog.LevelInfo
-	if s := Var("OLLAMA_DEBUG"); s != "" {
-		if b, _ := strconv.ParseBool(s); b {
-			level = slog.LevelDebug
-		} else if i, _ := strconv.ParseInt(s, 10, 64); i != 0 {
-			level = slog.Level(i * -4)
-		}
-	}
-
-	return level
-}
-
 var (
+	// Debug enabled additional debug information.
+	Debug = Bool("OLLAMA_DEBUG")
 	// FlashAttention enables the experimental flash attention feature.
 	FlashAttention = Bool("OLLAMA_FLASH_ATTENTION")
 	// KvCacheType is the quantization type for the K/V cache.
@@ -183,10 +170,6 @@ var (
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
 	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
-	// Auth enables authentication between the Ollama client and server
-	UseAuth = Bool("OLLAMA_AUTH")
-	// Enable the new memory estimation logic
-	NewMemoryEstimates = Bool("OLLAMA_NEW_ESTIMATES")
 )

 func String(s string) func() string {
@@ -221,11 +204,13 @@ func Uint(key string, defaultValue uint) func() uint {

 var (
 	// NumParallel sets the number of parallel model requests. NumParallel can be configured via the OLLAMA_NUM_PARALLEL environment variable.
-	NumParallel = Uint("OLLAMA_NUM_PARALLEL", 1)
+	NumParallel = Uint("OLLAMA_NUM_PARALLEL", 0)
 	// MaxRunners sets the maximum number of loaded models. MaxRunners can be configured via the OLLAMA_MAX_LOADED_MODELS environment variable.
 	MaxRunners = Uint("OLLAMA_MAX_LOADED_MODELS", 0)
 	// MaxQueue sets the maximum number of queued requests. MaxQueue can be configured via the OLLAMA_MAX_QUEUE environment variable.
 	MaxQueue = Uint("OLLAMA_MAX_QUEUE", 512)
+	// MaxVRAM sets a maximum VRAM override in bytes. MaxVRAM can be configured via the OLLAMA_MAX_VRAM environment variable.
+	MaxVRAM = Uint("OLLAMA_MAX_VRAM", 0)
 )

 func Uint64(key string, defaultValue uint64) func() uint64 {
@@ -253,7 +238,7 @@ type EnvVar struct {

 func AsMap() map[string]EnvVar {
 	ret := map[string]EnvVar{
-		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", LogLevel(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
+		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
 		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
 		"OLLAMA_KV_CACHE_TYPE":     {"OLLAMA_KV_CACHE_TYPE", KvCacheType(), "Quantization type for the K/V cache (default: f16)"},
 		"OLLAMA_GPU_OVERHEAD":      {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},
@@ -272,7 +257,6 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
 		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
-		"OLLAMA_NEW_ESTIMATES":     {"OLLAMA_NEW_ESTIMATES", NewMemoryEstimates(), "Enable the new memory estimation logic"},

 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
--- a/envconfig/config_test.go
+++ b/envconfig/config_test.go
@@ -1,13 +1,11 @@
 package envconfig

 import (
-	"log/slog"
 	"math"
 	"testing"
 	"time"

 	"github.com/google/go-cmp/cmp"
-	"github.com/ollama/ollama/logutil"
 )

 func TestHost(t *testing.T) {
@@ -294,34 +292,3 @@ func TestContextLength(t *testing.T) {
 		})
 	}
 }
-
-func TestLogLevel(t *testing.T) {
-	cases := map[string]slog.Level{
-		// Default to INFO
-		"":      slog.LevelInfo,
-		"false": slog.LevelInfo,
-		"f":     slog.LevelInfo,
-		"0":     slog.LevelInfo,
-
-		// True values enable Debug
-		"true": slog.LevelDebug,
-		"t":    slog.LevelDebug,
-
-		// Positive values increase verbosity
-		"1": slog.LevelDebug,
-		"2": logutil.LevelTrace,
-
-		// Negative values decrease verbosity
-		"-1": slog.LevelWarn,
-		"-2": slog.LevelError,
-	}
-
-	for k, v := range cases {
-		t.Run(k, func(t *testing.T) {
-			t.Setenv("OLLAMA_DEBUG", k)
-			if i := LogLevel(); i != v {
-				t.Errorf("%s: expected %d, got %d", k, v, i)
-			}
-		})
-	}
-}
--- a/fs/config.go
+++ b/fs/config.go
@@ -10,5 +10,4 @@ type Config interface {
 	Strings(string, ...[]string) []string
 	Ints(string, ...[]int32) []int32
 	Floats(string, ...[]float32) []float32
-	Bools(string, ...[]bool) []bool
 }
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -1,7 +1,6 @@
 package ggml

 import (
-	"cmp"
 	"encoding/binary"
 	"errors"
 	"fmt"
@@ -16,7 +15,6 @@ import (
 type GGML struct {
 	container
 	model
-	Length int64
 }

 type model interface {
@@ -35,8 +33,7 @@ func (kv KV) Kind() string {
 }

 func (kv KV) ParameterCount() uint64 {
-	val, _ := keyValue(kv, "general.parameter_count", uint64(0))
-	return val
+	return keyValue(kv, "general.parameter_count", uint64(0))
 }

 func (kv KV) FileType() FileType {
@@ -55,27 +52,16 @@ func (kv KV) EmbeddingLength() uint64 {
 	return uint64(kv.Uint("embedding_length"))
 }

-func (kv KV) HeadCountMax() uint64 {
-	// TODO(drifkin): using the max value can cause an overestimation. In the
-	// future if array values become more popular, we can adapt the more invasive
-	// <https://github.com/ollama/ollama/pull/10225>
-	return uint64(kv.UintOrMaxArrayValue("attention.head_count", 1))
+func (kv KV) HeadCount() uint64 {
+	return uint64(kv.Uint("attention.head_count"))
 }

-func (kv KV) HeadCountMin() uint64 {
-	return uint64(kv.UintOrMinArrayValue("attention.head_count", 1))
+func (kv KV) HeadCountKV() uint64 {
+	return uint64(kv.Uint("attention.head_count_kv", 1))
 }

-func (kv KV) HeadCountKVMax() uint64 {
-	return uint64(kv.UintOrMaxArrayValue("attention.head_count_kv", 1))
-}
-
-func (kv KV) HeadCountKVMin() uint64 {
-	return uint64(kv.UintOrMinArrayValue("attention.head_count_kv", 1))
-}
-
-func (kv KV) EmbeddingHeadCountMax() uint64 {
-	if heads := kv.HeadCountMin(); heads > 0 {
+func (kv KV) EmbeddingHeadCount() uint64 {
+	if heads := kv.HeadCount(); heads > 0 {
 		return kv.EmbeddingLength() / heads
 	}

@@ -83,11 +69,15 @@ func (kv KV) EmbeddingHeadCountMax() uint64 {
 }

 func (kv KV) EmbeddingHeadCountK() uint64 {
-	return uint64(kv.Uint("attention.key_length", uint32(kv.EmbeddingHeadCountMax())))
+	return uint64(kv.Uint("attention.key_length", uint32(kv.EmbeddingHeadCount())))
 }

 func (kv KV) EmbeddingHeadCountV() uint64 {
-	return uint64(kv.Uint("attention.value_length", uint32(kv.EmbeddingHeadCountMax())))
+	return uint64(kv.Uint("attention.value_length", uint32(kv.EmbeddingHeadCount())))
+}
+
+func (kv KV) GQA() uint64 {
+	return kv.HeadCount() / kv.HeadCountKV()
 }

 func (kv KV) ContextLength() uint64 {
@@ -99,88 +89,42 @@ func (kv KV) ChatTemplate() string {
 }

 func (kv KV) String(key string, defaultValue ...string) string {
-	val, _ := keyValue(kv, key, append(defaultValue, "")...)
-	return val
+	return keyValue(kv, key, append(defaultValue, "")...)
 }

 func (kv KV) Uint(key string, defaultValue ...uint32) uint32 {
-	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
-	return val
+	return keyValue(kv, key, append(defaultValue, 0)...)
 }

 func (kv KV) Float(key string, defaultValue ...float32) float32 {
-	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
-	return val
+	return keyValue(kv, key, append(defaultValue, 0)...)
 }

 func (kv KV) Bool(key string, defaultValue ...bool) bool {
-	val, _ := keyValue(kv, key, append(defaultValue, false)...)
-	return val
-}
-
-func (kv KV) UintOrMaxArrayValue(key string, defaultValue uint32) uint32 {
-	_, max := kv.UintOrArrayValue(key, defaultValue)
-	return max
-}
-
-func (kv KV) UintOrMinArrayValue(key string, defaultValue uint32) uint32 {
-	min, _ := kv.UintOrArrayValue(key, defaultValue)
-	return min
-}
-
-func (kv KV) UintOrArrayValue(key string, defaultValue uint32) (uint32, uint32) {
-	if u32, ok := keyValue(kv, key, uint32(0)); ok {
-		return u32, u32
-	} else if u32s, ok := keyValue(kv, key, &array[uint32]{}); ok {
-		min := slices.Min(u32s.values)
-		max := slices.Max(u32s.values)
-		return min, max
-	} else if i32s, ok := keyValue(kv, key, &array[int32]{}); ok {
-		min := slices.Min(i32s.values)
-		max := slices.Max(i32s.values)
-		if min < 0 || max < 0 {
-			slog.Warn("array values are unexpectedly negative", "key", key, "min", min, "max", max)
-		}
-		return uint32(min), uint32(max)
-	}
-
-	return defaultValue, defaultValue
+	return keyValue(kv, key, append(defaultValue, false)...)
 }

 func (kv KV) Strings(key string, defaultValue ...[]string) []string {
-	val, _ := keyValue(kv, key, &array[string]{values: append(defaultValue, []string(nil))[0]})
-	return val.values
+	return keyValue(kv, key, &array[string]{values: append(defaultValue, []string(nil))[0]}).values
 }

 func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
-	val, _ := keyValue(kv, key, &array[int32]{values: append(defaultValue, []int32(nil))[0]})
-	return val.values
+	return keyValue(kv, key, &array[int32]{values: append(defaultValue, []int32(nil))[0]}).values
 }

 func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
-	val, _ := keyValue(kv, key, &array[uint32]{values: append(defaultValue, []uint32(nil))[0]})
-	return val.values
+	return keyValue(kv, key, &array[uint32]{values: append(defaultValue, []uint32(nil))[0]}).values
 }

 func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
-	val, _ := keyValue(kv, key, &array[float32]{values: append(defaultValue, []float32(nil))[0]})
-	return val.values
-}
-
-func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
-	val, _ := keyValue(kv, key, &array[bool]{values: append(defaultValue, []bool(nil))[0]})
-	return val.values
+	return keyValue(kv, key, &array[float32]{values: append(defaultValue, []float32(nil))[0]}).values
 }

 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
 		"gemma3",
-		"gemma3n",
 		"mistral3",
 		"llama4",
-		"mllama",
-		"qwen25vl",
-		"gptoss", "gpt-oss",
 	}, kv.Architecture())
 }

@@ -196,17 +140,17 @@ type arrayValueTypes interface {
 		*array[string] | *array[float32] | *array[float64] | *array[bool]
 }

-func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) (T, bool) {
+func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) T {
 	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
 		key = kv.Architecture() + "." + key
 	}

-	if val, ok := kv[key].(T); ok {
-		return val, true
+	if val, ok := kv[key]; ok {
+		return val.(T)
 	}

-	slog.Debug("key with type not found", "key", key, "default", defaultValue[0])
-	return defaultValue[0], false
+	slog.Debug("key not found", "key", key, "default", defaultValue[0])
+	return defaultValue[0]
 }

 type Tensors struct {
@@ -282,7 +226,7 @@ func (t Tensor) block() (n int) {
 }

 func (t Tensor) blockSize() uint64 {
-	return TensorType(t.Kind).BlockSize()
+	return (TensorType)(t.Kind).BlockSize()
 }

 func (t TensorType) BlockSize() uint64 {
@@ -300,7 +244,6 @@ func (t TensorType) BlockSize() uint64 {
 	case
 		2,  // Q4_0
 		3,  // Q4_1
-		4,  // MXFP4
 		6,  // Q5_0
 		7,  // Q5_1
 		8,  // Q8_0
@@ -328,8 +271,6 @@ func (t TensorType) TypeSize() uint64 {
 		return 2 + blockSize/2
 	case TensorTypeQ4_1:
 		return 2 + 2 + blockSize/2
-	case TensorTypeMXFP4, 39:
-		return 1 + blockSize/2
 	case TensorTypeQ5_0:
 		return 2 + 4 + blockSize/2
 	case TensorTypeQ5_1:
@@ -443,12 +384,12 @@ func DetectContentType(b []byte) string {
 //
 // It collects array values for arrays with a size less than or equal to
 // maxArraySize. If the maxArraySize is negative, all arrays are collected.
-func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
+func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)

 	var magic uint32
 	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
-		return nil, err
+		return nil, 0, err
 	}

 	var c container
@@ -458,47 +399,42 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	case FILE_MAGIC_GGUF_BE:
 		c = &containerGGUF{ByteOrder: binary.BigEndian, maxArraySize: maxArraySize}
 	default:
-		return nil, errors.New("invalid file magic")
+		return nil, 0, errors.New("invalid file magic")
 	}

 	model, err := c.Decode(rs)
 	if err != nil {
-		return nil, err
+		return nil, 0, err
 	}

 	offset, err := rs.Seek(0, io.SeekCurrent)
 	if err != nil {
-		return nil, err
+		return nil, 0, err
 	}

 	// final model type
 	return &GGML{
 		container: c,
 		model:     model,
-		Length:    offset,
-	}, nil
+	}, offset, nil
 }

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

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

 	layers := f.Tensors().GroupLayers()

 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	var kvTotal uint64
 	kv = make([]uint64, f.KV().BlockCount())
 	for i := range kv {
 		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
-		kvTotal += kv[i]
 	}

 	switch f.KV().Architecture() {
@@ -564,7 +500,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 			// vocab graph
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
-	case "gemma", "gemma2", "gemma3", "gemma3n":
+	case "gemma", "gemma2", "gemma3":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
 			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
@@ -577,11 +513,6 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 				embedding*embeddingHeadsK*heads*9/16,
 		)

-		if f.KV().Architecture() == "gemma3n" {
-			fullOffload *= 4
-			partialOffload *= 4
-		}
-
 		// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
 		// engine. Gemma3 always uses the Ollama engine.
 		if f.KV().Architecture() == "gemma3" {
@@ -667,17 +598,6 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 					4*qkvBias.Shape[0],
 			)
 		}
-	case "gptoss", "gpt-oss":
-		kv = make([]uint64, f.KV().BlockCount())
-		for i := range kv {
-			kv[i] = uint64(float64((embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
-			if i%2 == 0 {
-				kv[i] *= (uint64(numParallel)*4096 + batch)
-			} else {
-				kv[i] *= context
-			}
-		}
-		partialOffload = 2 * f.KV().HeadCountMax() / cmp.Or(f.KV().HeadCountKVMin(), 1) * kvTotal / 6
 	}

 	return
@@ -728,20 +648,6 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
 		graphSize = 4 * (imageSize*imageSize*numChannels +
 			embeddingLength*patchSize +
 			numPatches*numPatches*headCount)
-	case "qwen25vl":
-		maxPixels := uint64(llm.KV().Uint("vision.max_pixels", 28*28*1280))
-
-		numPatches := maxPixels / (patchSize * patchSize)
-
-		graphSize = 4 * (maxPixels*numChannels + // Original image storage
-			// Normalized pixels
-			maxPixels*numChannels +
-			// Patches storage (numPatches * channels * patchSize^2)
-			numPatches*numChannels*patchSize*patchSize +
-			// Self-attention calculations
-			numPatches*numPatches*headCount +
-			// Additional buffer for processing
-			embeddingLength*numPatches)
 	case "llama4":
 		// vision graph is computed independently in the same schedule
 		// and is negligible compared to the worst case text graph
@@ -752,11 +658,6 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {

 // SupportsKVCacheType checks if the requested cache type is supported
 func (f GGML) SupportsKVCacheType(cacheType string) bool {
-	if arch := f.KV().Architecture(); slices.Contains([]string{"gptoss", "gpt-oss"}, arch) {
-		// gpt-oss uses attention with sinks which does not support quantized cache types
-		slog.Warn("model only supports non-quantized cache types ", "mode", arch)
-		return cacheType == "f16"
-	}
 	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
 }

--- a/fs/ggml/ggml_test.go
+++ b/fs/ggml/ggml_test.go
@@ -269,33 +269,3 @@ func TestKeyValue(t *testing.T) {
 		t.Errorf("unexpected uint8s (-got +want):\n%s", diff)
 	}
 }
-
-func TestHeadCount(t *testing.T) {
-	valuesArray := []int32{1, 5, 3, 4}
-	cases := []struct {
-		kv   KV
-		want uint64
-	}{
-		{
-			kv: KV{
-				"general.architecture":     "abc",
-				"abc.attention.head_count": &array[int32]{values: valuesArray, size: len(valuesArray)},
-			},
-			want: uint64(5),
-		},
-		{
-			kv: KV{
-				"general.architecture":     "abc",
-				"abc.attention.head_count": uint32(3),
-			},
-			want: uint64(3),
-		},
-	}
-
-	for _, tt := range cases {
-		got := tt.kv.HeadCountMax()
-		if got != tt.want {
-			t.Errorf("unexpected max value: got=%d want=%d", got, tt.want)
-		}
-	}
-}
--- a/fs/ggml/gguf.go
+++ b/fs/ggml/gguf.go
@@ -527,17 +527,23 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		return err
 	}

-	for _, key := range slices.Sorted(maps.Keys(kv)) {
+	keys := slices.Collect(maps.Keys(kv))
+	slices.Sort(keys)
+
+	for _, key := range keys {
 		if err := ggufWriteKV(f, key, kv[key]); err != nil {
 			return err
 		}
 	}

 	slices.SortStableFunc(ts, func(a, b *Tensor) int {
-		if i, j := a.block(), b.block(); i > 0 && j > 0 {
+		if i, j := a.block(), b.block(); i < 0 && j > 0 {
+			return 1
+		} else if i > 0 && j < 0 {
+			return -1
+		} else {
 			return cmp.Compare(i, j)
 		}
-		return cmp.Compare(a.Name, b.Name)
 	})

 	var s uint64
@@ -609,10 +615,6 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 		err = writeGGUFArray(ws, ggufTypeString, v)
 	case *array[string]:
 		err = writeGGUFArray(ws, ggufTypeString, v.values)
-	case []bool:
-		err = writeGGUFArray(ws, ggufTypeBool, v)
-	case *array[bool]:
-		err = writeGGUFArray(ws, ggufTypeBool, v.values)
 	default:
 		return fmt.Errorf("improper type for '%s'", k)
 	}
--- a/fs/ggml/gguf_test.go
+++ b/fs/ggml/gguf_test.go
@@ -2,82 +2,62 @@ package ggml

 import (
 	"bytes"
-	"math/rand/v2"
 	"os"
-	"strings"
+	"slices"
 	"testing"

 	"github.com/google/go-cmp/cmp"
 )

 func TestWriteGGUF(t *testing.T) {
-	r := rand.New(rand.NewPCG(0, 0))
-	for range 8 {
-		t.Run("shuffle", func(t *testing.T) {
-			t.Parallel()
+	w, err := os.CreateTemp(t.TempDir(), "*.bin")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer w.Close()

-			ts := []*Tensor{
-				{Name: "token_embd.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.0.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.1.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.2.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.3.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.4.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.5.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "output_norm.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
-				{Name: "output.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
-			}
+	if err := WriteGGUF(w, KV{
+		"general.alignment": uint32(16),
+	}, []*Tensor{
+		{Name: "test.0", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.1", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.2", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.3", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.4", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.5", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+	}); err != nil {
+		t.Fatal(err)
+	}

-			r.Shuffle(len(ts), func(i, j int) {
-				ts[i], ts[j] = ts[j], ts[i]
-			})
+	r, err := os.Open(w.Name())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer r.Close()

-			w, err := os.CreateTemp(t.TempDir(), strings.ReplaceAll(t.Name(), "/", "_")+"*.bin")
-			if err != nil {
-				t.Fatal(err)
-			}
-			defer w.Close()
+	ff, _, err := Decode(r, 0)
+	if err != nil {
+		t.Fatal(err)
+	}

-			if err := WriteGGUF(w, KV{
-				"general.alignment": uint32(16),
-			}, ts); err != nil {
-				t.Fatal(err)
-			}
+	if diff := cmp.Diff(ff.KV(), KV{
+		"general.alignment":       uint32(16),
+		"general.parameter_count": uint64(36),
+	}); diff != "" {
+		t.Errorf("Mismatch (-want +got):\n%s", diff)
+	}

-			r, err := os.Open(w.Name())
-			if err != nil {
-				t.Fatal(err)
-			}
-			defer r.Close()
-
-			ff, err := Decode(r, 0)
-			if err != nil {
-				t.Fatal(err)
-			}
-
-			if diff := cmp.Diff(KV{
-				"general.alignment":       uint32(16),
-				"general.parameter_count": uint64(54),
-			}, ff.KV()); diff != "" {
-				t.Errorf("Mismatch (-want +got):\n%s", diff)
-			}
-
-			if diff := cmp.Diff(Tensors{
-				Offset: 608,
-				items: []*Tensor{
-					{Name: "blk.0.attn_norm.weight", Offset: 0, Shape: []uint64{2, 3}},
-					{Name: "blk.1.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},
-					{Name: "blk.2.attn_norm.weight", Offset: 64, Shape: []uint64{2, 3}},
-					{Name: "blk.3.attn_norm.weight", Offset: 96, Shape: []uint64{2, 3}},
-					{Name: "blk.4.attn_norm.weight", Offset: 128, Shape: []uint64{2, 3}},
-					{Name: "blk.5.attn_norm.weight", Offset: 160, Shape: []uint64{2, 3}},
-					{Name: "output.weight", Offset: 192, Shape: []uint64{3, 2}},
-					{Name: "output_norm.weight", Offset: 224, Shape: []uint64{3, 2}},
-					{Name: "token_embd.weight", Offset: 256, Shape: []uint64{2, 3}},
-				},
-			}, ff.Tensors(), cmp.AllowUnexported(Tensors{})); diff != "" {
-				t.Errorf("Mismatch (-want +got):\n%s", diff)
-			}
-		})
+	if diff := cmp.Diff(ff.Tensors(), Tensors{
+		Offset: 336,
+		items: []*Tensor{
+			{Name: "test.0", Offset: 0, Shape: []uint64{2, 3}},
+			{Name: "test.1", Offset: 32, Shape: []uint64{2, 3}},
+			{Name: "test.2", Offset: 64, Shape: []uint64{2, 3}},
+			{Name: "test.3", Offset: 96, Shape: []uint64{2, 3}},
+			{Name: "test.4", Offset: 128, Shape: []uint64{2, 3}},
+			{Name: "test.5", Offset: 160, Shape: []uint64{2, 3}},
+		},
+	}, cmp.AllowUnexported(Tensors{})); diff != "" {
+		t.Errorf("Mismatch (-want +got):\n%s", diff)
 	}
 }
--- a/fs/ggml/type.go
+++ b/fs/ggml/type.go
@@ -12,42 +12,42 @@ type FileType uint32
 const (
 	FileTypeF32 FileType = iota
 	FileTypeF16
-	fileTypeQ4_0
-	fileTypeQ4_1
-	fileTypeMXFP4 // originally fileTypeQ4_1_F16 // unused by GGML
-	fileTypeQ4_2  // unused by GGML
-	fileTypeQ4_3  // unused by GGML
+	FileTypeQ4_0
+	FileTypeQ4_1
+	fileTypeQ4_1_F16 // unused by GGML
+	fileTypeQ4_2     // unused by GGML
+	fileTypeQ4_3     // unused by GGML
 	FileTypeQ8_0
-	fileTypeQ5_0
-	fileTypeQ5_1
-	fileTypeQ2_K
-	fileTypeQ3_K_S
-	fileTypeQ3_K_M
-	fileTypeQ3_K_L
+	FileTypeQ5_0
+	FileTypeQ5_1
+	FileTypeQ2_K
+	FileTypeQ3_K_S
+	FileTypeQ3_K_M
+	FileTypeQ3_K_L
 	FileTypeQ4_K_S
 	FileTypeQ4_K_M
-	fileTypeQ5_K_S
-	fileTypeQ5_K_M
-	fileTypeQ6_K
-	fileTypeIQ2_XXS
-	fileTypeIQ2_XS
-	fileTypeQ2_K_S
-	fileTypeIQ3_XS
-	fileTypeIQ3_XXS
-	fileTypeIQ1_S
-	fileTypeIQ4_NL
-	fileTypeIQ3_S
-	fileTypeIQ3_M
-	fileTypeIQ2_S
-	fileTypeIQ2_M
-	fileTypeIQ4_XS
-	fileTypeIQ1_M
+	FileTypeQ5_K_S
+	FileTypeQ5_K_M
+	FileTypeQ6_K
+	fileTypeIQ2_XXS // not supported by ollama
+	fileTypeIQ2_XS  // not supported by ollama
+	FileTypeQ2_K_S
+	fileTypeIQ3_XS  // not supported by ollama
+	fileTypeIQ3_XXS // not supported by ollama
+	fileTypeIQ1_S   // not supported by ollama
+	fileTypeIQ4_NL  // not supported by ollama
+	fileTypeIQ3_S   // not supported by ollama
+	fileTypeIQ3_M   // not supported by ollama
+	fileTypeIQ2_S   // not supported by ollama
+	fileTypeIQ2_M   // not supported by ollama
+	fileTypeIQ4_XS  // not supported by ollama
+	fileTypeIQ1_M   // not supported by ollama
 	FileTypeBF16
 	fileTypeQ4_0_4_4 // unused by GGML
 	fileTypeQ4_0_4_8 // unused by GGML
 	fileTypeQ4_0_8_8 // unused by GGML
-	fileTypeTQ1_0
-	fileTypeTQ2_0
+	fileTypeTQ1_0    // not supported by ollama
+	fileTypeTQ2_0    // not supported by ollama

 	FileTypeUnknown = 1024
 )
@@ -60,12 +60,36 @@ func ParseFileType(s string) (FileType, error) {
 		return FileTypeF32, nil
 	case "F16":
 		return FileTypeF16, nil
+	case "Q4_0":
+		return FileTypeQ4_0, nil
+	case "Q4_1":
+		return FileTypeQ4_1, nil
 	case "Q8_0":
 		return FileTypeQ8_0, nil
+	case "Q5_0":
+		return FileTypeQ5_0, nil
+	case "Q5_1":
+		return FileTypeQ5_1, nil
+	case "Q2_K":
+		return FileTypeQ2_K, nil
+	case "Q3_K_S":
+		return FileTypeQ3_K_S, nil
+	case "Q3_K_M":
+		return FileTypeQ3_K_M, nil
+	case "Q3_K_L":
+		return FileTypeQ3_K_L, nil
 	case "Q4_K_S":
 		return FileTypeQ4_K_S, nil
 	case "Q4_K_M", "Q4_K":
 		return FileTypeQ4_K_M, nil
+	case "Q5_K_S":
+		return FileTypeQ5_K_S, nil
+	case "Q5_K_M", "Q5_K":
+		return FileTypeQ5_K_M, nil
+	case "Q6_K":
+		return FileTypeQ6_K, nil
+	case "Q2_K_S":
+		return FileTypeQ2_K_S, nil
 	case "BF16":
 		return FileTypeBF16, nil
 	default:
@@ -87,43 +111,40 @@ func ParseFileType(s string) (FileType, error) {
 }

 func (t FileType) String() string {
-	// Note: this routine will return a broader set of file types for existing models
 	switch t {
 	case FileTypeF32:
 		return "F32"
 	case FileTypeF16:
 		return "F16"
-	case fileTypeQ4_0:
+	case FileTypeQ4_0:
 		return "Q4_0"
-	case fileTypeQ4_1:
+	case FileTypeQ4_1:
 		return "Q4_1"
-	case fileTypeMXFP4:
-		return "MXFP4"
 	case FileTypeQ8_0:
 		return "Q8_0"
-	case fileTypeQ5_0:
+	case FileTypeQ5_0:
 		return "Q5_0"
-	case fileTypeQ5_1:
+	case FileTypeQ5_1:
 		return "Q5_1"
-	case fileTypeQ2_K:
+	case FileTypeQ2_K:
 		return "Q2_K"
-	case fileTypeQ3_K_S:
+	case FileTypeQ3_K_S:
 		return "Q3_K_S"
-	case fileTypeQ3_K_M:
+	case FileTypeQ3_K_M:
 		return "Q3_K_M"
-	case fileTypeQ3_K_L:
+	case FileTypeQ3_K_L:
 		return "Q3_K_L"
 	case FileTypeQ4_K_S:
 		return "Q4_K_S"
 	case FileTypeQ4_K_M:
 		return "Q4_K_M"
-	case fileTypeQ5_K_S:
+	case FileTypeQ5_K_S:
 		return "Q5_K_S"
-	case fileTypeQ5_K_M:
+	case FileTypeQ5_K_M:
 		return "Q5_K_M"
-	case fileTypeQ6_K:
+	case FileTypeQ6_K:
 		return "Q6_K"
-	case fileTypeQ2_K_S:
+	case FileTypeQ2_K_S:
 		return "Q2_K_S"
 	case FileTypeBF16:
 		return "BF16"
@@ -142,37 +163,35 @@ func (ftype FileType) ToTensorType() TensorType {
 		return TensorTypeF32
 	case FileTypeF16:
 		return TensorTypeF16
-	case fileTypeQ4_0:
+	case FileTypeQ4_0:
 		return TensorTypeQ4_0
-	case fileTypeQ4_1:
+	case FileTypeQ4_1:
 		return TensorTypeQ4_1
-	case fileTypeMXFP4:
-		return TensorTypeMXFP4 // Formerly unused tensorTypeQ4_2
 	case FileTypeQ8_0:
 		return TensorTypeQ8_0
-	case fileTypeQ5_0:
+	case FileTypeQ5_0:
 		return TensorTypeQ5_0
-	case fileTypeQ5_1:
+	case FileTypeQ5_1:
 		return TensorTypeQ5_1
-	case fileTypeQ2_K:
+	case FileTypeQ2_K:
 		return TensorTypeQ2_K
-	case fileTypeQ3_K_S:
+	case FileTypeQ3_K_S:
 		return TensorTypeQ3_K
-	case fileTypeQ3_K_M:
+	case FileTypeQ3_K_M:
 		return TensorTypeQ3_K
-	case fileTypeQ3_K_L:
+	case FileTypeQ3_K_L:
 		return TensorTypeQ3_K
 	case FileTypeQ4_K_S:
 		return TensorTypeQ4_K
 	case FileTypeQ4_K_M:
 		return TensorTypeQ4_K
-	case fileTypeQ5_K_S:
+	case FileTypeQ5_K_S:
 		return TensorTypeQ5_K
-	case fileTypeQ5_K_M:
+	case FileTypeQ5_K_M:
 		return TensorTypeQ5_K
-	case fileTypeQ6_K:
+	case FileTypeQ6_K:
 		return TensorTypeQ6_K
-	case fileTypeQ2_K_S:
+	case FileTypeQ2_K_S:
 		return TensorTypeQ2_K
 	case FileTypeBF16:
 		return TensorTypeBF16
@@ -191,8 +210,8 @@ const (
 	TensorTypeF16
 	TensorTypeQ4_0
 	TensorTypeQ4_1
-	TensorTypeMXFP4 // Formerly unused tensorTypeQ4_2
-	tensorTypeQ4_3  // unused by GGML
+	tensorTypeQ4_2 // unused by GGML
+	tensorTypeQ4_3 // unused by GGML
 	TensorTypeQ5_0
 	TensorTypeQ5_1
 	TensorTypeQ8_0
@@ -264,8 +283,6 @@ func ParseTensorType(s string) (TensorType, error) {
 		return TensorTypeF64, nil
 	case "BF16":
 		return TensorTypeBF16, nil
-	case "MXFP4":
-		return TensorTypeMXFP4, nil
 	default:
 		return 0, fmt.Errorf("unsupported quantization type %s", s)
 	}
@@ -318,8 +335,6 @@ func (t TensorType) String() string {
 		return "F64"
 	case TensorTypeBF16:
 		return "BF16"
-	case TensorTypeMXFP4:
-		return "MXFP4"
 	default:
 		return "unknown"
 	}
--- a/fs/gguf/gguf.go
+++ b/fs/gguf/gguf.go
@@ -1,347 +0,0 @@
-package gguf
-
-import (
-	"bytes"
-	"cmp"
-	"encoding/binary"
-	"errors"
-	"fmt"
-	"io"
-	"iter"
-	"os"
-	"slices"
-	"strings"
-)
-
-const (
-	typeUint8 uint32 = iota
-	typeInt8
-	typeUint16
-	typeInt16
-	typeUint32
-	typeInt32
-	typeFloat32
-	typeBool
-	typeString
-	typeArray
-	typeUint64
-	typeInt64
-	typeFloat64
-)
-
-var ErrUnsupported = errors.New("unsupported")
-
-type File struct {
-	Magic   [4]byte
-	Version uint32
-
-	keyValues *lazy[KeyValue]
-	tensors   *lazy[TensorInfo]
-	offset    int64
-
-	file   *os.File
-	reader *bufferedReader
-	bts    []byte
-}
-
-func Open(path string) (f *File, err error) {
-	f = &File{bts: make([]byte, 4096)}
-	f.file, err = os.Open(path)
-	if err != nil {
-		return nil, err
-	}
-
-	f.reader = newBufferedReader(f.file, 32<<10)
-
-	if err := binary.Read(f.reader, binary.LittleEndian, &f.Magic); err != nil {
-		return nil, err
-	}
-
-	if bytes.Equal(f.Magic[:], []byte("gguf")) {
-		return nil, fmt.Errorf("%w file type %v", ErrUnsupported, f.Magic)
-	}
-
-	if err := binary.Read(f.reader, binary.LittleEndian, &f.Version); err != nil {
-		return nil, err
-	}
-
-	if f.Version < 2 {
-		return nil, fmt.Errorf("%w version %v", ErrUnsupported, f.Version)
-	}
-
-	f.tensors, err = newLazy(f, f.readTensor)
-	if err != nil {
-		return nil, err
-	}
-
-	f.tensors.successFunc = func() error {
-		offset := f.reader.offset
-
-		alignment := cmp.Or(f.KeyValue("general.alignment").Int(), 32)
-		f.offset = offset + (alignment-offset%alignment)%alignment
-		return nil
-	}
-
-	f.keyValues, err = newLazy(f, f.readKeyValue)
-	if err != nil {
-		return nil, err
-	}
-
-	return f, nil
-}
-
-func (f *File) readTensor() (TensorInfo, error) {
-	name, err := readString(f)
-	if err != nil {
-		return TensorInfo{}, err
-	}
-
-	dims, err := read[uint32](f)
-	if err != nil {
-		return TensorInfo{}, err
-	}
-
-	shape := make([]uint64, dims)
-	for i := range dims {
-		shape[i], err = read[uint64](f)
-		if err != nil {
-			return TensorInfo{}, err
-		}
-	}
-
-	type_, err := read[uint32](f)
-	if err != nil {
-		return TensorInfo{}, err
-	}
-
-	offset, err := read[uint64](f)
-	if err != nil {
-		return TensorInfo{}, err
-	}
-
-	return TensorInfo{
-		Name:   name,
-		Offset: offset,
-		Shape:  shape,
-		Type:   TensorType(type_),
-	}, nil
-}
-
-func (f *File) readKeyValue() (KeyValue, error) {
-	key, err := readString(f)
-	if err != nil {
-		return KeyValue{}, err
-	}
-
-	t, err := read[uint32](f)
-	if err != nil {
-		return KeyValue{}, err
-	}
-
-	value, err := func() (any, error) {
-		switch t {
-		case typeUint8:
-			return read[uint8](f)
-		case typeInt8:
-			return read[int8](f)
-		case typeUint16:
-			return read[uint16](f)
-		case typeInt16:
-			return read[int16](f)
-		case typeUint32:
-			return read[uint32](f)
-		case typeInt32:
-			return read[int32](f)
-		case typeUint64:
-			return read[uint64](f)
-		case typeInt64:
-			return read[int64](f)
-		case typeFloat32:
-			return read[float32](f)
-		case typeFloat64:
-			return read[float64](f)
-		case typeBool:
-			return read[bool](f)
-		case typeString:
-			return readString(f)
-		case typeArray:
-			return readArray(f)
-		default:
-			return nil, fmt.Errorf("%w type %d", ErrUnsupported, t)
-		}
-	}()
-	if err != nil {
-		return KeyValue{}, err
-	}
-
-	return KeyValue{
-		Key:   key,
-		Value: Value{value},
-	}, nil
-}
-
-func read[T any](f *File) (t T, err error) {
-	err = binary.Read(f.reader, binary.LittleEndian, &t)
-	return t, err
-}
-
-func readString(f *File) (string, error) {
-	n, err := read[uint64](f)
-	if err != nil {
-		return "", err
-	}
-
-	if int(n) > len(f.bts) {
-		f.bts = make([]byte, n)
-	}
-
-	bts := f.bts[:n]
-	if _, err := io.ReadFull(f.reader, bts); err != nil {
-		return "", err
-	}
-	defer clear(bts)
-
-	return string(bts), nil
-}
-
-func readArray(f *File) (any, error) {
-	t, err := read[uint32](f)
-	if err != nil {
-		return nil, err
-	}
-
-	n, err := read[uint64](f)
-	if err != nil {
-		return nil, err
-	}
-
-	switch t {
-	case typeUint8:
-		return readArrayData[uint8](f, n)
-	case typeInt8:
-		return readArrayData[int8](f, n)
-	case typeUint16:
-		return readArrayData[uint16](f, n)
-	case typeInt16:
-		return readArrayData[int16](f, n)
-	case typeUint32:
-		return readArrayData[uint32](f, n)
-	case typeInt32:
-		return readArrayData[int32](f, n)
-	case typeUint64:
-		return readArrayData[uint64](f, n)
-	case typeInt64:
-		return readArrayData[int64](f, n)
-	case typeFloat32:
-		return readArrayData[float32](f, n)
-	case typeFloat64:
-		return readArrayData[float64](f, n)
-	case typeBool:
-		return readArrayData[bool](f, n)
-	case typeString:
-		return readArrayString(f, n)
-	default:
-		return nil, fmt.Errorf("%w type %d", ErrUnsupported, t)
-	}
-}
-
-func readArrayData[T any](f *File, n uint64) (s []T, err error) {
-	s = make([]T, n)
-	for i := range n {
-		e, err := read[T](f)
-		if err != nil {
-			return nil, err
-		}
-
-		s[i] = e
-	}
-
-	return s, nil
-}
-
-func readArrayString(f *File, n uint64) (s []string, err error) {
-	s = make([]string, n)
-	for i := range n {
-		e, err := readString(f)
-		if err != nil {
-			return nil, err
-		}
-
-		s[i] = e
-	}
-
-	return s, nil
-}
-
-func (f *File) Close() error {
-	f.keyValues.stop()
-	f.tensors.stop()
-	return f.file.Close()
-}
-
-func (f *File) KeyValue(key string) KeyValue {
-	if !strings.HasPrefix(key, "general.") && !strings.HasPrefix(key, "tokenizer.") {
-		key = f.KeyValue("general.architecture").String() + "." + key
-	}
-
-	if index := slices.IndexFunc(f.keyValues.values, func(kv KeyValue) bool {
-		return kv.Key == key
-	}); index >= 0 {
-		return f.keyValues.values[index]
-	}
-
-	for keyValue, ok := f.keyValues.next(); ok; keyValue, ok = f.keyValues.next() {
-		if keyValue.Key == key {
-			return keyValue
-		}
-	}
-
-	return KeyValue{}
-}
-
-func (f *File) NumKeyValues() int {
-	return int(f.keyValues.count)
-}
-
-func (f *File) KeyValues() iter.Seq2[int, KeyValue] {
-	return f.keyValues.All()
-}
-
-func (f *File) TensorInfo(name string) TensorInfo {
-	if index := slices.IndexFunc(f.tensors.values, func(t TensorInfo) bool {
-		return t.Name == name
-	}); index >= 0 {
-		return f.tensors.values[index]
-	}
-
-	// fast-forward through key values if we haven't already
-	_ = f.keyValues.rest()
-	for tensor, ok := f.tensors.next(); ok; tensor, ok = f.tensors.next() {
-		if tensor.Name == name {
-			return tensor
-		}
-	}
-
-	return TensorInfo{}
-}
-
-func (f *File) NumTensors() int {
-	return int(f.tensors.count)
-}
-
-func (f *File) TensorInfos() iter.Seq2[int, TensorInfo] {
-	// fast forward through key values if we haven't already
-	f.keyValues.rest()
-	return f.tensors.All()
-}
-
-func (f *File) TensorReader(name string) (TensorInfo, io.Reader, error) {
-	t := f.TensorInfo(name)
-	if t.NumBytes() == 0 {
-		return TensorInfo{}, nil, fmt.Errorf("tensor %s not found", name)
-	}
-
-	// fast forward through tensor info if we haven't already
-	_ = f.tensors.rest()
-	return t, io.NewSectionReader(f.file, f.offset+int64(t.Offset), t.NumBytes()), nil
-}
--- a/fs/gguf/gguf_test.go
+++ b/fs/gguf/gguf_test.go
@@ -1,249 +0,0 @@
-package gguf_test
-
-import (
-	"bytes"
-	"os"
-	"strconv"
-	"strings"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-	"github.com/google/go-cmp/cmp/cmpopts"
-	"github.com/ollama/ollama/fs/ggml"
-	"github.com/ollama/ollama/fs/gguf"
-)
-
-func createBinFile(tb testing.TB) string {
-	tb.Helper()
-	f, err := os.CreateTemp(tb.TempDir(), "")
-	if err != nil {
-		tb.Fatal(err)
-	}
-	defer f.Close()
-
-	kv := ggml.KV{
-		"general.architecture":                   "llama",
-		"llama.block_count":                      uint32(8),
-		"llama.embedding_length":                 uint32(3),
-		"llama.attention.head_count":             uint32(2),
-		"llama.attention.head_count_kv":          uint32(2),
-		"llama.attention.key_length":             uint32(3),
-		"llama.rope.dimension_count":             uint32(4),
-		"llama.rope.freq_base":                   float32(10000.0),
-		"llama.rope.freq_scale":                  float32(1.0),
-		"llama.attention.layer_norm_rms_epsilon": float32(1e-6),
-		"tokenizer.ggml.eos_token_id":            uint32(0),
-		"tokenizer.ggml.eos_token_ids":           []int32{1, 2, 3},
-		"tokenizer.ggml.tokens":                  []string{"hello", "world"},
-		"tokenizer.ggml.scores":                  []float32{0, 1},
-	}
-
-	tensors := []*ggml.Tensor{
-		{
-			Name:     "token_embd.weight",
-			Kind:     0,
-			Shape:    []uint64{2, 3},
-			WriterTo: bytes.NewBuffer(make([]byte, 4*2*3)),
-		},
-		{
-			Name:     "output.weight",
-			Kind:     0,
-			Shape:    []uint64{3, 2},
-			WriterTo: bytes.NewBuffer(make([]byte, 4*3*2)),
-		},
-	}
-
-	for i := range 8 {
-		tensors = append(tensors, &ggml.Tensor{
-			Name:     "blk." + strconv.Itoa(i) + ".attn_q.weight",
-			Kind:     0,
-			Shape:    []uint64{3, 3},
-			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
-		}, &ggml.Tensor{
-			Name:     "blk." + strconv.Itoa(i) + ".attn_k.weight",
-			Kind:     0,
-			Shape:    []uint64{3, 3},
-			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
-		}, &ggml.Tensor{
-			Name:     "blk." + strconv.Itoa(i) + ".attn_v.weight",
-			Kind:     0,
-			Shape:    []uint64{3, 3},
-			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
-		}, &ggml.Tensor{
-			Name:     "blk." + strconv.Itoa(i) + ".attn_output.weight",
-			Kind:     0,
-			Shape:    []uint64{3, 3},
-			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
-		})
-	}
-
-	if err := ggml.WriteGGUF(f, kv, tensors); err != nil {
-		tb.Fatal(err)
-	}
-
-	return f.Name()
-}
-
-func TestRead(t *testing.T) {
-	f, err := gguf.Open(createBinFile(t))
-	if err != nil {
-		t.Fatal(err)
-	}
-	defer f.Close()
-
-	if got := f.KeyValue("does.not.exist").Valid(); got {
-		t.Errorf(`KeyValue("does.not.exist").Exists() = %v, want false`, got)
-	}
-
-	if got := f.KeyValue("general.architecture").String(); got != "llama" {
-		t.Errorf(`KeyValue("general.architecture").String() = %q, want %q`, got, "llama")
-	}
-
-	if got := f.TensorInfo("token_embd.weight"); got.Name != "token_embd.weight" {
-		t.Errorf(`TensorInfo("token_embd.weight").Name = %q, want %q`, got.Name, "token_embd.weight")
-	} else if diff := cmp.Diff(got.Shape, []uint64{2, 3}); diff != "" {
-		t.Errorf(`TensorInfo("token_embd.weight").Shape mismatch (-got +want):\n%s`, diff)
-	} else if got.Type != gguf.TensorTypeF32 {
-		t.Errorf(`TensorInfo("token_embd.weight").Type = %d, want %d`, got.Type, gguf.TensorTypeF32)
-	}
-
-	if got := f.KeyValue("block_count").Uint(); got != 8 {
-		t.Errorf(`KeyValue("block_count").Uint() = %d, want %d`, got, 8)
-	}
-
-	if diff := cmp.Diff(f.KeyValue("tokenizer.ggml.tokens").Strings(), []string{"hello", "world"}); diff != "" {
-		t.Errorf("KeyValue(\"tokenizer.ggml.tokens\").Strings() mismatch (-got +want):\n%s", diff)
-	}
-
-	if diff := cmp.Diff(f.KeyValue("tokenizer.ggml.scores").Floats(), []float64{0, 1}); diff != "" {
-		t.Errorf("KeyValue(\"tokenizer.ggml.scores\").Ints() mismatch (-got +want):\n%s", diff)
-	}
-
-	var kvs []string
-	for _, kv := range f.KeyValues() {
-		if !kv.Valid() {
-			t.Error("found invalid key-value pair:", kv)
-		}
-
-		kvs = append(kvs, kv.Key)
-	}
-
-	if len(kvs) != f.NumKeyValues() {
-		t.Errorf("iterated key count = %d, want %d", len(kvs), f.NumKeyValues())
-	}
-
-	if diff := cmp.Diff(kvs, []string{
-		"general.architecture",
-		"llama.block_count",
-		"llama.embedding_length",
-		"llama.attention.head_count",
-		"llama.attention.head_count_kv",
-		"llama.attention.key_length",
-		"llama.rope.dimension_count",
-		"llama.rope.freq_base",
-		"llama.rope.freq_scale",
-		"llama.attention.layer_norm_rms_epsilon",
-		"tokenizer.ggml.eos_token_id",
-		"tokenizer.ggml.eos_token_ids",
-		"tokenizer.ggml.tokens",
-		"tokenizer.ggml.scores",
-	}, cmpopts.SortSlices(strings.Compare)); diff != "" {
-		t.Errorf("KeyValues() mismatch (-got +want):\n%s", diff)
-	}
-
-	var tis []string
-	for _, ti := range f.TensorInfos() {
-		if !ti.Valid() {
-			t.Error("found invalid tensor info:", ti)
-		}
-
-		tis = append(tis, ti.Name)
-	}
-
-	if len(tis) != f.NumTensors() {
-		t.Errorf("iterated tensor count = %d, want %d", len(tis), f.NumTensors())
-	}
-
-	if diff := cmp.Diff(tis, []string{
-		"token_embd.weight",
-		"output.weight",
-		"blk.0.attn_q.weight",
-		"blk.0.attn_k.weight",
-		"blk.0.attn_v.weight",
-		"blk.0.attn_output.weight",
-		"blk.1.attn_q.weight",
-		"blk.1.attn_k.weight",
-		"blk.1.attn_v.weight",
-		"blk.1.attn_output.weight",
-		"blk.2.attn_q.weight",
-		"blk.2.attn_k.weight",
-		"blk.2.attn_v.weight",
-		"blk.2.attn_output.weight",
-		"blk.3.attn_q.weight",
-		"blk.3.attn_k.weight",
-		"blk.3.attn_v.weight",
-		"blk.3.attn_output.weight",
-		"blk.4.attn_q.weight",
-		"blk.4.attn_k.weight",
-		"blk.4.attn_v.weight",
-		"blk.4.attn_output.weight",
-		"blk.5.attn_q.weight",
-		"blk.5.attn_k.weight",
-		"blk.5.attn_v.weight",
-		"blk.5.attn_output.weight",
-		"blk.6.attn_q.weight",
-		"blk.6.attn_k.weight",
-		"blk.6.attn_v.weight",
-		"blk.6.attn_output.weight",
-		"blk.7.attn_q.weight",
-		"blk.7.attn_k.weight",
-		"blk.7.attn_v.weight",
-		"blk.7.attn_output.weight",
-	}, cmpopts.SortSlices(strings.Compare)); diff != "" {
-		t.Errorf("TensorInfos() mismatch (-got +want):\n%s", diff)
-	}
-
-	ti, r, err := f.TensorReader("output.weight")
-	if err != nil {
-		t.Fatalf(`TensorReader("output.weight") error: %v`, err)
-	}
-
-	if ti.Name != "output.weight" {
-		t.Errorf(`TensorReader("output.weight").Name = %q, want %q`, ti.Name, "output.weight")
-	} else if diff := cmp.Diff(ti.Shape, []uint64{3, 2}); diff != "" {
-		t.Errorf(`TensorReader("output.weight").Shape mismatch (-got +want):\n%s`, diff)
-	} else if ti.Type != gguf.TensorTypeF32 {
-		t.Errorf(`TensorReader("output.weight").Type = %d, want %d`, ti.Type, gguf.TensorTypeF32)
-	}
-
-	var b bytes.Buffer
-	if _, err := b.ReadFrom(r); err != nil {
-		t.Fatalf(`ReadFrom TensorReader("output.weight") error: %v`, err)
-	}
-
-	if b.Len() != int(ti.NumBytes()) {
-		t.Errorf(`ReadFrom TensorReader("output.weight") length = %d, want %d`, b.Len(), ti.NumBytes())
-	}
-}
-
-func BenchmarkRead(b *testing.B) {
-	b.ReportAllocs()
-
-	p := createBinFile(b)
-	for b.Loop() {
-		f, err := gguf.Open(p)
-		if err != nil {
-			b.Fatal(err)
-		}
-
-		if got := f.KeyValue("general.architecture").String(); got != "llama" {
-			b.Errorf("got = %q, want %q", got, "llama")
-		}
-
-		// Iterate through some tensors
-		for range f.TensorInfos() {
-		}
-
-		f.Close()
-	}
-}
--- a/fs/gguf/keyvalue.go
+++ b/fs/gguf/keyvalue.go
@@ -1,90 +0,0 @@
-package gguf
-
-import (
-	"reflect"
-	"slices"
-)
-
-type KeyValue struct {
-	Key string
-	Value
-}
-
-func (kv KeyValue) Valid() bool {
-	return kv.Key != "" && kv.Value.value != nil
-}
-
-type Value struct {
-	value any
-}
-
-func value[T any](v Value, kinds ...reflect.Kind) (t T) {
-	vv := reflect.ValueOf(v.value)
-	if slices.Contains(kinds, vv.Kind()) {
-		t = vv.Convert(reflect.TypeOf(t)).Interface().(T)
-	}
-	return
-}
-
-func values[T any](v Value, kinds ...reflect.Kind) (ts []T) {
-	switch vv := reflect.ValueOf(v.value); vv.Kind() {
-	case reflect.Slice:
-		if slices.Contains(kinds, vv.Type().Elem().Kind()) {
-			ts = make([]T, vv.Len())
-			for i := range vv.Len() {
-				ts[i] = vv.Index(i).Convert(reflect.TypeOf(ts[i])).Interface().(T)
-			}
-		}
-	}
-	return
-}
-
-// Int returns Value as a signed integer. If it is not a signed integer, it returns 0.
-func (v Value) Int() int64 {
-	return value[int64](v, reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64)
-}
-
-// Ints returns Value as a signed integer slice. If it is not a signed integer slice, it returns nil.
-func (v Value) Ints() (i64s []int64) {
-	return values[int64](v, reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64)
-}
-
-// Uint converts an unsigned integer value to uint64. If the value is not a unsigned integer, it returns 0.
-func (v Value) Uint() uint64 {
-	return value[uint64](v, reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64)
-}
-
-// Uints returns Value as a unsigned integer slice. If it is not a unsigned integer slice, it returns nil.
-func (v Value) Uints() (u64s []uint64) {
-	return values[uint64](v, reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64)
-}
-
-// Float returns Value as a float. If it is not a float, it returns 0.
-func (v Value) Float() float64 {
-	return value[float64](v, reflect.Float32, reflect.Float64)
-}
-
-// Floats returns Value as a float slice. If it is not a float slice, it returns nil.
-func (v Value) Floats() (f64s []float64) {
-	return values[float64](v, reflect.Float32, reflect.Float64)
-}
-
-// Bool returns Value as a boolean. If it is not a boolean, it returns false.
-func (v Value) Bool() bool {
-	return value[bool](v, reflect.Bool)
-}
-
-// Bools returns Value as a boolean slice. If it is not a boolean slice, it returns nil.
-func (v Value) Bools() (bools []bool) {
-	return values[bool](v, reflect.Bool)
-}
-
-// String returns Value as a string. If it is not a string, it returns an empty string.
-func (v Value) String() string {
-	return value[string](v, reflect.String)
-}
-
-// Strings returns Value as a string slice. If it is not a string slice, it returns nil.
-func (v Value) Strings() (strings []string) {
-	return values[string](v, reflect.String)
-}
--- a/fs/gguf/keyvalue_test.go
+++ b/fs/gguf/keyvalue_test.go
@@ -1,208 +0,0 @@
-package gguf
-
-import (
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-)
-
-func split(name string, values map[string][]any) (matched []any, unmatched []any) {
-	for key, value := range values {
-		if key == name {
-			matched = value
-		} else {
-			unmatched = append(unmatched, value...)
-		}
-	}
-	return
-}
-
-func TestValue(t *testing.T) {
-	values := map[string][]any{
-		"int64":   {int(42), int8(42), int16(42), int32(42), int64(42)},
-		"uint64":  {uint(42), uint8(42), uint16(42), uint32(42), uint64(42)},
-		"float64": {float32(42), float64(42)},
-		"string":  {"42", "hello"},
-		"bool":    {true, false},
-	}
-
-	t.Run("int64", func(t *testing.T) {
-		matched, unmatched := split("int64", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if i64 := kv.Int(); i64 != 42 {
-				t.Errorf("expected 42, got %d", i64)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if i64 := kv.Int(); i64 != 0 {
-				t.Errorf("expected 42, got %d", i64)
-			}
-		}
-	})
-
-	t.Run("uint64", func(t *testing.T) {
-		matched, unmatched := split("uint64", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if u64 := kv.Uint(); u64 != 42 {
-				t.Errorf("expected 42, got %d", u64)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if u64 := kv.Uint(); u64 != 0 {
-				t.Errorf("expected 42, got %d", u64)
-			}
-		}
-	})
-
-	t.Run("float64", func(t *testing.T) {
-		matched, unmatched := split("float64", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if f64 := kv.Float(); f64 != 42 {
-				t.Errorf("expected 42, got %f", f64)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if f64 := kv.Float(); f64 != 0 {
-				t.Errorf("expected 42, got %f", f64)
-			}
-		}
-	})
-
-	t.Run("string", func(t *testing.T) {
-		matched, unmatched := split("string", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if s := kv.String(); s != v {
-				t.Errorf("expected 42, got %s", s)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if s := kv.String(); s != "" {
-				t.Errorf("expected 42, got %s", s)
-			}
-		}
-	})
-
-	t.Run("bool", func(t *testing.T) {
-		matched, unmatched := split("bool", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if b := kv.Bool(); b != v {
-				t.Errorf("expected true, got %v", b)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if b := kv.Bool(); b != false {
-				t.Errorf("expected false, got %v", b)
-			}
-		}
-	})
-}
-
-func TestValues(t *testing.T) {
-	values := map[string][]any{
-		"int64s":   {[]int{42}, []int8{42}, []int16{42}, []int32{42}, []int64{42}},
-		"uint64s":  {[]uint{42}, []uint8{42}, []uint16{42}, []uint32{42}, []uint64{42}},
-		"float64s": {[]float32{42}, []float64{42}},
-		"strings":  {[]string{"42"}, []string{"hello"}},
-		"bools":    {[]bool{true}, []bool{false}},
-	}
-
-	t.Run("int64s", func(t *testing.T) {
-		matched, unmatched := split("int64s", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if diff := cmp.Diff(kv.Ints(), []int64{42}); diff != "" {
-				t.Errorf("diff: %s", diff)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if i64s := kv.Ints(); i64s != nil {
-				t.Errorf("expected nil, got %v", i64s)
-			}
-		}
-	})
-
-	t.Run("uint64s", func(t *testing.T) {
-		matched, unmatched := split("uint64s", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if diff := cmp.Diff(kv.Uints(), []uint64{42}); diff != "" {
-				t.Errorf("diff: %s", diff)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if u64s := kv.Uints(); u64s != nil {
-				t.Errorf("expected nil, got %v", u64s)
-			}
-		}
-	})
-
-	t.Run("float64s", func(t *testing.T) {
-		matched, unmatched := split("float64s", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if diff := cmp.Diff(kv.Floats(), []float64{42}); diff != "" {
-				t.Errorf("diff: %s", diff)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if f64s := kv.Floats(); f64s != nil {
-				t.Errorf("expected nil, got %v", f64s)
-			}
-		}
-	})
-
-	t.Run("strings", func(t *testing.T) {
-		matched, unmatched := split("strings", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if diff := cmp.Diff(kv.Strings(), v); diff != "" {
-				t.Errorf("diff: %s", diff)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if s := kv.Strings(); s != nil {
-				t.Errorf("expected nil, got %v", s)
-			}
-		}
-	})
-
-	t.Run("bools", func(t *testing.T) {
-		matched, unmatched := split("bools", values)
-		for _, v := range matched {
-			kv := KeyValue{"key", Value{v}}
-			if diff := cmp.Diff(kv.Bools(), v); diff != "" {
-				t.Errorf("diff: %s", diff)
-			}
-		}
-
-		for _, v := range unmatched {
-			kv := KeyValue{"key", Value{v}}
-			if b := kv.Bools(); b != nil {
-				t.Errorf("expected nil, got %v", b)
-			}
-		}
-	})
-}
--- a/fs/gguf/lazy.go
+++ b/fs/gguf/lazy.go
@@ -1,89 +0,0 @@
-package gguf
-
-import (
-	"encoding/binary"
-	"iter"
-	"log/slog"
-)
-
-type lazy[T any] struct {
-	count  uint64
-	next   func() (T, bool)
-	stop   func()
-	values []T
-
-	// successFunc is called when all values have been successfully read.
-	successFunc func() error
-}
-
-func newLazy[T any](f *File, fn func() (T, error)) (*lazy[T], error) {
-	it := lazy[T]{}
-	if err := binary.Read(f.reader, binary.LittleEndian, &it.count); err != nil {
-		return nil, err
-	}
-
-	it.values = make([]T, 0)
-	it.next, it.stop = iter.Pull(func(yield func(T) bool) {
-		for i := range it.count {
-			t, err := fn()
-			if err != nil {
-				slog.Error("error reading tensor", "index", i, "error", err)
-				return
-			}
-
-			it.values = append(it.values, t)
-			if !yield(t) {
-				break
-			}
-		}
-
-		if it.successFunc != nil {
-			it.successFunc()
-		}
-	})
-
-	return &it, nil
-}
-
-func (g *lazy[T]) Values() iter.Seq[T] {
-	return func(yield func(T) bool) {
-		for _, v := range g.All() {
-			if !yield(v) {
-				break
-			}
-		}
-	}
-}
-
-func (g *lazy[T]) All() iter.Seq2[int, T] {
-	return func(yield func(int, T) bool) {
-		for i := range int(g.count) {
-			if i < len(g.values) {
-				if !yield(i, g.values[i]) {
-					break
-				}
-			} else {
-				t, ok := g.next()
-				if !ok {
-					break
-				}
-
-				if !yield(i, t) {
-					break
-				}
-			}
-		}
-	}
-}
-
-func (g *lazy[T]) rest() (collected bool) {
-	for {
-		_, ok := g.next()
-		collected = collected || ok
-		if !ok {
-			break
-		}
-	}
-
-	return collected
-}
--- a/fs/gguf/reader.go
+++ b/fs/gguf/reader.go
@@ -1,23 +0,0 @@
-package gguf
-
-import (
-	"bufio"
-	"io"
-)
-
-type bufferedReader struct {
-	offset int64
-	*bufio.Reader
-}
-
-func newBufferedReader(rs io.ReadSeeker, size int) *bufferedReader {
-	return &bufferedReader{
-		Reader: bufio.NewReaderSize(rs, size),
-	}
-}
-
-func (rs *bufferedReader) Read(p []byte) (n int, err error) {
-	n, err = rs.Reader.Read(p)
-	rs.offset += int64(n)
-	return n, err
-}
--- a/fs/gguf/tensor.go
+++ b/fs/gguf/tensor.go
@@ -1,288 +0,0 @@
-package gguf
-
-import (
-	"log/slog"
-	"strings"
-)
-
-type TensorInfo struct {
-	Name   string
-	Offset uint64
-	Shape  []uint64
-	Type   TensorType
-}
-
-func (ti TensorInfo) Valid() bool {
-	return ti.Name != "" && ti.NumBytes() > 0
-}
-
-func (ti TensorInfo) NumValues() int64 {
-	var numItems int64 = 1
-	for _, dim := range ti.Shape {
-		numItems *= int64(dim)
-	}
-	return numItems
-}
-
-// NumBytes returns the number of bytes in the tensor.
-func (ti TensorInfo) NumBytes() int64 {
-	return int64(float64(ti.NumValues()) * ti.Type.NumBytes())
-}
-
-func (ti TensorInfo) LogValue() slog.Value {
-	return slog.GroupValue(
-		slog.String("name", ti.Name),
-		slog.Int64("offset", int64(ti.Offset)),
-		slog.Any("shape", ti.Shape),
-		slog.Int64("num_values", ti.NumValues()),
-		slog.Int64("num_bytes", ti.NumBytes()),
-		slog.Any("type", ti.Type),
-	)
-}
-
-type TensorType uint32
-
-const (
-	TensorTypeF32 TensorType = iota
-	TensorTypeF16
-	TensorTypeQ4_0
-	TensorTypeQ4_1
-
-	// unexported // unused in gguf
-	tensorTypeQ4_2
-	tensorTypeQ4_3
-
-	TensorTypeQ5_0
-	TensorTypeQ5_1
-	TensorTypeQ8_0
-	TensorTypeQ8_1
-	TensorTypeQ2_K
-	TensorTypeQ3_K
-	TensorTypeQ4_K
-	TensorTypeQ5_K
-	TensorTypeQ6_K
-	TensorTypeQ8_K
-
-	// unexported // unquantizable by ollama
-	tensorTypeIQ2_XXS
-	tensorTypeIQ2_XS
-	tensorTypeIQ3_XXS
-	tensorTypeIQ1_S
-	tensorTypeIQ4_NL
-	tensorTypeIQ3_S
-	tensorTypeIQ2_S
-	tensorTypeIQ4_XS
-
-	TensorTypeI8
-	TensorTypeI16
-	TensorTypeI32
-	TensorTypeI64
-	TensorTypeF64
-
-	// unexported // unquantizable by ollama
-	tensorTypeIQ1_M
-
-	TensorTypeBF16
-
-	// unexported // unused in gguf
-	tensorTypeQ4_0_4_4
-	tensorTypeQ4_0_4_8
-	tensorTypeQ4_0_8_8
-
-	// unexported // unquantizable by ollama
-	tensorTypeTQ1_0
-	tensorTypeTQ2_0
-
-	// unexported // unused in gguf
-	tensorTypeIQ4_NL_4_4
-	tensorTypeIQ4_NL_4_8
-	tensorTypeIQ4_NL_8_8
-)
-
-func (tt TensorType) NumBytes() float64 {
-	return float64(tt.typeSize()) / float64(tt.blockSize())
-}
-
-func (tt TensorType) typeSize() int64 {
-	switch tt {
-	case TensorTypeF32:
-		return 4
-	case TensorTypeF16:
-		return 2
-	case TensorTypeQ4_0:
-		return 2 + tt.blockSize()/2
-	case TensorTypeQ4_1:
-		return 2 + 2 + tt.blockSize()/2
-	case TensorTypeQ5_0:
-		return 2 + 4 + tt.blockSize()/2
-	case TensorTypeQ5_1:
-		return 2 + 2 + 4 + tt.blockSize()/2
-	case TensorTypeQ8_0:
-		return 2 + tt.blockSize()
-	case TensorTypeQ8_1:
-		return 2 + 2 + tt.blockSize()
-	case TensorTypeQ2_K:
-		return tt.blockSize()/16 + tt.blockSize()/4 + 2 + 2
-	case TensorTypeQ3_K:
-		return tt.blockSize()/8 + tt.blockSize()/4 + 12 + 2
-	case TensorTypeQ4_K:
-		return 2 + 2 + 12 + tt.blockSize()/2
-	case TensorTypeQ5_K:
-		return 2 + 2 + 12 + tt.blockSize()/8 + tt.blockSize()/2
-	case TensorTypeQ6_K:
-		return tt.blockSize()/2 + tt.blockSize()/4 + tt.blockSize()/16 + 2
-	case TensorTypeQ8_K:
-		return 4 + tt.blockSize() + 2*tt.blockSize()/16
-	case tensorTypeIQ2_XXS:
-		return 2 + 2*tt.blockSize()/8
-	case tensorTypeIQ2_XS:
-		return 2 + 2*tt.blockSize()/8 + tt.blockSize()/32
-	case tensorTypeIQ3_XXS:
-		return 2 + tt.blockSize()/4 + tt.blockSize()/8
-	case tensorTypeIQ1_S:
-		return 2 + tt.blockSize()/8 + tt.blockSize()/16
-	case tensorTypeIQ4_NL:
-		return 2 + tt.blockSize()/2
-	case tensorTypeIQ3_S:
-		return 2 + tt.blockSize()/4 + tt.blockSize()/8 + tt.blockSize()/32 + 4
-	case tensorTypeIQ2_S:
-		return 2 + tt.blockSize()/4 + tt.blockSize()/16
-	case tensorTypeIQ4_XS:
-		return 2 + 2 + tt.blockSize()/2 + tt.blockSize()/64
-	case TensorTypeI8:
-		return 1
-	case TensorTypeI16:
-		return 2
-	case TensorTypeI32:
-		return 4
-	case TensorTypeI64:
-		return 8
-	case TensorTypeF64:
-		return 8
-	case tensorTypeIQ1_M:
-		return tt.blockSize()/8 + tt.blockSize()/16 + tt.blockSize()/32
-	case TensorTypeBF16:
-		return 2
-	default:
-		return 0
-	}
-}
-
-func (tt TensorType) blockSize() int64 {
-	switch tt {
-	case TensorTypeF32,
-		TensorTypeF16,
-		TensorTypeI8,
-		TensorTypeI16,
-		TensorTypeI32,
-		TensorTypeI64,
-		TensorTypeF64,
-		TensorTypeBF16:
-		return 1
-	case TensorTypeQ4_0,
-		TensorTypeQ4_1,
-		TensorTypeQ5_0,
-		TensorTypeQ5_1,
-		TensorTypeQ8_0,
-		TensorTypeQ8_1,
-		tensorTypeIQ4_NL:
-		return 32
-	default:
-		return 256
-	}
-}
-
-func (tt TensorType) String() string {
-	switch tt {
-	case TensorTypeF32:
-		return "f32"
-	case TensorTypeF16:
-		return "f16"
-	case TensorTypeQ4_0:
-		return "q4_0"
-	case TensorTypeQ4_1:
-		return "q4_1"
-	case tensorTypeQ4_2:
-		return "q4_2"
-	case tensorTypeQ4_3:
-		return "q4_3"
-	case TensorTypeQ5_0:
-		return "q5_0"
-	case TensorTypeQ5_1:
-		return "q5_1"
-	case TensorTypeQ8_0:
-		return "q8_0"
-	case TensorTypeQ8_1:
-		return "q8_1"
-	case TensorTypeQ2_K:
-		return "q2_k"
-	case TensorTypeQ3_K:
-		return "q3_k"
-	case TensorTypeQ4_K:
-		return "q4_k"
-	case TensorTypeQ5_K:
-		return "q5_k"
-	case TensorTypeQ6_K:
-		return "q6_k"
-	case TensorTypeQ8_K:
-		return "q8_k"
-	case tensorTypeIQ2_XXS:
-		return "iq2_xxs"
-	case tensorTypeIQ2_XS:
-		return "iq2_xs"
-	case tensorTypeIQ3_XXS:
-		return "iq3_xxs"
-	case tensorTypeIQ1_S:
-		return "iq1_s"
-	case tensorTypeIQ4_NL:
-		return "iq4_nl"
-	case tensorTypeIQ3_S:
-		return "iq3_s"
-	case tensorTypeIQ2_S:
-		return "iq2_s"
-	case tensorTypeIQ4_XS:
-		return "iq4_xs"
-	case TensorTypeI8:
-		return "i8"
-	case TensorTypeI16:
-		return "i16"
-	case TensorTypeI32:
-		return "i32"
-	case TensorTypeI64:
-		return "i64"
-	case TensorTypeF64:
-		return "f64"
-	case tensorTypeIQ1_M:
-		return "iq1_m"
-	case TensorTypeBF16:
-		return "bf16"
-	case tensorTypeQ4_0_4_4:
-		return "q4_0_4_4"
-	case tensorTypeQ4_0_4_8:
-		return "q4_0_4_8"
-	case tensorTypeQ4_0_8_8:
-		return "q4_0_8_8"
-	case tensorTypeTQ1_0:
-		return "tq1_0"
-	case tensorTypeTQ2_0:
-		return "tq2_0"
-	case tensorTypeIQ4_NL_4_4:
-		return "iq4_nl_4_4"
-	case tensorTypeIQ4_NL_4_8:
-		return "iq4_nl_4_8"
-	case tensorTypeIQ4_NL_8_8:
-		return "iq4_nl_8_8"
-	default:
-		return "unknown"
-	}
-}
-
-func (tt TensorType) LogValue() slog.Value {
-	return slog.GroupValue(
-		slog.Uint64("value", uint64(tt)),
-		slog.String("name", strings.ToUpper(tt.String())),
-		slog.Int64("size", tt.typeSize()),
-		slog.Int64("block_size", tt.blockSize()),
-		slog.Float64("num_bytes", tt.NumBytes()),
-	)
-}
--- a/go.mod
+++ b/go.mod
@@ -19,13 +19,12 @@ require (
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/dlclark/regexp2 v1.11.4
 	github.com/emirpasic/gods/v2 v2.0.0-alpha
-	github.com/google/go-cmp v0.7.0
+	github.com/google/go-cmp v0.6.0
 	github.com/mattn/go-runewidth v0.0.14
 	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 v0.30.0
-	gonum.org/v1/gonum v0.15.0
 )

 require (
@@ -45,6 +44,7 @@ require (
 	github.com/xtgo/set v1.0.0 // indirect
 	go4.org/unsafe/assume-no-moving-gc v0.0.0-20231121144256-b99613f794b6 // indirect
 	golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 // indirect
+	gonum.org/v1/gonum v0.15.0 // indirect
 	gorgonia.org/vecf32 v0.9.0 // indirect
 	gorgonia.org/vecf64 v0.9.0 // indirect
 )
@@ -71,7 +71,7 @@ require (
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
 	golang.org/x/crypto v0.36.0
-	golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa // indirect
+	golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa
 	golang.org/x/net v0.38.0 // indirect
 	golang.org/x/sys v0.31.0
 	golang.org/x/term v0.30.0
--- a/go.sum
+++ b/go.sum
@@ -112,8 +112,8 @@ github.com/google/go-cmp v0.4.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/
 github.com/google/go-cmp v0.5.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.6/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
-github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
-github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
+github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
+github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
 github.com/google/uuid v1.1.2/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
--- a/integration/concurrency_test.go
+++ b/integration/concurrency_test.go
@@ -4,10 +4,7 @@ package integration

 import (
 	"context"
-	"fmt"
 	"log/slog"
-	"math"
-	"math/rand"
 	"os"
 	"strconv"
 	"sync"
@@ -17,163 +14,247 @@ import (
 	"github.com/stretchr/testify/require"

 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 )

-// Send multiple requests in parallel (concurrently) to a single model and ensure responses are expected
-func TestConcurrentGenerate(t *testing.T) {
-	// Assumes all requests have the same model
-	req, resp := GenerateRequests()
-	numParallel := int(envconfig.NumParallel() + 1)
-	iterLimit := 3
+func TestMultiModelConcurrency(t *testing.T) {
+	var (
+		req = [2]api.GenerateRequest{
+			{
+				Model:     "llama3.2:1b",
+				Prompt:    "why is the ocean blue?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}, {
+				Model:     "tinydolphin",
+				Prompt:    "what is the origin of the us thanksgiving holiday?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			},
+		}
+		resp = [2][]string{
+			{"sunlight"},
+			{"england", "english", "massachusetts", "pilgrims", "british", "festival"},
+		}
+	)
+	var wg sync.WaitGroup
+	wg.Add(len(req))
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*240)
+	defer cancel()

-	softTimeout, hardTimeout := getTimeouts(t)
-	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	for i := 0; i < len(req); i++ {
+		require.NoError(t, PullIfMissing(ctx, client, req[i].Model))
+	}
+
+	for i := 0; i < len(req); i++ {
+		go func(i int) {
+			defer wg.Done()
+			// Note: CPU based inference can crawl so don't give up too quickly
+			DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 30*time.Second)
+		}(i)
+	}
+	wg.Wait()
+}
+
+func TestIntegrationConcurrentPredict(t *testing.T) {
+	req, resp := GenerateRequests()
+	reqLimit := len(req)
+	iterLimit := 5
+
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err := strconv.ParseUint(s, 10, 64)
+		require.NoError(t, err)
+		// Don't hammer on small VRAM cards...
+		if maxVram < 4*format.GibiByte {
+			reqLimit = min(reqLimit, 2)
+			iterLimit = 2
+		}
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 9*time.Minute)
 	defer cancel()
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()

 	// Get the server running (if applicable) warm the model up with a single initial request
-	slog.Info("loading", "model", req[0].Model)
-	err := client.Generate(ctx,
-		&api.GenerateRequest{Model: req[0].Model, KeepAlive: &api.Duration{Duration: 10 * time.Second}},
-		func(response api.GenerateResponse) error { return nil },
-	)
-	if err != nil {
-		t.Fatalf("failed to load model %s: %s", req[0].Model, err)
-	}
+	DoGenerate(ctx, t, client, req[0], resp[0], 60*time.Second, 10*time.Second)

 	var wg sync.WaitGroup
-	r := rand.New(rand.NewSource(0))
-	wg.Add(numParallel)
-	for i := range numParallel {
+	wg.Add(reqLimit)
+	for i := 0; i < reqLimit; i++ {
 		go func(i int) {
 			defer wg.Done()
 			for j := 0; j < iterLimit; j++ {
-				if time.Now().Sub(started) > softTimeout {
-					slog.Info("exceeded soft timeout, winding down test")
-					return
-				}
-				k := r.Int() % len(req)
-				slog.Info("Starting", "thread", i, "iter", j)
+				slog.Info("Starting", "req", i, "iter", j)
 				// On slower GPUs it can take a while to process the concurrent requests
 				// so we allow a much longer initial timeout
-				DoGenerate(ctx, t, client, req[k], resp[k], 120*time.Second, 20*time.Second)
+				DoGenerate(ctx, t, client, req[i], resp[i], 120*time.Second, 20*time.Second)
 			}
 		}(i)
 	}
 	wg.Wait()
 }

-// Stress the scheduler and attempt to load more models than will fit to cause thrashing
-// This test will always load at least 2 models even on CPU based systems
+// Stress the system if we know how much VRAM it has, and attempt to load more models than will fit
 func TestMultiModelStress(t *testing.T) {
-	s := os.Getenv("OLLAMA_MAX_VRAM")
+	s := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
 	if s == "" {
-		s = "0"
+		t.Skip("OLLAMA_MAX_VRAM not specified, can't pick the right models for the stress test")
 	}

 	maxVram, err := strconv.ParseUint(s, 10, 64)
 	if err != nil {
 		t.Fatal(err)
 	}
-
-	smallModels := []string{
-		"llama3.2:1b",
-		"qwen3:0.6b",
-		"gemma:2b",
-		"deepseek-r1:1.5b",
-		"starcoder2:3b",
-	}
-	mediumModels := []string{
-		"qwen3:8b",
-		"llama2",
-		"deepseek-r1:7b",
-		"mistral",
-		"dolphin-mistral",
-		"gemma:7b",
-		"codellama:7b",
+	if maxVram < 2*format.GibiByte {
+		t.Skip("VRAM less than 2G, skipping model stress tests")
 	}

-	var chosenModels []string
+	type model struct {
+		name string
+		size uint64 // Approximate amount of VRAM they typically use when fully loaded in VRAM
+	}
+
+	smallModels := []model{
+		{
+			name: "llama3.2:1b",
+			size: 2876 * format.MebiByte,
+		},
+		{
+			name: "phi",
+			size: 2616 * format.MebiByte,
+		},
+		{
+			name: "gemma:2b",
+			size: 2364 * format.MebiByte,
+		},
+		{
+			name: "stable-code:3b",
+			size: 2608 * format.MebiByte,
+		},
+		{
+			name: "starcoder2:3b",
+			size: 2166 * format.MebiByte,
+		},
+	}
+	mediumModels := []model{
+		{
+			name: "llama2",
+			size: 5118 * format.MebiByte,
+		},
+		{
+			name: "mistral",
+			size: 4620 * format.MebiByte,
+		},
+		{
+			name: "orca-mini:7b",
+			size: 5118 * format.MebiByte,
+		},
+		{
+			name: "dolphin-mistral",
+			size: 4620 * format.MebiByte,
+		},
+		{
+			name: "gemma:7b",
+			size: 5000 * format.MebiByte,
+		},
+		{
+			name: "codellama:7b",
+			size: 5118 * format.MebiByte,
+		},
+	}
+
+	// These seem to be too slow to be useful...
+	// largeModels := []model{
+	// 	{
+	// 		name: "llama2:13b",
+	// 		size: 7400 * format.MebiByte,
+	// 	},
+	// 	{
+	// 		name: "codellama:13b",
+	// 		size: 7400 * format.MebiByte,
+	// 	},
+	// 	{
+	// 		name: "orca-mini:13b",
+	// 		size: 7400 * format.MebiByte,
+	// 	},
+	// 	{
+	// 		name: "gemma:7b",
+	// 		size: 5000 * format.MebiByte,
+	// 	},
+	// 	{
+	// 		name: "starcoder2:15b",
+	// 		size: 9100 * format.MebiByte,
+	// 	},
+	// }
+
+	var chosenModels []model
 	switch {
 	case maxVram < 10000*format.MebiByte:
 		slog.Info("selecting small models")
 		chosenModels = smallModels
+	// case maxVram < 30000*format.MebiByte:
 	default:
 		slog.Info("selecting medium models")
 		chosenModels = mediumModels
+		// default:
+		// 	slog.Info("selecting large models")
+		// 	chosenModels = largeModels
 	}

-	softTimeout, hardTimeout := getTimeouts(t)
-	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	req, resp := GenerateRequests()
+
+	for i := range req {
+		if i > len(chosenModels) {
+			break
+		}
+		req[i].Model = chosenModels[i].name
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute) // TODO baseline -- 10m too short
 	defer cancel()
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()

 	// Make sure all the models are pulled before we get started
-	for _, model := range chosenModels {
-		require.NoError(t, PullIfMissing(ctx, client, model))
+	for _, r := range req {
+		require.NoError(t, PullIfMissing(ctx, client, r.Model))
 	}

-	// Determine how many models we can load in parallel before we exceed VRAM
-	// The intent is to go 1 over what can fit so we force the scheduler to thrash
-	targetLoadCount := 0
-	slog.Info("Loading models to find how many can fit in VRAM before overflowing")
-	for i, model := range chosenModels {
-		req := &api.GenerateRequest{Model: model}
-		slog.Info("loading", "model", model)
-		err = client.Generate(ctx, req, func(response api.GenerateResponse) error { return nil })
-		if err != nil {
-			t.Fatalf("failed to load model %s: %s", model, err)
-		}
-		targetLoadCount++
-		if i > 0 {
-			models, err := client.ListRunning(ctx)
-			if err != nil {
-				t.Fatalf("failed to list running models: %s", err)
-			}
-			if len(models.Models) < targetLoadCount {
-				loaded := []string{}
-				for _, m := range models.Models {
-					loaded = append(loaded, m.Name)
-				}
-				slog.Info("found model load capacity", "target", targetLoadCount, "current", loaded, "chosen", chosenModels[:targetLoadCount])
-				break
-			}
-		}
-	}
-	if targetLoadCount == len(chosenModels) {
-		// TODO consider retrying the medium models
-		slog.Warn("all models being used without exceeding VRAM, set OLLAMA_MAX_VRAM so test can pick larger models")
-	}
-
-	r := rand.New(rand.NewSource(0))
 	var wg sync.WaitGroup
-	for i := range targetLoadCount {
+	consumed := uint64(256 * format.MebiByte) // Assume some baseline usage
+	for i := 0; i < len(req); i++ {
+		// Always get at least 2 models, but don't overshoot VRAM too much or we'll take too long
+		if i > 1 && consumed > maxVram {
+			slog.Info("achieved target vram exhaustion", "count", i, "vram", format.HumanBytes2(maxVram), "models", format.HumanBytes2(consumed))
+			break
+		}
+		consumed += chosenModels[i].size
+		slog.Info("target vram", "count", i, "vram", format.HumanBytes2(maxVram), "models", format.HumanBytes2(consumed))
+
 		wg.Add(1)
 		go func(i int) {
 			defer wg.Done()
-			reqs, resps := GenerateRequests()
 			for j := 0; j < 3; j++ {
-				if time.Now().Sub(started) > softTimeout {
-					slog.Info("exceeded soft timeout, winding down test")
-					return
-				}
-				k := r.Int() % len(reqs)
-				reqs[k].Model = chosenModels[i]
-				slog.Info("Starting", "model", reqs[k].Model, "iteration", j, "request", reqs[k].Prompt)
-				DoGenerate(ctx, t, client, reqs[k], resps[k],
-					120*time.Second, // Be extra patient for the model to load initially
-					10*time.Second,  // Once results start streaming, fail if they stall
-				)
+				slog.Info("Starting", "req", i, "iter", j, "model", req[i].Model)
+				DoGenerate(ctx, t, client, req[i], resp[i], 120*time.Second, 5*time.Second)
 			}
 		}(i)
 	}
 	go func() {
 		for {
-			time.Sleep(10 * time.Second)
+			time.Sleep(2 * time.Second)
 			select {
 			case <-ctx.Done():
 				return
@@ -184,21 +265,7 @@ func TestMultiModelStress(t *testing.T) {
 					continue
 				}
 				for _, m := range models.Models {
-					var procStr string
-					switch {
-					case m.SizeVRAM == 0:
-						procStr = "100% CPU"
-					case m.SizeVRAM == m.Size:
-						procStr = "100% GPU"
-					case m.SizeVRAM > m.Size || m.Size == 0:
-						procStr = "Unknown"
-					default:
-						sizeCPU := m.Size - m.SizeVRAM
-						cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
-						procStr = fmt.Sprintf("%d%%/%d%%", int(cpuPercent), int(100-cpuPercent))
-					}
-
-					slog.Info("loaded model snapshot", "model", m.Name, "CPU/GPU", procStr, "expires", format.HumanTime(m.ExpiresAt, "Never"))
+					slog.Info("loaded model snapshot", "model", m)
 				}
 			}
 		}
--- a/integration/context_test.go
+++ b/integration/context_test.go
@@ -4,8 +4,6 @@ package integration

 import (
 	"context"
-	"log/slog"
-	"sync"
 	"testing"
 	"time"

@@ -65,51 +63,3 @@ func TestContextExhaustion(t *testing.T) {
 	}
 	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived"}, 120*time.Second, 10*time.Second)
 }
-
-// Send multiple requests with prior context and ensure the response is coherant and expected
-func TestGenerateWithHistory(t *testing.T) {
-	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
-	req, resp := GenerateRequests()
-	numParallel := 2
-	iterLimit := 2
-
-	softTimeout, hardTimeout := getTimeouts(t)
-	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	// Get the server running (if applicable) warm the model up with a single initial request
-	slog.Info("loading", "model", modelOverride)
-	err := client.Generate(ctx,
-		&api.GenerateRequest{Model: modelOverride, KeepAlive: &api.Duration{Duration: 10 * time.Second}},
-		func(response api.GenerateResponse) error { return nil },
-	)
-	if err != nil {
-		t.Fatalf("failed to load model %s: %s", modelOverride, err)
-	}
-
-	var wg sync.WaitGroup
-	wg.Add(numParallel)
-	for i := range numParallel {
-		go func(i int) {
-			defer wg.Done()
-			k := i % len(req)
-			req[k].Model = modelOverride
-			for j := 0; j < iterLimit; j++ {
-				if time.Now().Sub(started) > softTimeout {
-					slog.Info("exceeded soft timeout, winding down test")
-					return
-				}
-				slog.Info("Starting", "thread", i, "iter", j)
-				// On slower GPUs it can take a while to process the concurrent requests
-				// so we allow a much longer initial timeout
-				c := DoGenerate(ctx, t, client, req[k], resp[k], 120*time.Second, 20*time.Second)
-				req[k].Context = c
-				req[k].Prompt = "tell me more!"
-			}
-		}(i)
-	}
-	wg.Wait()
-
-}
--- a/integration/library_models_test.go
+++ b/integration/library_models_test.go
@@ -1,57 +0,0 @@
-//go:build integration && library
-
-package integration
-
-import (
-	"context"
-	"log/slog"
-	"testing"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-// First run of this scenario on a target system will take a long time to download
-// ~1.5TB of models.  Set a sufficiently large -timeout for your network speed
-func TestLibraryModelsGenerate(t *testing.T) {
-	softTimeout, hardTimeout := getTimeouts(t)
-	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
-	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	chatModels := libraryChatModels
-	for _, model := range chatModels {
-		t.Run(model, func(t *testing.T) {
-			if time.Now().Sub(started) > softTimeout {
-				t.Skip("skipping remaining tests to avoid excessive runtime")
-			}
-			if err := PullIfMissing(ctx, client, model); err != nil {
-				t.Fatalf("pull failed %s", err)
-			}
-			req := api.GenerateRequest{
-				Model:     model,
-				Prompt:    "why is the sky blue?",
-				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
-					"temperature": 0.1,
-					"seed":        123,
-				},
-			}
-			anyResp := []string{"rayleigh", "scatter", "atmosphere", "nitrogen", "oxygen", "wavelength"}
-			// Special cases
-			if model == "duckdb-nsql" {
-				anyResp = []string{"select", "from"}
-			} else if model == "granite3-guardian" || model == "shieldgemma" || model == "llama-guard3" || model == "bespoke-minicheck" {
-				anyResp = []string{"yes", "no", "safe", "unsafe"}
-			} else if model == "openthinker" || model == "nexusraven" {
-				anyResp = []string{"plugin", "im_sep", "components", "function call"}
-			} else if model == "starcoder" || model == "starcoder2" || model == "magicoder" || model == "deepseek-coder" {
-				req.Prompt = "def fibonacci():"
-				anyResp = []string{"f(n)", "sequence", "n-1", "main()", "__main__", "while"}
-			}
-			DoGenerate(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
-		})
-	}
-}
--- a/integration/llm_image_test.go
+++ b/integration/llm_image_test.go
@@ -19,7 +19,7 @@ func TestVisionModels(t *testing.T) {
 	}
 	testCases := []testCase{
 		{
-			model: "qwen2.5vl",
+			model: "llava:7b",
 		},
 		{
 			model: "llama3.2-vision",
@@ -60,7 +60,6 @@ func TestVisionModels(t *testing.T) {
 }

 func TestIntegrationSplitBatch(t *testing.T) {
-	skipUnderMinVRAM(t, 6)
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
 	require.NoError(t, err)
 	req := api.GenerateRequest{
--- a/integration/model_arch_test.go
+++ b/integration/model_arch_test.go
@@ -19,6 +19,35 @@ import (
 	"github.com/ollama/ollama/format"
 )

+var (
+	started    = time.Now()
+	chatModels = []string{
+		"granite3-moe:latest",
+		"granite-code:latest",
+		"nemotron-mini:latest",
+		"command-r:latest",
+		"gemma2:latest",
+		"gemma:latest",
+		"internlm2:latest",
+		"phi3.5:latest",
+		"phi3:latest",
+		// "phi:latest", // flaky, sometimes generates no response on first query
+		"stablelm2:latest", // Predictions are off, crashes on small VRAM GPUs
+		"falcon:latest",
+		"falcon2:latest",
+		"minicpm-v:latest",
+		"mistral:latest",
+		"orca-mini:latest",
+		"llama2:latest",
+		"llama3.1:latest",
+		"llama3.2:latest",
+		"llama3.2-vision:latest",
+		"qwen2.5-coder:latest",
+		"qwen:latest",
+		"solar-pro:latest",
+	}
+)
+
 func TestModelsGenerate(t *testing.T) {
 	softTimeout, hardTimeout := getTimeouts(t)
 	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
@@ -39,13 +68,6 @@ func TestModelsGenerate(t *testing.T) {
 		slog.Warn("No VRAM info available, testing all models, so larger ones might timeout...")
 	}

-	var chatModels []string
-	if s := os.Getenv("OLLAMA_NEW_ENGINE"); s != "" {
-		chatModels = ollamaEngineChatModels
-	} else {
-		chatModels = append(ollamaEngineChatModels, llamaRunnerChatModels...)
-	}
-
 	for _, model := range chatModels {
 		t.Run(model, func(t *testing.T) {
 			if time.Now().Sub(started) > softTimeout {
--- a/integration/model_perf_test.go
+++ b/integration/model_perf_test.go
@@ -1,266 +0,0 @@
-//go:build integration && perf
-
-package integration
-
-import (
-	"context"
-	"fmt"
-	"io/ioutil"
-	"log/slog"
-	"math"
-	"os"
-	"path/filepath"
-	"strconv"
-	"strings"
-	"testing"
-	"time"
-
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/format"
-)
-
-var (
-	// Models that don't work reliably with the large context prompt in this test case
-	longContextFlakes = []string{
-		"granite-code:latest",
-		"nemotron-mini:latest",
-		"falcon:latest",  // 2k model
-		"falcon2:latest", // 2k model
-		"minicpm-v:latest",
-		"qwen:latest",
-		"solar-pro:latest",
-	}
-)
-
-// Note: this test case can take a long time to run, particularly on models with
-// large contexts.  Run with -timeout set to a large value to get reasonable coverage
-// Example usage:
-//
-// go test --tags=integration,perf -count 1 ./integration -v -timeout 90m -run TestModelsPerf 2>&1 | tee int.log
-// cat int.log | grep MODEL_PERF_HEADER | head -1| cut -f2- -d: > perf.csv
-// cat int.log | grep MODEL_PERF_DATA | cut -f2- -d: >> perf.csv
-func TestModelsPerf(t *testing.T) {
-	softTimeout, hardTimeout := getTimeouts(t)
-	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
-	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	// TODO use info API eventually
-	var maxVram uint64
-	var err error
-	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
-		maxVram, err = strconv.ParseUint(s, 10, 64)
-		if err != nil {
-			t.Fatalf("invalid  OLLAMA_MAX_VRAM %v", err)
-		}
-	} else {
-		slog.Warn("No VRAM info available, testing all models, so larger ones might timeout...")
-	}
-
-	data, err := ioutil.ReadFile(filepath.Join("testdata", "shakespeare.txt"))
-	if err != nil {
-		t.Fatalf("failed to open test data file: %s", err)
-	}
-	longPrompt := "summarize the following: " + string(data)
-
-	var chatModels []string
-	if s := os.Getenv("OLLAMA_NEW_ENGINE"); s != "" {
-		chatModels = ollamaEngineChatModels
-	} else {
-		chatModels = append(ollamaEngineChatModels, llamaRunnerChatModels...)
-	}
-
-	for _, model := range chatModels {
-		t.Run(model, func(t *testing.T) {
-			if time.Now().Sub(started) > softTimeout {
-				t.Skip("skipping remaining tests to avoid excessive runtime")
-			}
-			if err := PullIfMissing(ctx, client, model); err != nil {
-				t.Fatalf("pull failed %s", err)
-			}
-			var maxContext int
-
-			resp, err := client.Show(ctx, &api.ShowRequest{Model: model})
-			if err != nil {
-				t.Fatalf("show failed: %s", err)
-			}
-			arch := resp.ModelInfo["general.architecture"].(string)
-			maxContext = int(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))
-
-			if maxVram > 0 {
-				resp, err := client.List(ctx)
-				if err != nil {
-					t.Fatalf("list models failed %v", err)
-				}
-				for _, m := range resp.Models {
-					// For these tests we want to exercise a some amount of overflow on the CPU
-					if m.Name == model && float32(m.Size)*0.75 > float32(maxVram) {
-						t.Skipf("model %s is too large %s for available VRAM %s", model, format.HumanBytes(m.Size), format.HumanBytes(int64(maxVram)))
-					}
-				}
-			}
-			slog.Info("scneario", "model", model, "max_context", maxContext)
-			loaded := false
-			defer func() {
-				// best effort unload once we're done with the model
-				if loaded {
-					client.Generate(ctx, &api.GenerateRequest{Model: model, KeepAlive: &api.Duration{Duration: 0}}, func(rsp api.GenerateResponse) error { return nil })
-				}
-			}()
-
-			// Some models don't handle the long context data well so skip them to avoid flaky test results
-			longContextFlake := false
-			for _, flake := range longContextFlakes {
-				if model == flake {
-					longContextFlake = true
-					break
-				}
-			}
-
-			// iterate through a few context sizes for coverage without excessive runtime
-			var contexts []int
-			keepGoing := true
-			if maxContext > 16384 {
-				contexts = []int{4096, 8192, 16384, maxContext}
-			} else if maxContext > 8192 {
-				contexts = []int{4096, 8192, maxContext}
-			} else if maxContext > 4096 {
-				contexts = []int{4096, maxContext}
-			} else if maxContext > 0 {
-				contexts = []int{maxContext}
-			} else {
-				t.Fatal("unknown max context size")
-			}
-			for _, numCtx := range contexts {
-				if !keepGoing && numCtx > 8192 { // Always try up to 8k before bailing out
-					break
-				}
-				skipLongPrompt := false
-
-				// Workaround bug 11172 temporarily...
-				maxPrompt := longPrompt
-				// If we fill the context too full with the prompt, many models
-				// quickly hit context shifting and go bad.
-				if len(maxPrompt) > numCtx*2 { // typically yields ~1/2 full context
-					maxPrompt = maxPrompt[:numCtx*2]
-				}
-
-				testCases := []struct {
-					prompt  string
-					anyResp []string
-				}{
-					{"why is the sky blue?", []string{"rayleigh", "scattering", "atmosphere", "nitrogen", "oxygen"}},
-					{maxPrompt, []string{"shakespeare", "oppression", "sorrows", "gutenberg", "child", "license", "sonnet", "melancholy"}},
-				}
-				var gpuPercent int
-				for _, tc := range testCases {
-					if len(tc.prompt) > 100 && (longContextFlake || skipLongPrompt) {
-						slog.Info("skipping long prompt", "model", model, "num_ctx", numCtx, "gpu_percent", gpuPercent)
-						continue
-					}
-					req := api.GenerateRequest{
-						Model:     model,
-						Prompt:    tc.prompt,
-						KeepAlive: &api.Duration{Duration: 20 * time.Second}, // long enough to ensure a ps returns
-						Options: map[string]interface{}{
-							"temperature": 0,
-							"seed":        123,
-							"num_ctx":     numCtx,
-						},
-					}
-					atLeastOne := false
-					var resp api.GenerateResponse
-
-					stream := false
-					req.Stream = &stream
-
-					// Avoid potentially getting stuck indefinitely
-					limit := 5 * time.Minute
-					genCtx, cancel := context.WithDeadlineCause(
-						ctx,
-						time.Now().Add(limit),
-						fmt.Errorf("generate on model %s with ctx %d took longer than %v", model, numCtx, limit),
-					)
-					defer cancel()
-
-					err = client.Generate(genCtx, &req, func(rsp api.GenerateResponse) error {
-						resp = rsp
-						return nil
-					})
-					if err != nil {
-						// Avoid excessive test runs, but don't consider a failure with massive context
-						if numCtx > 16384 && strings.Contains(err.Error(), "took longer") {
-							slog.Warn("max context was taking too long, skipping", "error", err)
-							keepGoing = false
-							skipLongPrompt = true
-							continue
-						}
-						t.Fatalf("generate error: ctx:%d err:%s", numCtx, err)
-					}
-					loaded = true
-					for _, expResp := range tc.anyResp {
-						if strings.Contains(strings.ToLower(resp.Response), expResp) {
-							atLeastOne = true
-							break
-						}
-					}
-					if !atLeastOne {
-						t.Fatalf("response didn't contain expected values: ctx:%d  expected:%v response:%s ", numCtx, tc.anyResp, resp.Response)
-					}
-					models, err := client.ListRunning(ctx)
-					if err != nil {
-						slog.Warn("failed to list running models", "error", err)
-						continue
-					}
-					if len(models.Models) > 1 {
-						slog.Warn("multiple models loaded, may impact performance results", "loaded", models.Models)
-					}
-					for _, m := range models.Models {
-						if m.Name == model {
-							if m.SizeVRAM == 0 {
-								slog.Info("Model fully loaded into CPU")
-								gpuPercent = 0
-								keepGoing = false
-								skipLongPrompt = true
-							} else if m.SizeVRAM == m.Size {
-								slog.Info("Model fully loaded into GPU")
-								gpuPercent = 100
-							} else {
-								sizeCPU := m.Size - m.SizeVRAM
-								cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
-								gpuPercent = int(100 - cpuPercent)
-								slog.Info("Model split between CPU/GPU", "CPU", cpuPercent, "GPU", gpuPercent)
-								keepGoing = false
-
-								// Heuristic to avoid excessive test run time
-								if gpuPercent < 90 {
-									skipLongPrompt = true
-								}
-							}
-						}
-					}
-					fmt.Fprintf(os.Stderr, "MODEL_PERF_HEADER:%s,%s,%s,%s,%s,%s,%s\n",
-						"MODEL",
-						"CONTEXT",
-						"GPU PERCENT",
-						"PROMPT COUNT",
-						"LOAD TIME",
-						"PROMPT EVAL TPS",
-						"EVAL TPS",
-					)
-					fmt.Fprintf(os.Stderr, "MODEL_PERF_DATA:%s,%d,%d,%d,%0.2f,%0.2f,%0.2f\n",
-						model,
-						numCtx,
-						gpuPercent,
-						resp.PromptEvalCount,
-						float64(resp.LoadDuration)/1000000000.0,
-						float64(resp.PromptEvalCount)/(float64(resp.PromptEvalDuration)/1000000000.0),
-						float64(resp.EvalCount)/(float64(resp.EvalDuration)/1000000000.0),
-					)
-				}
-			}
-		})
-	}
-}
--- a/integration/testdata/embed.json
+++ b/integration/testdata/embed.json
--- a/integration/testdata/shakespeare.txt
+++ b/integration/testdata/shakespeare.txt
--- a/integration/utils_test.go
+++ b/integration/utils_test.go
@@ -28,242 +28,12 @@ import (
 	"github.com/stretchr/testify/require"
 )

-var (
+const (
 	smol = "llama3.2:1b"
 )

-var (
-	started = time.Now()
-
-	// Note: add newer models at the top of the list to test them first
-	ollamaEngineChatModels = []string{
-		"gpt-oss:20b",
-		"gemma3n:e2b",
-		"mistral-small3.2:latest",
-		"deepseek-r1:1.5b",
-		"llama3.2-vision:latest",
-		"qwen2.5-coder:latest",
-		"qwen2.5vl:3b",
-		"qwen3:0.6b", // dense
-		"qwen3:30b",  // MOE
-		"gemma3:1b",
-		"llama3.1:latest",
-		"llama3.2:latest",
-		"gemma2:latest",
-		"minicpm-v:latest",    // arch=qwen2
-		"granite-code:latest", // arch=llama
-	}
-	llamaRunnerChatModels = []string{
-		"mistral:latest",
-		"falcon3:latest",
-		"granite3-moe:latest",
-		"command-r:latest",
-		"nemotron-mini:latest",
-		"phi3.5:latest",
-		"solar-pro:latest",
-		"internlm2:latest",
-		"codellama:latest", // arch=llama
-		"phi3:latest",
-		"falcon2:latest",
-		"gemma:latest",
-		"llama2:latest",
-		"nous-hermes:latest",
-		"orca-mini:latest",
-		"qwen:latest",
-		"stablelm2:latest", // Predictions are off, crashes on small VRAM GPUs
-		"falcon:latest",
-	}
-
-	// Some library models are quite large - ensure large VRAM and sufficient disk space
-	// before running scenarios based on this set
-	libraryChatModels = []string{
-		"alfred",
-		"athene-v2",
-		"aya-expanse",
-		"aya",
-		"bakllava",
-		"bespoke-minicheck",
-		"codebooga",
-		"codegeex4",
-		"codegemma",
-		"codellama",
-		"codeqwen",
-		"codestral",
-		"codeup",
-		"cogito",
-		"command-a",
-		"command-r-plus",
-		"command-r",
-		"command-r7b-arabic",
-		"command-r7b",
-		"dbrx",
-		"deepcoder",
-		"deepscaler",
-		"deepseek-coder-v2",
-		"deepseek-coder",
-		"deepseek-llm",
-		"deepseek-r1",
-		// "deepseek-v2.5", // requires 155 GB VRAM
-		"deepseek-v2",
-		// "deepseek-v3", // requires 482 GB VRAM
-		"devstral",
-		"dolphin-llama3",
-		"dolphin-mistral",
-		"dolphin-mixtral",
-		"dolphin-phi",
-		"dolphin3",
-		"dolphincoder",
-		"duckdb-nsql",
-		"everythinglm",
-		"exaone-deep",
-		"exaone3.5",
-		"falcon",
-		"falcon2",
-		"falcon3",
-		"firefunction-v2",
-		"gemma",
-		"gemma2",
-		"gemma3",
-		"gemma3n",
-		"glm4",
-		"goliath",
-		"gpt-oss:20b",
-		"granite-code",
-		"granite3-dense",
-		"granite3-guardian",
-		"granite3-moe",
-		"granite3.1-dense",
-		"granite3.1-moe",
-		"granite3.2-vision",
-		"granite3.2",
-		"granite3.3",
-		"hermes3",
-		"internlm2",
-		"llama-guard3",
-		"llama-pro",
-		"llama2-chinese",
-		"llama2-uncensored",
-		"llama2",
-		"llama3-chatqa",
-		"llama3-gradient",
-		"llama3-groq-tool-use",
-		"llama3.1",
-		"llama3.2-vision",
-		"llama3.2",
-		"llama3.3",
-		"llama3",
-		"llama4",
-		"llava-llama3",
-		"llava-phi3",
-		"llava",
-		"magicoder",
-		"magistral",
-		"marco-o1",
-		"mathstral",
-		"meditron",
-		"medllama2",
-		"megadolphin",
-		"minicpm-v",
-		"mistral-large",
-		"mistral-nemo",
-		"mistral-openorca",
-		"mistral-small",
-		"mistral-small3.1",
-		"mistral-small3.2",
-		"mistral",
-		"mistrallite",
-		"mixtral",
-		"moondream",
-		"nemotron-mini",
-		"nemotron",
-		"neural-chat",
-		"nexusraven",
-		"notus",
-		"nous-hermes",
-		"nous-hermes2-mixtral",
-		"nous-hermes2",
-		"nuextract",
-		"olmo2",
-		"open-orca-platypus2",
-		"openchat",
-		"opencoder",
-		"openhermes",
-		"openthinker",
-		"orca-mini",
-		"orca2",
-		// "phi", // unreliable
-		"phi3.5",
-		"phi3",
-		"phi4-mini-reasoning",
-		"phi4-mini",
-		"phi4-reasoning",
-		"phi4",
-		"phind-codellama",
-		"qwen",
-		"qwen2-math",
-		"qwen2.5-coder",
-		"qwen2.5",
-		"qwen2.5vl",
-		"qwen2",
-		"qwen3:0.6b", // dense
-		"qwen3:30b",  // MOE
-		"qwq",
-		"r1-1776",
-		"reader-lm",
-		"reflection",
-		"sailor2",
-		"samantha-mistral",
-		"shieldgemma",
-		"smallthinker",
-		"smollm",
-		"smollm2",
-		"solar-pro",
-		"solar",
-		"sqlcoder",
-		"stable-beluga",
-		"stable-code",
-		"stablelm-zephyr",
-		"stablelm2",
-		"starcoder",
-		"starcoder2",
-		"starling-lm",
-		"tinydolphin",
-		"tinyllama",
-		"tulu3",
-		"vicuna",
-		"wizard-math",
-		"wizard-vicuna-uncensored",
-		"wizard-vicuna",
-		"wizardcoder",
-		"wizardlm-uncensored",
-		"wizardlm2",
-		"xwinlm",
-		"yarn-llama2",
-		"yarn-mistral",
-		"yi-coder",
-		"yi",
-		"zephyr",
-	}
-	libraryEmbedModels = []string{
-		"all-minilm",
-		"bge-large",
-		"bge-m3",
-		"granite-embedding",
-		"mxbai-embed-large",
-		"nomic-embed-text",
-		"paraphrase-multilingual",
-		"snowflake-arctic-embed",
-		"snowflake-arctic-embed2",
-	}
-)
-
-func init() {
+func Init() {
 	lifecycle.InitLogging()
-	custom := os.Getenv("OLLAMA_TEST_SMOL_MODEL")
-	if custom != "" {
-		slog.Info("setting smol test model to " + custom)
-		smol = custom
-	}
 }

 func FindPort() string {
@@ -472,19 +242,15 @@ func GenerateTestHelper(ctx context.Context, t *testing.T, genReq api.GenerateRe
 	DoGenerate(ctx, t, client, genReq, anyResp, 30*time.Second, 10*time.Second)
 }

-func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq api.GenerateRequest, anyResp []string, initialTimeout, streamTimeout time.Duration) []int {
+func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq api.GenerateRequest, anyResp []string, initialTimeout, streamTimeout time.Duration) {
 	stallTimer := time.NewTimer(initialTimeout)
 	var buf bytes.Buffer
-	var context []int
 	fn := func(response api.GenerateResponse) error {
 		// fmt.Print(".")
 		buf.Write([]byte(response.Response))
 		if !stallTimer.Reset(streamTimeout) {
 			return errors.New("stall was detected while streaming response, aborting")
 		}
-		if len(response.Context) > 0 {
-			context = response.Context
-		}
 		return nil
 	}

@@ -505,10 +271,6 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 			t.Errorf("generate stalled.  Response so far:%s", buf.String())
 		}
 	case <-done:
-		if genErr != nil && strings.Contains(genErr.Error(), "model requires more system memory") {
-			slog.Warn("model is too large for the target test system", "model", genReq.Model, "error", genErr)
-			return context
-		}
 		require.NoError(t, genErr, "failed with %s request prompt %s ", genReq.Model, genReq.Prompt)
 		// Verify the response contains the expected data
 		response := buf.String()
@@ -524,7 +286,6 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 	case <-ctx.Done():
 		t.Error("outer test context done while waiting for generate")
 	}
-	return context
 }

 // Generate a set of requests
@@ -533,35 +294,55 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 	return []api.GenerateRequest{
 			{
 				Model:     smol,
-				Prompt:    "why is the ocean blue? Be brief but factual in your reply",
+				Prompt:    "why is the ocean blue?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
 			}, {
 				Model:     smol,
-				Prompt:    "why is the color of dirt brown? Be brief but factual in your reply",
+				Prompt:    "why is the color of dirt brown?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
 			}, {
 				Model:     smol,
-				Prompt:    "what is the origin of the US thanksgiving holiday? Be brief but factual in your reply",
+				Prompt:    "what is the origin of the us thanksgiving holiday?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
 			}, {
 				Model:     smol,
-				Prompt:    "what is the origin of independence day? Be brief but factual in your reply",
+				Prompt:    "what is the origin of independence day?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
 			}, {
 				Model:     smol,
-				Prompt:    "what is the composition of air? Be brief but factual in your reply",
+				Prompt:    "what is the composition of air?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
 			},
 		},
 		[][]string{
-			{"sunlight", "scattering", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorbs", "wavelength"},
-			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigments", "particles", "iron oxide", "rust", "air", "water", "mixture", "mixing"},
-			{"england", "english", "massachusetts", "pilgrims", "colonists", "independence", "british", "feast", "family", "gatherings", "traditions", "turkey", "colonial", "period", "harvest", "agricultural", "european settlers", "american revolution", "civil war", "16th century", "17th century", "native american", "united states"},
+			{"sunlight"},
+			{"soil", "organic", "earth", "black", "tan"},
+			{"england", "english", "massachusetts", "pilgrims", "british"},
 			{"fourth", "july", "declaration", "independence"},
 			{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}
--- a/kvcache/causal.go
+++ b/kvcache/causal.go
@@ -19,32 +19,17 @@ type shiftFn func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, e
 // 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
-
-	// swaWindowSize is the number of tokens that will be included in the mask
-	// during attention operations. swaMemorySize is the number of tokens that
-	// will be retained in memory for partial prefix caching. Set to math.MaxInt32
-	// for unlimited or if sliding window attention is not being used.
-	swaWindowSize int32
-	swaMemorySize int32
-
-	chunkSize int32
+	DType      ml.DType
+	windowSize int32
+	chunkSize  int32

 	opts CausalOptions

-	// maxBatch is the largest batch that we might receive
-	maxBatch int
-
 	// config controls mostly backend-specific optimizations
 	config *ml.CacheConfig

 	// ** current forward pass **

-	// curReserve indicates that this forward pass is only for
-	// memory reservation and we should not update our metadata
-	// based on it.
-	curReserve bool
-
 	// the active layer for Get and Put
 	curLayer int

@@ -95,41 +80,32 @@ type cellRange struct {

 func NewCausalCache(shift shiftFn) *Causal {
 	return &Causal{
-		shiftFn: shift,
-		ctxs:    make(map[int]ml.Context),
-		keys:    make(map[int]ml.Tensor),
-		values:  make(map[int]ml.Tensor),
+		windowSize: math.MaxInt32,
+		shiftFn:    shift,
+		ctxs:       make(map[int]ml.Context),
+		keys:       make(map[int]ml.Tensor),
+		values:     make(map[int]ml.Tensor),
 	}
 }

 func NewSWACache(windowSize int32, shift shiftFn) *Causal {
 	return &Causal{
-		swaWindowSize: windowSize,
-		shiftFn:       shift,
-		ctxs:          make(map[int]ml.Context),
-		keys:          make(map[int]ml.Tensor),
-		values:        make(map[int]ml.Tensor),
-	}
-}
-
-func NewSWAMemCache(windowSize int32, memorySize int32, shift shiftFn) *Causal {
-	return &Causal{
-		swaWindowSize: windowSize,
-		swaMemorySize: memorySize,
-		shiftFn:       shift,
-		ctxs:          make(map[int]ml.Context),
-		keys:          make(map[int]ml.Tensor),
-		values:        make(map[int]ml.Tensor),
+		windowSize: windowSize,
+		shiftFn:    shift,
+		ctxs:       make(map[int]ml.Context),
+		keys:       make(map[int]ml.Tensor),
+		values:     make(map[int]ml.Tensor),
 	}
 }

 func NewChunkedAttentionCache(chunkSize int32, shift shiftFn) *Causal {
 	return &Causal{
-		chunkSize: chunkSize,
-		shiftFn:   shift,
-		ctxs:      make(map[int]ml.Context),
-		keys:      make(map[int]ml.Tensor),
-		values:    make(map[int]ml.Tensor),
+		windowSize: math.MaxInt32,
+		chunkSize:  chunkSize,
+		shiftFn:    shift,
+		ctxs:       make(map[int]ml.Context),
+		keys:       make(map[int]ml.Tensor),
+		values:     make(map[int]ml.Tensor),
 	}
 }

@@ -154,25 +130,11 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
 		c.config.MaskDType = ml.DTypeF32
 	}

-	if c.swaWindowSize == 0 {
-		c.swaWindowSize = math.MaxInt32
-	}
-	if c.swaMemorySize == 0 {
-		c.swaMemorySize = c.swaWindowSize
-	}
-	if int(c.swaMemorySize) > capacity {
-		c.swaMemorySize = math.MaxInt32
-	}
-
-	if c.swaMemorySize < c.swaWindowSize {
-		panic(fmt.Errorf("sliding window memory (%v) must be at least as large as the window (%v)", c.swaMemorySize, c.swaWindowSize))
-	}
-
 	var cacheSize int
-	if c.swaMemorySize == math.MaxInt32 {
+	if c.windowSize == math.MaxInt32 || capacity < int(c.windowSize) {
 		cacheSize = maxSequences * capacity
 	} else {
-		cacheSize = (maxSequences * int(c.swaMemorySize)) + maxBatch
+		cacheSize = (maxSequences * int(c.windowSize)) + maxBatch
 	}
 	cacheSize = roundUp(cacheSize, c.config.CachePadding)
 	c.cells = make([]cacheCell, cacheSize)
@@ -180,7 +142,6 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
 	c.DType = dtype
 	c.cellRanges = make(map[int]cellRange)
 	c.backend = backend
-	c.maxBatch = maxBatch
 }

 func (c *Causal) SetConfig(config ml.CacheConfig) {
@@ -198,13 +159,12 @@ func (c *Causal) Close() {
 }

 func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
-	c.curReserve = reserve
 	c.curBatchSize = len(batch.Positions)
 	c.curSequences = batch.Sequences
 	c.curPositions = batch.Positions
 	c.opts.Except = nil

-	if !c.curReserve {
+	if !reserve {
 		c.updateSlidingWindow()

 		var err error
@@ -214,10 +174,10 @@ func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) e
 			c.curLoc, err = c.findStartLoc()
 		}
 		if err != nil {
-			slog.Warn("unable to find a kv cache slot", "cache", c)
 			return err
 		}

+		c.curCellRange = newRange()
 		for i, pos := range batch.Positions {
 			seq := batch.Sequences[i]

@@ -228,12 +188,19 @@ func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) e
 				seqRange = newRange()
 			}

-			seqRange.min = min(seqRange.min, c.curLoc+i)
-			c.curCellRange.min = min(c.curCellRange.min, c.curLoc+i)
-
-			seqRange.max = max(seqRange.max, c.curLoc+i)
-			c.curCellRange.max = max(c.curCellRange.max, c.curLoc+i)
+			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
 		}
 	} else {
@@ -244,9 +211,10 @@ func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) e
 		c.curCellRange.max = len(c.cells) - 1
 	}

-	c.curMask = c.buildMask(ctx)
+	var err error
+	c.curMask, err = c.buildMask(ctx)

-	return nil
+	return err
 }

 func newRange() cellRange {
@@ -275,16 +243,7 @@ func (c *Causal) findStartLoc() (int, error) {
 }

 func (c *Causal) updateSlidingWindow() {
-	c.curCellRange = newRange()
-
-	if c.swaMemorySize == math.MaxInt32 {
-		for _, seq := range c.curSequences {
-			if seqRange, ok := c.cellRanges[seq]; ok {
-				c.curCellRange.min = min(c.curCellRange.min, seqRange.min)
-				c.curCellRange.max = max(c.curCellRange.max, seqRange.max)
-			}
-		}
-
+	if c.windowSize == math.MaxInt32 {
 		return
 	}

@@ -314,16 +273,12 @@ func (c *Causal) updateSlidingWindow() {

 		for i := oldRange.min; i <= oldRange.max; i++ {
 			if slices.Contains(c.cells[i].sequences, seq) {
-				if c.cells[i].pos < pos-c.swaMemorySize {
+				if c.cells[i].pos < pos-c.windowSize {
 					c.cells[i].sequences = slices.DeleteFunc(c.cells[i].sequences, func(s int) bool { return s == seq })
 				} else {
 					newRange.min = min(newRange.min, i)
 					newRange.max = max(newRange.max, i)
 				}
-				if c.cells[i].pos >= pos-c.swaWindowSize {
-					c.curCellRange.min = min(c.curCellRange.min, i)
-					c.curCellRange.max = max(c.curCellRange.max, i)
-				}
 			}
 		}

@@ -342,7 +297,7 @@ func roundUp(length, pad int) int {
 // 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) ml.Tensor {
+func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 	// Align and pad the two dimensions as required by the backend
 	batchSize := roundUp(c.curBatchSize, c.config.MaskBatchPadding)

@@ -350,11 +305,6 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 	c.curCellRange.max = roundUp(c.curCellRange.max+1, c.config.CachePadding) - 1

 	length := c.curCellRange.max - c.curCellRange.min + 1
-
-	if c.curReserve {
-		return ctx.Input().Empty(c.config.MaskDType, length, batchSize)
-	}
-
 	mask := make([]float32, batchSize*length)

 	for i := range c.curBatchSize {
@@ -363,7 +313,7 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 			if !slices.Contains(c.cells[j].sequences, c.curSequences[i]) ||
 				(enabled && c.cells[j].pos > c.curPositions[i]) ||
 				c.chunkSize > 0 && c.cells[j].pos < c.curPositions[i]-c.curPositions[i]%c.chunkSize ||
-				c.cells[j].pos < c.curPositions[i]-c.swaWindowSize {
+				c.cells[j].pos < c.curPositions[i]-c.windowSize {
 				mask[i*length+(j-c.curCellRange.min)] = float32(math.Inf(-1))
 			}
 		}
@@ -375,7 +325,10 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 		mask[i] = float32(math.Inf(-1))
 	}

-	maskTensor := ctx.Input().FromFloatSlice(mask, length, batchSize)
+	maskTensor, err := ctx.Input().FromFloatSlice(mask, length, batchSize)
+	if err != nil {
+		return nil, err
+	}

 	if c.config.MaskDType != ml.DTypeF32 {
 		out := ctx.Input().Empty(c.config.MaskDType, maskTensor.Shape()...)
@@ -383,7 +336,7 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 		maskTensor = out
 	}

-	return maskTensor
+	return maskTensor, nil
 }

 func (c *Causal) moveCells(ctx ml.Context, src, dst, length int) {
@@ -521,8 +474,6 @@ func (c *Causal) defrag() {

 		c.cellRanges[seq] = seqRange
 	}
-
-	c.updateSlidingWindow()
 }

 func (c *Causal) SetLayer(layer int) {
@@ -540,7 +491,12 @@ func (c *Causal) SetCausal(ctx ml.Context, opts CausalOptions) {
 	if !slices.Equal(c.opts.Except, opts.Except) {
 		c.opts = opts
 		if ctx != nil {
-			c.curMask = c.buildMask(ctx)
+			var err error
+			c.curMask, err = c.buildMask(ctx)
+			if err != nil {
+				// This error should never occur because we have previously built a mask with the same shape
+				panic(fmt.Errorf("SetCausal: %w", err))
+			}
 		}
 	}
 }
@@ -648,7 +604,7 @@ func (c *Causal) CopyPrefix(srcSeq, dstSeq int, len int32) {
 }

 func (c *Causal) CanResume(seq int, pos int32) bool {
-	if c.swaMemorySize == math.MaxInt32 {
+	if c.windowSize == math.MaxInt32 {
 		return true
 	}

@@ -670,8 +626,8 @@ func (c *Causal) CanResume(seq int, pos int32) bool {
 		return false
 	}

-	lastWindowStart := max(0, last-c.swaMemorySize)
-	posWindowStart := max(0, pos-c.swaWindowSize)
+	lastWindowStart := max(0, last-c.windowSize)
+	posWindowStart := max(0, pos-c.windowSize)

 	return posWindowStart >= lastWindowStart
 }
@@ -681,64 +637,51 @@ func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 		return ErrNotSupported
 	}

+	ctx := c.backend.NewContext()
+	defer ctx.Close()
+
 	seqRange := c.cellRanges[seq]
+	size := seqRange.max - seqRange.min + 1

-	for start := seqRange.min; start <= seqRange.max; start += c.maxBatch {
-		size := min(seqRange.max-start+1, c.maxBatch)
-		offsets := make([]int32, size)
+	offsets := make([]int32, size)
+	for i := range offsets {
+		cell := c.cells[seqRange.min+i]

-		var batchFirst, batchLast int
-
-		batchFirst = -1
-		for i := range offsets {
-			cell := c.cells[start+i]
-
-			if slices.Contains(cell.sequences, seq) && cell.pos >= beginIndex {
-				offsets[i] = offset
-				if batchFirst < 0 {
-					batchFirst = i
-				}
-				batchLast = i
-			}
+		if slices.Contains(cell.sequences, seq) && cell.pos >= beginIndex {
+			offsets[i] = offset
 		}
+	}

-		if batchFirst < 0 {
+	kShift, err := ctx.Input().FromIntSlice(offsets, len(offsets))
+	if err != nil {
+		return err
+	}
+
+	for i, key := range c.keys {
+		if key == nil {
 			continue
 		}

-		offsets = offsets[batchFirst : batchLast+1]
+		kHeadDim := key.Dim(0)
+		numKVHeads := key.Dim(1)
+		rowSize := key.Stride(2)

-		ctx := c.backend.NewContext()
-		kShift := ctx.Input().FromIntSlice(offsets, len(offsets))
+		key = key.View(ctx, rowSize*seqRange.min,
+			kHeadDim, key.Stride(1),
+			numKVHeads, key.Stride(2),
+			size,
+		)

-		for i, key := range c.keys {
-			if key == nil {
-				continue
-			}
-
-			kHeadDim := key.Dim(0)
-			numKVHeads := key.Dim(1)
-			rowSize := key.Stride(2)
-
-			key = key.View(ctx, rowSize*(start+batchFirst),
-				kHeadDim, key.Stride(1),
-				numKVHeads, key.Stride(2),
-				len(offsets),
-			)
-
-			roped, err := c.shiftFn(ctx, i, key, kShift)
-			if err != nil {
-				ctx.Close()
-				return err
-			}
-
-			ctx.Forward(roped.Copy(ctx, key))
+		roped, err := c.shiftFn(ctx, i, key, kShift)
+		if err != nil {
+			return err
 		}

-		ctx.Compute()
-		ctx.Close()
+		ctx.Forward(roped.Copy(ctx, key))
 	}

+	ctx.Compute()
+
 	return nil
 }

--- a/kvcache/causal_test.go
+++ b/kvcache/causal_test.go
@@ -60,8 +60,6 @@ func TestSWA(t *testing.T) {

 	cache.Init(backend, ml.DTypeF16, 1, 16, 16)

-	x := float32(math.Inf(-1))
-
 	tests := []testCase{
 		{
 			name:          "FirstBatch",
@@ -71,12 +69,7 @@ func TestSWA(t *testing.T) {
 			pos:           []int32{0, 1, 2, 3},
 			expected:      []float32{1, 2, 3, 4},
 			expectedShape: []int{1, 1, 4},
-			expectedMask: []float32{
-				0, x, x, x,
-				0, 0, x, x,
-				x, 0, 0, x,
-				x, x, 0, 0,
-			},
+			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},
 		},
 		{
 			name:          "SecondBatch",
@@ -86,53 +79,7 @@ func TestSWA(t *testing.T) {
 			pos:           []int32{4, 5},
 			expected:      []float32{5, 6, 3, 4},
 			expectedShape: []int{1, 1, 4},
-			expectedMask: []float32{
-				0, x, x, 0,
-				0, 0, x, x,
-			},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestSWAMem(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewSWAMemCache(1, 3, nil)
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
-
-	x := float32(math.Inf(-1))
-
-	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, x, x, x,
-				0, 0, x, x,
-				x, 0, 0, x,
-				x, x, 0, 0,
-			},
-		},
-		{
-			name:          "SecondBatch",
-			in:            []float32{5, 6},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{0, 0},
-			pos:           []int32{4, 5},
-			expected:      []float32{4, 5, 6},
-			expectedShape: []int{1, 1, 3},
-			expectedMask: []float32{
-				0, 0, x,
-				x, 0, 0,
-			},
+			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1))},
 		},
 	}

@@ -397,7 +344,7 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 			}

 			cache.SetLayer(0)
-			tensor := context.FromFloatSlice(test.in, test.inShape...)
+			tensor, _ := context.FromFloatSlice(test.in, test.inShape...)
 			cache.Put(context, tensor, tensor)

 			out, _, mask := cache.Get(context)
@@ -439,7 +386,7 @@ func TestCanResume(t *testing.T) {
 	}

 	cache.SetLayer(0)
-	tensor := context.FromFloatSlice([]float32{1, 2, 3, 4}, 1, 1, 4)
+	tensor, _ := context.FromFloatSlice([]float32{1, 2, 3, 4}, 1, 1, 4)
 	cache.Put(context, tensor, tensor)

 	// with window size 4, nothing has slid out of the window yet
@@ -466,7 +413,7 @@ func TestCanResume(t *testing.T) {
 	}

 	cache.SetLayer(0)
-	tensor = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
+	tensor, _ = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
 	cache.Put(context, tensor, tensor)

 	// only the latest position has overlapping windows
@@ -490,70 +437,6 @@ func TestCanResume(t *testing.T) {
 	}
 }

-func TestCanResumeSWAMem(t *testing.T) {
-	backend := &testBackend{}
-	windowSize := int32(4)
-	memSize := int32(5)
-	cache := NewSWAMemCache(windowSize, memSize, nil)
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
-
-	context := backend.NewContext()
-	defer context.Close()
-
-	err := cache.StartForward(context, input.Batch{
-		Positions: []int32{0, 1, 2, 3, 4, 5},
-		Sequences: []int{0, 0, 0, 0, 0, 0},
-	}, false)
-	if err != nil {
-		t.Fatalf("StartForward failed: %v", err)
-	}
-
-	cache.SetLayer(0)
-	tensor := context.FromFloatSlice([]float32{1, 2, 3, 4, 5, 6}, 1, 1, 6)
-	cache.Put(context, tensor, tensor)
-
-	// shift window by adding position 6
-	err = cache.StartForward(context, input.Batch{
-		Positions: []int32{6, 7},
-		Sequences: []int{0, 0},
-	}, false)
-	if err != nil {
-		t.Fatalf("StartForward failed: %v", err)
-	}
-
-	cache.SetLayer(0)
-	tensor = context.FromFloatSlice([]float32{7, 8}, 1, 1, 2)
-	cache.Put(context, tensor, tensor)
-
-	// only the latest position has overlapping windows
-	if cache.CanResume(0, 0) {
-		t.Errorf("after shift: CanResume(0, 0) = true, want false (outside window)")
-	}
-	if cache.CanResume(0, 1) {
-		t.Errorf("after shift: CanResume(0, 1) = true, want false (outside window)")
-	}
-	if cache.CanResume(0, 2) {
-		t.Errorf("after shift: CanResume(0, 2) = true, want false (outside window)")
-	}
-	if cache.CanResume(0, 3) {
-		t.Errorf("after shift: CanResume(0, 3) = true, want false (outside window)")
-	}
-	if cache.CanResume(0, 4) {
-		t.Errorf("after shift: CanResume(0, 4) = true, want false (outside window)")
-	}
-	if cache.CanResume(0, 5) {
-		t.Errorf("after shift: CanResume(0, 5) = true, want false (outside window)")
-	}
-	if !cache.CanResume(0, 6) {
-		t.Errorf("after shift: CanResume(0, 6) = false, want true (inside window)")
-	}
-	if !cache.CanResume(0, 7) {
-		t.Errorf("after shift: CanResume(0, 7) = false, want true (latest position)")
-	}
-}
-
 type testBackend struct {
 	ml.Backend
 }
@@ -587,24 +470,24 @@ func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
 	return c.Empty(dtype, shape...)
 }

-func (c *testContext) FromFloatSlice(s []float32, shape ...int) ml.Tensor {
+func (c *testContext) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
 	t := c.Empty(ml.DTypeF32, shape...).(*testTensor)

 	copy(t.data, s)

-	return t
+	return t, nil
 }

-func (c *testContext) FromIntSlice(s []int32, shape ...int) ml.Tensor {
+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, _ := c.FromFloatSlice(f, shape...)
 	out.(*testTensor).dtype = ml.DTypeI32

-	return out
+	return out, nil
 }

 func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
@@ -613,7 +496,7 @@ func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tenso
 		s = append(s, i)
 	}

-	out := c.FromFloatSlice(s, len(s))
+	out, _ := c.FromFloatSlice(s, len(s))
 	out.(*testTensor).dtype = dtype
 	return out
 }
@@ -625,7 +508,7 @@ func (c *testContext) Forward(...ml.Tensor) ml.Context { return c }

 func (c *testContext) Compute(...ml.Tensor) {}

-func (c *testContext) Reserve() {}
+func (c *testContext) Reserve() error { return nil }

 func (c *testContext) MaxGraphNodes() int {
 	return 10
--- a/llama/build-info.cpp
+++ b/llama/build-info.cpp
@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "e54d41befcc1575f4c898c5ff4ef43970cead75f";
+char const *LLAMA_COMMIT = "e1e8e0991ffd9e99a445c6812bb519d5bac9f4b5";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";
--- a/llama/llama.cpp/.rsync-filter
+++ b/llama/llama.cpp/.rsync-filter
@@ -1,32 +1,23 @@
-protect .rsync-filter
-protect *.go
-include /common/
-include /common/base64.*
-include /common/common.*
-include /common/json-schema-to-grammar.*
-include /common/json.*
-include /common/log.*
-include /common/sampling.*
-include /include/
-include /include/llama.*
-include /include/llama-*.*
-include /tools/
-include /tools/mtmd/
-include /tools/mtmd/*.h
-include /tools/mtmd/clip.cpp
-include /tools/mtmd/mtmd.cpp
-include /tools/mtmd/mtmd-audio.cpp
-include /tools/mtmd/mtmd-helper.cpp
-include /src/
-include /src/llama.*
-include /src/llama-*.*
-include /src/unicode-data.*
-include /src/unicode.*
-include /vendor/
-include /vendor/miniaudio/
-include /vendor/miniaudio/*.h
-include /vendor/nlohmann/
-include /vendor/nlohmann/*.hpp
-include /vendor/stb/
-include /vendor/stb/*.h
-hide *
+protect **/*.go
+include common/
+include common/base64.*
+include common/common.*
+include common/json-schema-to-grammar.*
+include common/json.*
+include common/log.*
+include common/sampling.*
+include common/stb_image.*
+include include/
+include include/llama.*
+include include/llama-*.*
+include examples/
+include examples/llava/
+include examples/llava/clip.*
+include examples/llava/clip-impl.*
+include examples/llava/llava.*
+include src/
+include src/llama.*
+include src/llama-*.*
+include src/unicode-data.*
+include src/unicode.*
+exclude *
--- a/llama/llama.cpp/common/common.cpp
+++ b/llama/llama.cpp/common/common.cpp
@@ -203,7 +203,6 @@ bool set_process_priority(enum ggml_sched_priority prio) {

    DWORD p = NORMAL_PRIORITY_CLASS;
    switch (prio) {
-        case GGML_SCHED_PRIO_LOW:      p = BELOW_NORMAL_PRIORITY_CLASS; break;
        case GGML_SCHED_PRIO_NORMAL:   p = NORMAL_PRIORITY_CLASS;       break;
        case GGML_SCHED_PRIO_MEDIUM:   p = ABOVE_NORMAL_PRIORITY_CLASS; break;
        case GGML_SCHED_PRIO_HIGH:     p = HIGH_PRIORITY_CLASS;         break;
@@ -229,7 +228,6 @@ bool set_process_priority(enum ggml_sched_priority prio) {

    int p = 0;
    switch (prio) {
-        case GGML_SCHED_PRIO_LOW:      p =  5;  break;
        case GGML_SCHED_PRIO_NORMAL:   p =  0;  break;
        case GGML_SCHED_PRIO_MEDIUM:   p = -5;  break;
        case GGML_SCHED_PRIO_HIGH:     p = -10; break;
@@ -445,37 +443,9 @@ void string_replace_all(std::string & s, const std::string & search, const std::
    s = std::move(builder);
 }

-bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
-    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-}
-
-bool string_remove_suffix(std::string & str, const std::string_view & suffix) {
-    bool has_suffix = string_ends_with(str, suffix);
-    if (has_suffix) {
-        str = str.substr(0, str.size() - suffix.size());
-    }
-    return has_suffix;
-}
-
-size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
-    if (!str.empty() && !stop.empty()) {
-        const char text_last_char = str.back();
-        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
-            if (stop[char_index] == text_last_char) {
-                const auto current_partial = stop.substr(0, char_index + 1);
-                if (string_ends_with(str, current_partial)) {
-                    return str.size() - char_index - 1;
-                }
-            }
-        }
-    }
-
-    return std::string::npos;
-}
-
 std::string regex_escape(const std::string & s) {
    static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
-    return std::regex_replace(s, special_chars, "\\$&");
+    return std::regex_replace(s, special_chars, "\\$0");
 }

 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
@@ -715,17 +685,11 @@ bool fs_validate_filename(const std::string & filename) {
        // disable C++17 deprecation warning for std::codecvt_utf8
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
-
        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;

 #if defined(__clang__)
 #    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
 #endif

        filename_utf32 = converter.from_bytes(filename);
@@ -782,9 +746,6 @@ bool fs_validate_filename(const std::string & filename) {
    return true;
 }

-#include <iostream>
-
-
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -802,16 +763,9 @@ bool fs_create_directory_with_parents(const std::string & path) {
    // process path from front to back, procedurally creating directories
    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
        const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();

-        pos_slash += 1;
-
-        // skip the drive letter, in some systems it can return an access denied error
-        if (subpath.length() == 2 && subpath[1] == ':') {
-            continue;
-        }
-
-        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
-
+        const bool success = CreateDirectoryW(test, NULL);
        if (!success) {
            const DWORD error = GetLastError();

@@ -825,6 +779,8 @@ bool fs_create_directory_with_parents(const std::string & path) {
                return false;
            }
        }
+
+        pos_slash += 1;
    }

    return true;
@@ -874,7 +830,7 @@ std::string fs_get_cache_directory() {
    if (getenv("LLAMA_CACHE")) {
        cache_directory = std::getenv("LLAMA_CACHE");
    } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
        if (std::getenv("XDG_CACHE_HOME")) {
            cache_directory = std::getenv("XDG_CACHE_HOME");
        } else {
@@ -920,6 +876,31 @@ struct common_init_result common_init_from_params(common_params & params) {

    const llama_vocab * vocab = llama_model_get_vocab(model);

+    if (params.reranking) {
+        bool ok = true;
+
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_model_free(model);
+
+            return iparams;
+        }
+    }
+
    auto cparams = common_context_params_to_llama(params);

    llama_context * lctx = llama_init_from_model(model, cparams);
@@ -929,7 +910,7 @@ struct common_init_result common_init_from_params(common_params & params) {
        return iparams;
    }

-    if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
+    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
        LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
        params.ctx_shift = false;
    }
@@ -961,35 +942,6 @@ struct common_init_result common_init_from_params(common_params & params) {
        }
    }

-    if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
-        bool ok = true;
-
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
-        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
-            ok = false;
-        } else if (!has_eos) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (!ok) {
-            llama_free(lctx);
-            llama_model_free(model);
-
-            return iparams;
-        }
-    }
-
    // load and optionally apply lora adapters
    for (auto & la : params.lora_adapters) {
        llama_adapter_lora_ptr lora;
@@ -1014,19 +966,13 @@ struct common_init_result common_init_from_params(common_params & params) {
        params.sampling.ignore_eos = false;
    }

-    // initialize once
-    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-        if (llama_vocab_is_eog(vocab, i)) {
-            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
-            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
-        }
-    }
-
    if (params.sampling.ignore_eos) {
-        // add EOG biases to the active set of logit biases
-        params.sampling.logit_bias.insert(
-                params.sampling.logit_bias.end(),
-                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+        for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+            if (llama_vocab_is_eog(vocab, i)) {
+                LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
+                params.sampling.logit_bias.push_back({i, -INFINITY});
+            }
+        }
    }

    if (params.sampling.penalty_last_n == -1) {
@@ -1071,7 +1017,7 @@ struct common_init_result common_init_from_params(common_params & params) {
        if (llama_model_has_decoder(model)) {
            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
        }
-        llama_memory_clear(llama_get_memory(lctx), true);
+        llama_kv_self_clear(lctx);
        llama_synchronize(lctx);
        llama_perf_context_reset(lctx);
        llama_set_warmup(lctx, false);
@@ -1122,7 +1068,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
    mparams.use_mmap        = params.use_mmap;
    mparams.use_mlock       = params.use_mlock;
    mparams.check_tensors   = params.check_tensors;
-    mparams.use_extra_bufts = !params.no_extra_bufts;

    if (params.kv_overrides.empty()) {
        mparams.kv_overrides = NULL;
@@ -1138,9 +1083,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
        mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
    }

-    mparams.progress_callback           = params.load_progress_callback;
-    mparams.progress_callback_user_data = params.load_progress_callback_user_data;
-
    return mparams;
 }

@@ -1154,6 +1096,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.n_threads         = params.cpuparams.n_threads;
    cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
                                params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
+    cparams.logits_all        = params.logits_all;
    cparams.embeddings        = params.embedding;
    cparams.rope_scaling_type = params.rope_scaling_type;
    cparams.rope_freq_base    = params.rope_freq_base;
@@ -1171,9 +1114,11 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.offload_kqv       = !params.no_kv_offload;
    cparams.flash_attn        = params.flash_attn;
    cparams.no_perf           = params.no_perf;
-    cparams.op_offload        = !params.no_op_offload;
-    cparams.swa_full          = params.swa_full;
-    cparams.kv_unified        = params.kv_unified;
+
+    if (params.reranking) {
+        cparams.embeddings    = true;
+        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
+    }

    cparams.type_k = params.cache_type_k;
    cparams.type_v = params.cache_type_v;
@@ -1307,9 +1252,6 @@ std::vector<llama_token> common_tokenize(
    int n_tokens = text.length() + 2 * add_special;
    std::vector<llama_token> result(n_tokens);
    n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
-    if (n_tokens == std::numeric_limits<int32_t>::min()) {
-        throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit");
-    }
    if (n_tokens < 0) {
        result.resize(-n_tokens);
        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
@@ -1364,6 +1306,81 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto
    return text;
 }

+//
+// KV cache utils
+//
+
+void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
+        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        int seq_count = 0;
+        for (int j = 0; j < view.n_seq_max; j++) {
+            if (cs_curr[j] >= 0) { seq_count++; }
+        }
+        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
+    }
+
+    printf("\n=== Done dumping\n");
+}
+
+void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
+        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    std::unordered_map<llama_seq_id, size_t> seqs;
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+        for (int j = 0; j < view.n_seq_max; j++) {
+            if (cs_curr[j] < 0) { continue; }
+            if (seqs.find(cs_curr[j]) == seqs.end()) {
+                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+                const size_t sz = seqs.size();
+                seqs[cs_curr[j]] = sz;
+            }
+        }
+        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+    }
+
+    printf("=== Sequence legend: ");
+    for (const auto & it : seqs) {
+        printf("%zu=%d, ", it.second, it.first);
+    }
+    printf("'+'=other sequence ids");
+
+    c_curr = view.cells;
+    cs_curr = view.cells_sequences;
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        for (int j = 0; j < view.n_seq_max; j++) {
+            if (cs_curr[j] >= 0) {
+                const auto & it = seqs.find(cs_curr[j]);
+                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
+            } else {
+                putchar('.');
+            }
+        }
+        putchar(' ');
+    }
+
+    printf("\n=== Done dumping\n");
+}
+
 //
 // Embedding utils
 //
@@ -1548,20 +1565,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c

    return result;
 }
-
-ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
-    const int64_t ne_datapoint = llama_n_ctx(ctx);
-    const int64_t ndata        = (tokens.size() - ne_datapoint - 1) / stride;
-    ggml_opt_dataset_t result = ggml_opt_dataset_init(
-        GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
-
-    llama_token * data   = (llama_token *) ggml_opt_dataset_data(result)->data;
-    llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
-
-    for (int64_t idata = 0; idata < ndata; ++idata) {
-        memcpy(data   + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
-        memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
-    }
-
-    return result;
-}
--- a/llama/llama.cpp/common/common.go
+++ b/llama/llama.cpp/common/common.go
@@ -1,6 +1,6 @@
 package common

-// #cgo CXXFLAGS: -std=c++17
-// #cgo CPPFLAGS: -I${SRCDIR}/../include -I${SRCDIR}/../vendor
+// #cgo CXXFLAGS: -std=c++11
+// #cgo CPPFLAGS: -I${SRCDIR}/../include
 // #cgo CPPFLAGS: -I${SRCDIR}/../../../ml/backend/ggml/ggml/include
 import "C"
--- a/llama/llama.cpp/common/common.h
+++ b/llama/llama.cpp/common/common.h
@@ -6,9 +6,7 @@

 #include <set>
 #include <string>
-#include <string_view>
 #include <vector>
-#include <map>
 #include <sstream>

 #ifdef _WIN32
@@ -68,6 +66,7 @@ enum llama_example {
    LLAMA_EXAMPLE_COMMON,
    LLAMA_EXAMPLE_SPECULATIVE,
    LLAMA_EXAMPLE_MAIN,
+    LLAMA_EXAMPLE_INFILL,
    LLAMA_EXAMPLE_EMBEDDING,
    LLAMA_EXAMPLE_PERPLEXITY,
    LLAMA_EXAMPLE_RETRIEVAL,
@@ -77,11 +76,10 @@ enum llama_example {
    LLAMA_EXAMPLE_SERVER,
    LLAMA_EXAMPLE_CVECTOR_GENERATOR,
    LLAMA_EXAMPLE_EXPORT_LORA,
-    LLAMA_EXAMPLE_MTMD,
+    LLAMA_EXAMPLE_LLAVA,
    LLAMA_EXAMPLE_LOOKUP,
    LLAMA_EXAMPLE_PARALLEL,
    LLAMA_EXAMPLE_TTS,
-    LLAMA_EXAMPLE_DIFFUSION,

    LLAMA_EXAMPLE_COUNT,
 };
@@ -98,7 +96,6 @@ enum common_sampler_type {
    COMMON_SAMPLER_TYPE_XTC         = 8,
    COMMON_SAMPLER_TYPE_INFILL      = 9,
    COMMON_SAMPLER_TYPE_PENALTIES   = 10,
-    COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11,
 };

 // dimensionality reduction methods, used by cvector-generator
@@ -117,7 +114,7 @@ enum common_grammar_trigger_type {
    COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN,
    COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
 };

 struct common_grammar_trigger {
@@ -164,7 +161,6 @@ struct common_params_sampling {
    std::vector<enum common_sampler_type> samplers = {
        COMMON_SAMPLER_TYPE_PENALTIES,
        COMMON_SAMPLER_TYPE_DRY,
-        COMMON_SAMPLER_TYPE_TOP_N_SIGMA,
        COMMON_SAMPLER_TYPE_TOP_K,
        COMMON_SAMPLER_TYPE_TYPICAL_P,
        COMMON_SAMPLER_TYPE_TOP_P,
@@ -178,8 +174,7 @@ struct common_params_sampling {
    std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
    std::set<llama_token>               preserved_tokens;

-    std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
-    std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
+    std::vector<llama_logit_bias> logit_bias; // logit biases to apply

    // print the parameters into a string
    std::string print() const;
@@ -201,10 +196,6 @@ struct common_params_speculative {
    int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
    float   p_split      =  0.1f; // speculative decoding split probability
    float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
-    std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
-
-    ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
-    ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V

    struct cpu_params cpuparams;
    struct cpu_params cpuparams_batch;
@@ -220,26 +211,9 @@ struct common_params_vocoder {
    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
 };

-struct common_params_diffusion {
-    int32_t steps         = 128;
-    bool    visual_mode   = false;
-
-    float   eps           = 0;        // epsilon for timesteps
-    int32_t block_length  = 0;        // block length for generation
-
-    int32_t algorithm     = 4;        // default algorithm: low-confidence
-    float   alg_temp      = 0.0f;     // algorithm temperature
-
-    float   cfg_scale     = 0;        // classifier-free guidance scale
-    bool    add_gumbel_noise = false; // add gumbel noise to the logits if temp > 0.0
-};
-
 enum common_reasoning_format {
    COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_AUTO,
-    COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
-    COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
-    COMMON_REASONING_FORMAT_GRANITE,         // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
+    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
 };

 struct common_params {
@@ -287,7 +261,6 @@ struct common_params {
    struct common_params_sampling    sampling;
    struct common_params_speculative speculative;
    struct common_params_vocoder     vocoder;
-    struct common_params_diffusion   diffusion;

    struct common_params_model model;

@@ -316,7 +289,6 @@ struct common_params {
    int32_t verbosity                  = 0;
    int32_t control_vector_layer_start = -1; // layer range for control vector
    int32_t control_vector_layer_end   = -1; // layer range for control vector
-    bool    offline                    = false;

    int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
    int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -349,19 +321,17 @@ struct common_params {
    bool flash_attn        = false; // flash attention
    bool no_perf           = false; // disable performance metrics
    bool ctx_shift         = true;  // context shift on inifinite text generation
-    bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
-    bool kv_unified        = false; // enable unified KV cache

    bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
+    bool logits_all        = false; // return logits for all tokens in the batch
    bool use_mmap          = true;  // use mmap for faster loads
    bool use_mlock         = false; // use mlock to keep model in memory
    bool verbose_prompt    = false; // print prompt tokens before generation
    bool display_prompt    = true;  // print prompt before generation
+    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
    bool no_kv_offload     = false; // disable KV offloading
    bool warmup            = true;  // warmup run
    bool check_tensors     = false; // validate tensor data
-    bool no_op_offload     = false; // globally disable offload host tensor operations to device
-    bool no_extra_bufts    = false; // disable extra buffer types (used for weight repacking)

    bool single_turn       = false; // single turn chat conversation

@@ -370,7 +340,7 @@ struct common_params {

    common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;

-    // multimodal models (see tools/mtmd)
+    // multimodal models (see examples/llava)
    struct common_params_model mmproj;
    bool mmproj_use_gpu = true;     // use GPU for multimodal model
    bool no_mmproj = false;         // explicitly disable multimodal model
@@ -381,7 +351,7 @@ struct common_params {
    int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
    std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
    std::string embd_sep   = "\n";  // separator of embeddings
-    std::string cls_sep    = "\t";  // separator of classification sequences
+    bool reranking         = false; // enable reranking support on server

    // server params
    int32_t port           = 8080;         // server listens on this network port
@@ -392,21 +362,16 @@ struct common_params {

    std::string hostname      = "127.0.0.1";
    std::string public_path   = "";                                                                         // NOLINT
-    std::string api_prefix    = "";                                                                         // NOLINT
    std::string chat_template = "";                                                                         // NOLINT
    bool use_jinja = false;                                                                                 // NOLINT
    bool enable_chat_template = true;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_AUTO;
-    int reasoning_budget = -1;
-    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;

    std::vector<std::string> api_keys;

    std::string ssl_file_key  = "";                                                                         // NOLINT
    std::string ssl_file_cert = "";                                                                         // NOLINT

-    std::map<std::string, std::string> default_template_kwargs;
-
    // "advanced" endpoints are disabled by default for better security
    bool webui            = true;
    bool endpoint_slots   = false;
@@ -441,19 +406,16 @@ struct common_params {
    int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
    int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
    int32_t i_chunk     =  0; // start processing from this chunk
-    int8_t  imat_dat    =  0; // whether the legacy imatrix.dat format should be output (gguf <= 0 < dat)

-    bool process_output  = false; // collect data for the output tensor
-    bool compute_ppl     = true;  // whether to compute perplexity
-    bool show_statistics = false; // show imatrix statistics per tensor
-    bool parse_special   = false; // whether to parse special tokens during imatrix tokenization
+    bool process_output = false; // collect data for the output tensor
+    bool compute_ppl    = true;  // whether to compute perplexity

    // cvector-generator params
    int n_pca_batch = 100;
    int n_pca_iterations = 1000;
    dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_positive_file = "tools/cvector-generator/positive.txt";
-    std::string cvector_negative_file = "tools/cvector-generator/negative.txt";
+    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
+    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";

    bool spm_infill = false; // suffix/prefix/middle pattern for infill

@@ -462,11 +424,6 @@ struct common_params {

    // common params
    std::string out_file; // output filename for all example programs
-    // optional callback for model loading progress and cancellation:
-    // called with a progress value between 0.0 and 1.0.
-    // return false from callback to abort model loading or true to continue
-    llama_progress_callback load_progress_callback = NULL;
-    void *                  load_progress_callback_user_data = NULL;
 };

 // call once at the start of a program if it uses libcommon
@@ -544,10 +501,10 @@ static bool string_starts_with(const std::string & str,
    return str.rfind(prefix, 0) == 0;
 }

-// While we wait for C++20's std::string::ends_with...
-bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
-bool string_remove_suffix(std::string & str, const std::string_view & suffix);
-size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
+static bool string_ends_with(const std::string & str,
+                               const std::string & suffix) {  // While we wait for C++20's std::string::ends_with...
+    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+}

 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
@@ -656,6 +613,16 @@ std::string common_detokenize(
        const std::vector<llama_token> & tokens,
                                  bool   special = true);

+//
+// KV cache utils
+//
+
+// Dump the KV cache view with the number of sequences per cell.
+void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
+
+// Dump the KV cache view showing individual sequences in each cell (long output).
+void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+
 //
 // Embedding utils
 //
@@ -697,9 +664,3 @@ const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

 }
-
-//
-// training utils
-//
-
-ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
--- a/llama/llama.cpp/common/json-schema-to-grammar.cpp
+++ b/llama/llama.cpp/common/json-schema-to-grammar.cpp
@@ -1,9 +1,8 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"

-#include <nlohmann/json.hpp>
-
 #include <algorithm>
+#include <fstream>
 #include <map>
 #include <regex>
 #include <sstream>
@@ -41,6 +40,49 @@ static std::string build_repetition(const std::string & item_rule, int min_items
    return result;
 }

+/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
+class string_view {
+    const std::string & _str;
+    const size_t _start;
+    const size_t _end;
+public:
+    string_view(const std::string & str, size_t start = 0, size_t end  = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
+
+    size_t size() const {
+        return _end - _start;
+    }
+
+    size_t length() const {
+        return size();
+    }
+
+    operator std::string() const {
+        return str();
+    }
+
+    std::string str() const {
+        return _str.substr(_start, _end - _start);
+    }
+
+    string_view substr(size_t pos, size_t len = std::string::npos) const {
+        return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
+    }
+
+    char operator[](size_t pos) const {
+        auto index = _start + pos;
+        if (index >= _end) {
+            throw std::out_of_range("string_view index out of range");
+        }
+        return _str[_start + pos];
+    }
+
+    bool operator==(const string_view & other) const {
+        std::string this_str = *this;
+        std::string other_str = other;
+        return this_str == other_str;
+    }
+};
+
 static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
    auto has_min = min_value != std::numeric_limits<int>::min();
    auto has_max = max_value != std::numeric_limits<int>::max();
@@ -69,14 +111,14 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
        }
        out << "}";
    };
-    std::function<void(const std::string_view &, const std::string_view &)> uniform_range =
-        [&](const std::string_view & from, const std::string_view & to) {
+    std::function<void(const string_view &, const string_view &)> uniform_range =
+        [&](const string_view & from, const string_view & to) {
            size_t i = 0;
            while (i < from.length() && i < to.length() && from[i] == to[i]) {
                i++;
            }
            if (i > 0) {
-                out << "\"" << from.substr(0, i) << "\"";
+                out << "\"" << from.substr(0, i).str() << "\"";
            }
            if (i < from.length() && i < to.length()) {
                if (i > 0) {
--- a/llama/llama.cpp/common/json-schema-to-grammar.h
+++ b/llama/llama.cpp/common/json-schema-to-grammar.h
@@ -1,9 +1,9 @@
 #pragma once

-#include <nlohmann/json_fwd.hpp>
-
-#include <functional>
-#include <string>
+#include "ggml.h"
+// Change JSON_ASSERT from assert() to GGML_ASSERT:
+#define JSON_ASSERT GGML_ASSERT
+#include "json.hpp"

 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                   bool force_gbnf = false);
--- a/llama/llama.cpp/vendor/nlohmann/json.hpp
+++ b/llama/llama.cpp/vendor/nlohmann/json.hpp
--- a/llama/llama.cpp/common/sampling.cpp
+++ b/llama/llama.cpp/common/sampling.cpp
@@ -1,7 +1,6 @@
 #include "sampling.h"

 #include "common.h"
-#include "log.h"

 #include <cmath>
 #include <unordered_map>
@@ -161,7 +160,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
        GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
    } else {
-        std::vector<std::string> trigger_patterns;
+        std::vector<std::string> patterns_at_start;
        std::vector<std::string> patterns_anywhere;
        std::vector<llama_token> trigger_tokens;
        for (const auto & trigger : params.grammar_triggers) {
@@ -173,13 +172,10 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                    break;
                }
                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START:
                {
-                    patterns_anywhere.push_back(trigger.value);
-                    break;
-                }
-                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL:
-                {
-                    trigger_patterns.push_back(trigger.value);
+                    const auto & pattern = trigger.value;
+                    (trigger.type == COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START ? patterns_at_start : patterns_anywhere).push_back(pattern);
                    break;
                }
                case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
@@ -193,6 +189,10 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
            }
        }

+        std::vector<std::string> trigger_patterns;
+        if (!patterns_at_start.empty()) {
+            trigger_patterns.push_back("^(" + string_join(patterns_at_start, "|") + ")[\\s\\S]*");
+        }
        if (!patterns_anywhere.empty()) {
            trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
        }
@@ -229,48 +229,51 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                params.logit_bias.data()));

    if (params.mirostat == 0) {
-        for (const auto & cnstr : params.samplers) {
-            switch (cnstr) {
-                case COMMON_SAMPLER_TYPE_DRY:
-                    {
-                        std::vector<const char *> c_breakers;
-                        c_breakers.reserve(params.dry_sequence_breakers.size());
-                        for (const auto & str : params.dry_sequence_breakers) {
-                            c_breakers.push_back(str.c_str());
-                        }
+        if (params.top_n_sigma >= 0) {
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
+        } else {
+            for (const auto & cnstr : params.samplers) {
+                switch (cnstr) {
+                    case COMMON_SAMPLER_TYPE_DRY:
+                        {
+                            std::vector<const char *> c_breakers;
+                            c_breakers.reserve(params.dry_sequence_breakers.size());
+                            for (const auto & str : params.dry_sequence_breakers) {
+                                c_breakers.push_back(str.c_str());
+                            }

-                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
-                    }
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_K:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
-                    break;
-                case COMMON_SAMPLER_TYPE_MIN_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_XTC:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                    break;
-                case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                    break;
-                case COMMON_SAMPLER_TYPE_INFILL:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill      (vocab));
-                    break;
-                case COMMON_SAMPLER_TYPE_PENALTIES:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties   (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
-                    break;
-                default:
-                    GGML_ASSERT(false && "unknown sampler type");
+                            llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        }
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_K:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_MIN_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_XTC:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                        break;
+                    case COMMON_SAMPLER_TYPE_INFILL:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
+                        break;
+                    case COMMON_SAMPLER_TYPE_PENALTIES:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                        break;
+                    default:
+                        GGML_ASSERT(false && "unknown sampler type");
+                }
            }
        }
        llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
@@ -472,7 +475,6 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
        case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
        case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
-        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return 's';
        case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
        case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
        case COMMON_SAMPLER_TYPE_XTC:         return 'x';
@@ -488,7 +490,6 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
        case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
        case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
-        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return "top_n_sigma";
        case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
        case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
        case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
@@ -503,7 +504,6 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
        { "dry",         COMMON_SAMPLER_TYPE_DRY },
        { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
        { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
-        { "top_n_sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
        { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
        { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
@@ -517,7 +517,6 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
    std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
        { "top-k",       COMMON_SAMPLER_TYPE_TOP_K },
        { "top-p",       COMMON_SAMPLER_TYPE_TOP_P },
-        { "top-n-sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { "nucleus",     COMMON_SAMPLER_TYPE_TOP_P },
        { "typical-p",   COMMON_SAMPLER_TYPE_TYPICAL_P },
        { "typical",     COMMON_SAMPLER_TYPE_TYPICAL_P },
@@ -534,16 +533,14 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
        auto sampler = sampler_canonical_name_map.find(name);
        if (sampler != sampler_canonical_name_map.end()) {
            samplers.push_back(sampler->second);
-            continue;
-        }
-        if (allow_alt_names) {
-            sampler = sampler_alt_name_map.find(name);
-            if (sampler != sampler_alt_name_map.end()) {
-                samplers.push_back(sampler->second);
-                continue;
+        } else {
+            if (allow_alt_names) {
+                sampler = sampler_alt_name_map.find(name);
+                if (sampler != sampler_alt_name_map.end()) {
+                    samplers.push_back(sampler->second);
+                }
            }
        }
-        LOG_WRN("%s: unable to match sampler by name '%s'\n", __func__, name.c_str());
    }

    return samplers;
@@ -555,7 +552,6 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
-        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_N_SIGMA), COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
@@ -570,8 +566,6 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
        const auto sampler = sampler_name_map.find(c);
        if (sampler != sampler_name_map.end()) {
            samplers.push_back(sampler->second);
-        } else {
-            LOG_WRN("%s: unable to match sampler by char '%c'\n", __func__, c);
        }
    }

--- a/llama/llama.cpp/vendor/stb/stb_image.h
+++ b/llama/llama.cpp/vendor/stb/stb_image.h
--- a/Show More
+++ b/Show More