logs: quiet down context canceled on completion and scheduler noise (#12553 )

* logs: quiet down context canceled on completion If the client closes the connection before Completion finishes, we were logging at error level implying the runner crashed which was misleading. time=2025-10-08T22:59:20.566-07:00 level=ERROR source=server.go:1490 msg="post predict" error="Post \"http://127.0.0.1:57736/completion\": context canceled" * quiet down scheduler log error on expected case Since we don't hold the lock while performing memory load calculations, other runners can unload in parallel, so finding no runner to unload is a valid scenario which we shouldn't log at error level.
routes: structured outputs for gpt-oss (#12460 )
2026-01-08 07:30:53 -05:00 · 2025-10-09 10:37:47 -07:00 · 2025-10-08 19:13:38 -07:00 · 2025-10-08 18:21:01 -07:00 · 2025-10-08 17:57:57 -07:00 · 2025-10-08 17:05:05 -07:00
488 changed files with 45893 additions and 27513 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -65,14 +65,36 @@ jobs:
            arch: amd64
            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-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
            cuda-version: '12.8'
            flags: ''
+            runner_dir: 'cuda_v12'
+          - os: windows
+            arch: amd64
+            preset: 'CUDA 13'
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+              - '"crt"'
+              - '"nvvm"'
+              - '"nvptxcompiler"'
+            cuda-version: '13.0'
+            flags: ''
+            runner_dir: 'cuda_v13'
          - 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"'
+            runner_dir: 'rocm'
    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
    environment: release
    env:
@@ -96,7 +118,7 @@ jobs:
          $ErrorActionPreference = "Stop"
          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            $subpackages = @("cudart", "nvcc", "cublas", "cublas_dev") | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
+            $subpackages = @(${{ join(matrix.cuda-components, ', ') }}) | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
            Start-Process -FilePath .\install.exe -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
          }

@@ -138,9 +160,10 @@ jobs:
        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 }}" ${{ matrix.flags }}
+          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} -DOLLAMA_RUNNER_DIR="${{ matrix.runner_dir }}"
          cmake --build --parallel --preset "${{ matrix.preset }}"
          cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || 'CPU' }}" --strip --parallel 8
+          Remove-Item -Path dist\lib\ollama\rocm\rocblas\library\*gfx906* -ErrorAction SilentlyContinue
        env:
          CMAKE_GENERATOR: Ninja
      - uses: actions/upload-artifact@v4
@@ -153,19 +176,19 @@ jobs:
      matrix:
        os: [windows]
        arch: [amd64, arm64]
+        include:
+        - os: windows
+          arch: amd64
+          llvmarch: x86_64
+        - os: windows
+          arch: arm64
+          llvmarch: aarch64
    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
    environment: release
    needs: [setup-environment]
    env:
      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
    steps:
-      - name: Install AMD64 system dependencies
-        if: matrix.arch == 'amd64'
-        run: |
-          $ErrorActionPreference = "Stop"
-          Start-Process "C:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
-          echo "C:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
      - name: Install ARM64 system dependencies
        if: matrix.arch == 'arm64'
        run: |
@@ -177,15 +200,29 @@ jobs:

          choco install -y --no-progress git gzip
          echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-
-          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-aarch64.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt-aarch64.zip"
-          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt-aarch64.zip -DestinationPath "C:\Program Files\"
-          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt-aarch64").path
-          echo $installPath\bin | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install clang and gcc-compat
+        run: |
+          $ErrorActionPreference = "Stop"
+          Set-ExecutionPolicy Bypass -Scope Process -Force
+          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-${{ matrix.llvmarch }}.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
+          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt.zip -DestinationPath "C:\Program Files\"
+          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt*").path
+          echo "$installPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
      - uses: actions/checkout@v4
      - uses: actions/setup-go@v5
        with:
          go-version-file: go.mod
+      - name: Verify gcc is actually clang
+        run: |
+          $ErrorActionPreference='Continue'
+          $version=& gcc -v 2>&1
+          $version=$version -join "`n"
+          echo "gcc is $version"
+          if ($version -notmatch 'clang') {
+            echo "ERROR: GCC must be clang for proper utf16 handling"
+            exit 1
+          }
+          $ErrorActionPreference='Stop'
      - run: |
          go build -o dist/${{ matrix.os }}-${{ matrix.arch }}/ .
      - uses: actions/upload-artifact@v4
@@ -232,7 +269,7 @@ jobs:
            case "$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_v*)        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 ;;
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -46,7 +46,7 @@ jobs:
        include:
          - preset: CPU
          - preset: CUDA
-            container: nvidia/cuda:12.8.1-devel-ubuntu22.04
+            container: nvidia/cuda:13.0.0-devel-ubuntu22.04
            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
          - preset: ROCm
            container: rocm/dev-ubuntu-22.04:6.1.2
@@ -78,8 +78,17 @@ jobs:
        include:
          - preset: CPU
          - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+              - '"crt"'
+              - '"nvvm"'
+              - '"nvptxcompiler"'
+            cuda-version: '13.0'
          - 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"'
@@ -102,7 +111,8 @@ jobs:
          $ErrorActionPreference = "Stop"
          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_12.8", "nvcc_12.8", "cublas_12.8", "cublas_dev_12.8")) -NoNewWindow -Wait
+            $subpackages = @(${{ join(matrix.cuda-components, ', ') }}) | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
          }

          $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
--- a/.gitignore
+++ b/.gitignore
@@ -6,6 +6,7 @@
 dist
 build
 .cache
+.gocache
 *.exe
 .idea
 test_data
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -38,7 +38,7 @@ if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
 endif()

 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
-set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama)
+set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama/${OLLAMA_RUNNER_DIR})

 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
@@ -81,7 +81,7 @@ if(CMAKE_CUDA_COMPILER)
    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
    install(TARGETS ggml-cuda
        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_LIBRARY_DIR}
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
            PRE_INCLUDE_REGEXES cublas cublasLt cudart
            PRE_EXCLUDE_REGEXES ".*"
        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
@@ -89,23 +89,26 @@ if(CMAKE_CUDA_COMPILER)
    )
 endif()

-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a|1200|1201):xnack[+-]$"
+set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(908|90a|1200|1201):xnack[+-]$"
    CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a|1200|1201):xnack[+-]$\"."
+    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(908|90a|1200|1201):xnack[+-]$\"."
 )

 check_language(HIP)
 if(CMAKE_HIP_COMPILER)
    set(HIP_PLATFORM "amd")

-    find_package(hip REQUIRED)
    if(NOT AMDGPU_TARGETS)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012]|120[01])$")
-    elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
+        find_package(hip REQUIRED)
+        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(94[012]|101[02]|1030|110[012]|120[01])$")
+    endif()
+
+    if(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
        list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
    endif()

    if(AMDGPU_TARGETS)
+        find_package(hip REQUIRED)
        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)

        if (WIN32)
@@ -114,7 +117,6 @@ if(CMAKE_HIP_COMPILER)

        target_compile_definitions(ggml-hip PRIVATE GGML_HIP_NO_VMM)

-        set(OLLAMA_HIP_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/rocm)
        install(TARGETS ggml-hip
            RUNTIME_DEPENDENCY_SET rocm
            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
@@ -125,13 +127,13 @@ if(CMAKE_HIP_COMPILER)
                PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register drm drm_amdgpu numa elf
                PRE_EXCLUDE_REGEXES ".*"
                POST_EXCLUDE_REGEXES "system32"
-            RUNTIME DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP
+            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
+            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
        )

        foreach(HIP_LIB_BIN_INSTALL_DIR IN ITEMS ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR})
            if(EXISTS ${HIP_LIB_BIN_INSTALL_DIR}/rocblas)
-                install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP)
+                install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP)
                break()
            endif()
        endforeach()
--- a/CMakePresets.json
+++ b/CMakePresets.json
@@ -18,6 +18,14 @@
      "name": "CUDA",
      "inherits": [ "Default" ]
    },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "50-virtual;60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
+      }
+    },
    {
      "name": "CUDA 12",
      "inherits": [ "CUDA" ],
@@ -26,6 +34,14 @@
        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
      }
    },
+    {
+      "name": "CUDA 13",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;110-virtual;120-virtual;121-virtual",
+        "CMAKE_CUDA_FLAGS": "-t 2"
+      }
+    },
    {
      "name": "JetPack 5",
      "inherits": [ "CUDA" ],
@@ -52,7 +68,7 @@
      "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-"
+        "AMDGPU_TARGETS": "gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
      }
    }
  ],
@@ -72,11 +88,21 @@
      "configurePreset": "CUDA",
      "targets": [ "ggml-cuda" ]
    },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 11"
+    },
    {
      "name": "CUDA 12",
      "inherits": [ "CUDA" ],
      "configurePreset": "CUDA 12"
    },
+    {
+      "name": "CUDA 13",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 13"
+    },
    {
      "name": "JetPack 5",
      "inherits": [ "CUDA" ],
--- a/69
+++ b/69
@@ -1,6 +1,7 @@
 # vim: filetype=dockerfile

 ARG FLAVOR=${TARGETARCH}
+ARG PARALLEL=8

 ARG ROCMVERSION=6.3.3
 ARG JETPACK5VERSION=r35.4.1
@@ -34,26 +35,52 @@ ENV LDFLAGS=-s
 FROM base AS cpu
 RUN dnf install -y gcc-toolset-11-gcc gcc-toolset-11-gcc-c++
 ENV PATH=/opt/rh/gcc-toolset-11/root/usr/bin:$PATH
+ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
    cmake --preset 'CPU' \
-        && cmake --build --parallel --preset 'CPU' \
-        && cmake --install build --component CPU --strip --parallel 8
+        && cmake --build --parallel ${PARALLEL} --preset 'CPU' \
+        && cmake --install build --component CPU --strip --parallel ${PARALLEL}
+
+FROM base AS cuda-11
+ARG CUDA11VERSION=11.8
+RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
+ENV PATH=/usr/local/cuda-11/bin:$PATH
+ARG PARALLEL
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 11' -DOLLAMA_RUNNER_DIR="cuda_v11" \
+        && cmake --build --parallel ${PARALLEL} --preset 'CUDA 11' \
+        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}

 FROM base AS cuda-12
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
 ENV PATH=/usr/local/cuda-12/bin:$PATH
+ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 12' \
-        && cmake --build --parallel --preset 'CUDA 12' \
-        && cmake --install build --component CUDA --strip --parallel 8
+    cmake --preset 'CUDA 12' -DOLLAMA_RUNNER_DIR="cuda_v12"\
+        && cmake --build --parallel ${PARALLEL} --preset 'CUDA 12' \
+        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}
+
+
+FROM base AS cuda-13
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-}
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+ARG PARALLEL
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 13' -DOLLAMA_RUNNER_DIR="cuda_v13" \
+        && cmake --build --parallel ${PARALLEL} --preset 'CUDA 13' \
+        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}
+

 FROM base AS rocm-6
 ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
+ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'ROCm 6' \
-        && cmake --build --parallel --preset 'ROCm 6' \
-        && cmake --install build --component HIP --strip --parallel 8
+    cmake --preset 'ROCm 6' -DOLLAMA_RUNNER_DIR="rocm" \
+        && cmake --build --parallel ${PARALLEL} --preset 'ROCm 6' \
+        && cmake --install build --component HIP --strip --parallel ${PARALLEL}
+RUN rm -f dist/lib/ollama/rocm/rocblas/library/*gfx90[06]*

 FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
 ARG CMAKEVERSION
@@ -61,10 +88,11 @@ RUN apt-get update && apt-get install -y curl ccache \
    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
 COPY CMakeLists.txt CMakePresets.json .
 COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 5' \
-        && cmake --build --parallel --preset 'JetPack 5' \
-        && cmake --install build --component CUDA --strip --parallel 8
+    cmake --preset 'JetPack 5' -DOLLAMA_RUNNER_DIR="cuda_jetpack5" \
+        && cmake --build --parallel ${PARALLEL} --preset 'JetPack 5' \
+        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}

 FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK6VERSION} AS jetpack-6
 ARG CMAKEVERSION
@@ -72,10 +100,11 @@ RUN apt-get update && apt-get install -y curl ccache \
    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
 COPY CMakeLists.txt CMakePresets.json .
 COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 6' \
-        && cmake --build --parallel --preset 'JetPack 6' \
-        && cmake --install build --component CUDA --strip --parallel 8
+    cmake --preset 'JetPack 6' -DOLLAMA_RUNNER_DIR="cuda_jetpack6" \
+        && cmake --build --parallel ${PARALLEL} --preset 'JetPack 6' \
+        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}

 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -92,12 +121,16 @@ 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-11 dist/lib/ollama/ /lib/ollama/
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/

 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-11 dist/lib/ollama/ /lib/ollama/
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
+COPY --from=jetpack-5 dist/lib/ollama/ /lib/ollama/
+COPY --from=jetpack-6 dist/lib/ollama/ /lib/ollama/

 FROM scratch AS rocm
 COPY --from=rocm-6 dist/lib/ollama /lib/ollama
--- a/Makefile.sync
+++ b/Makefile.sync
@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=e54d41befcc1575f4c898c5ff4ef43970cead75f
+FETCH_HEAD=364a7a6d4a786e98947c8a90430ea581213c0ba9

 .PHONY: help
 help:
--- a/README.md
+++ b/README.md
@@ -411,6 +411,10 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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.)
+- [Serene Pub](https://github.com/doolijb/serene-pub) (Beginner friendly, open source AI Roleplaying App for Windows, Mac OS and Linux. Search, download and use models with Ollama all inside the app.)
+- [Andes](https://github.com/aqerd/andes) (A Visual Studio Code extension that provides a local UI interface for Ollama models)
+- [Clueless](https://github.com/KashyapTan/clueless) (Open Source & Local Cluely: A desktop application LLM assistant to help you talk to anything on your screen using locally served Ollama models. Also undetectable to screenshare)
+- [ollama-co2](https://github.com/carbonatedWaterOrg/ollama-co2) (FastAPI web interface for monitoring and managing local and remote Ollama servers with real-time model monitoring and concurrent downloads)

 ### Cloud

@@ -537,6 +541,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
+- [any-llm](https://github.com/mozilla-ai/any-llm) (A single interface to use different llm providers by [mozilla.ai](https://www.mozilla.ai/))
+- [any-agent](https://github.com/mozilla-ai/any-agent) (A single interface to use and evaluate different agent frameworks by [mozilla.ai](https://www.mozilla.ai/))
+- [Neuro SAN](https://github.com/cognizant-ai-lab/neuro-san-studio) (Data-driven multi-agent orchestration framework) with [example](https://github.com/cognizant-ai-lab/neuro-san-studio/blob/main/docs/user_guide.md#ollama)

 ### Mobile

@@ -597,6 +604,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
+- [NOMYO Router](https://github.com/nomyo-ai/nomyo-router) (A transparent Ollama proxy with model deployment aware routing which auto-manages multiple Ollama instances in a given network)

 ### Supported backends

--- a/api/client.go
+++ b/api/client.go
@@ -45,6 +45,12 @@ func checkError(resp *http.Response, body []byte) error {
 		return nil
 	}

+	if resp.StatusCode == http.StatusUnauthorized {
+		authError := AuthorizationError{StatusCode: resp.StatusCode}
+		json.Unmarshal(body, &authError)
+		return authError
+	}
+
 	apiError := StatusError{StatusCode: resp.StatusCode}

 	err := json.Unmarshal(body, &apiError)
@@ -214,7 +220,8 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	scanner.Buffer(scanBuf, maxBufferSize)
 	for scanner.Scan() {
 		var errorResponse struct {
-			Error string `json:"error,omitempty"`
+			Error     string `json:"error,omitempty"`
+			SigninURL string `json:"signin_url,omitempty"`
 		}

 		bts := scanner.Bytes()
@@ -222,7 +229,13 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			return fmt.Errorf("unmarshal: %w", err)
 		}

-		if response.StatusCode >= http.StatusBadRequest {
+		if response.StatusCode == http.StatusUnauthorized {
+			return AuthorizationError{
+				StatusCode: response.StatusCode,
+				Status:     response.Status,
+				SigninURL:  errorResponse.SigninURL,
+			}
+		} else if response.StatusCode >= http.StatusBadRequest {
 			return StatusError{
 				StatusCode:   response.StatusCode,
 				Status:       response.Status,
@@ -428,3 +441,21 @@ func (c *Client) Version(ctx context.Context) (string, error) {

 	return version.Version, nil
 }
+
+// Signout will signout a client for a local ollama server.
+func (c *Client) Signout(ctx context.Context) error {
+	return c.do(ctx, http.MethodPost, "/api/signout", nil, nil)
+}
+
+// Disconnect will disconnect an ollama instance from ollama.com.
+func (c *Client) Disconnect(ctx context.Context, encodedKey string) error {
+	return c.do(ctx, http.MethodDelete, fmt.Sprintf("/api/user/keys/%s", encodedKey), nil, nil)
+}
+
+func (c *Client) Whoami(ctx context.Context) (*UserResponse, error) {
+	var resp UserResponse
+	if err := c.do(ctx, http.MethodPost, "/api/me", nil, &resp); err != nil {
+		return nil, err
+	}
+	return &resp, nil
+}
--- a/api/types.go
+++ b/api/types.go
@@ -11,6 +11,8 @@ import (
 	"strings"
 	"time"

+	"github.com/google/uuid"
+
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/types/model"
 )
@@ -36,6 +38,19 @@ func (e StatusError) Error() string {
 	}
 }

+type AuthorizationError struct {
+	StatusCode int
+	Status     string
+	SigninURL  string `json:"signin_url"`
+}
+
+func (e AuthorizationError) Error() string {
+	if e.Status != "" {
+		return e.Status
+	}
+	return "something went wrong, please see the ollama server logs for details"
+}
+
 // ImageData represents the raw binary data of an image file.
 type ImageData []byte

@@ -90,6 +105,10 @@ type GenerateRequest struct {
 	// (request that thinking _not_ be used) and unset (use the old behavior
 	// before this option was introduced)
 	Think *ThinkValue `json:"think,omitempty"`
+
+	// DebugRenderOnly is a debug option that, when set to true, returns the rendered
+	// template instead of calling the model.
+	DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
 }

 // ChatRequest describes a request sent by [Client.Chat].
@@ -120,6 +139,10 @@ type ChatRequest struct {
 	// responding. Can be a boolean (true/false) or a string ("high", "medium", "low")
 	// for supported models.
 	Think *ThinkValue `json:"think,omitempty"`
+
+	// DebugRenderOnly is a debug option that, when set to true, returns the rendered
+	// template instead of calling the model.
+	DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
 }

 type Tools []Tool
@@ -278,16 +301,23 @@ func mapToTypeScriptType(jsonType string) string {
 	}
 }

+type ToolFunctionParameters struct {
+	Type       string                  `json:"type"`
+	Defs       any                     `json:"$defs,omitempty"`
+	Items      any                     `json:"items,omitempty"`
+	Required   []string                `json:"required"`
+	Properties map[string]ToolProperty `json:"properties"`
+}
+
+func (t *ToolFunctionParameters) String() string {
+	bts, _ := json.Marshal(t)
+	return string(bts)
+}
+
 type ToolFunction struct {
-	Name        string `json:"name"`
-	Description string `json:"description"`
-	Parameters  struct {
-		Type       string                  `json:"type"`
-		Defs       any                     `json:"$defs,omitempty"`
-		Items      any                     `json:"items,omitempty"`
-		Required   []string                `json:"required"`
-		Properties map[string]ToolProperty `json:"properties"`
-	} `json:"parameters"`
+	Name        string                 `json:"name"`
+	Description string                 `json:"description"`
+	Parameters  ToolFunctionParameters `json:"parameters"`
 }

 func (t *ToolFunction) String() string {
@@ -298,16 +328,38 @@ func (t *ToolFunction) String() string {
 // ChatResponse is the response returned by [Client.Chat]. Its fields are
 // similar to [GenerateResponse].
 type ChatResponse struct {
-	Model      string    `json:"model"`
-	CreatedAt  time.Time `json:"created_at"`
-	Message    Message   `json:"message"`
-	DoneReason string    `json:"done_reason,omitempty"`
+	// Model is the model name that generated the response.
+	Model string `json:"model"`

+	// RemoteModel is the name of the upstream model that generated the response.
+	RemoteModel string `json:"remote_model,omitempty"`
+
+	// RemoteHost is the URL of the upstream Ollama host that generated the response.
+	RemoteHost string `json:"remote_host,omitempty"`
+
+	// CreatedAt is the timestamp of the response.
+	CreatedAt time.Time `json:"created_at"`
+
+	// Message contains the message or part of a message from the model.
+	Message Message `json:"message"`
+
+	// Done specifies if the response is complete.
 	Done bool `json:"done"`

+	// DoneReason is the reason the model stopped generating text.
+	DoneReason string `json:"done_reason,omitempty"`
+
+	DebugInfo *DebugInfo `json:"_debug_info,omitempty"`
+
 	Metrics
 }

+// DebugInfo contains debug information for template rendering
+type DebugInfo struct {
+	RenderedTemplate string `json:"rendered_template"`
+	ImageCount       int    `json:"image_count,omitempty"`
+}
+
 type Metrics struct {
 	TotalDuration      time.Duration `json:"total_duration,omitempty"`
 	LoadDuration       time.Duration `json:"load_duration,omitempty"`
@@ -360,8 +412,12 @@ type EmbedRequest struct {
 	// this request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`

+	// Truncate truncates the input to fit the model's max sequence length.
 	Truncate *bool `json:"truncate,omitempty"`

+	// Dimensions truncates the output embedding to the specified dimension.
+	Dimensions int `json:"dimensions,omitempty"`
+
 	// Options lists model-specific options.
 	Options map[string]any `json:"options"`
 }
@@ -399,18 +455,47 @@ type EmbeddingResponse struct {

 // CreateRequest is the request passed to [Client.Create].
 type CreateRequest struct {
-	Model    string `json:"model"`
-	Stream   *bool  `json:"stream,omitempty"`
+	// Model is the model name to create.
+	Model string `json:"model"`
+
+	// Stream specifies whether the response is streaming; it is true by default.
+	Stream *bool `json:"stream,omitempty"`
+
+	// Quantize is the quantization format for the model; leave blank to not change the quantization level.
 	Quantize string `json:"quantize,omitempty"`

-	From       string            `json:"from,omitempty"`
-	Files      map[string]string `json:"files,omitempty"`
-	Adapters   map[string]string `json:"adapters,omitempty"`
-	Template   string            `json:"template,omitempty"`
-	License    any               `json:"license,omitempty"`
-	System     string            `json:"system,omitempty"`
-	Parameters map[string]any    `json:"parameters,omitempty"`
-	Messages   []Message         `json:"messages,omitempty"`
+	// From is the name of the model or file to use as the source.
+	From string `json:"from,omitempty"`
+
+	// RemoteHost is the URL of the upstream ollama API for the model (if any).
+	RemoteHost string `json:"remote_host,omitempty"`
+
+	// Files is a map of files include when creating the model.
+	Files map[string]string `json:"files,omitempty"`
+
+	// Adapters is a map of LoRA adapters to include when creating the model.
+	Adapters map[string]string `json:"adapters,omitempty"`
+
+	// Template is the template used when constructing a request to the model.
+	Template string `json:"template,omitempty"`
+
+	// License is a string or list of strings for licenses.
+	License any `json:"license,omitempty"`
+
+	// System is the system prompt for the model.
+	System string `json:"system,omitempty"`
+
+	// Parameters is a map of hyper-parameters which are applied to the model.
+	Parameters map[string]any `json:"parameters,omitempty"`
+
+	// Messages is a list of messages added to the model before chat and generation requests.
+	Messages []Message `json:"messages,omitempty"`
+
+	Renderer string `json:"renderer,omitempty"`
+	Parser   string `json:"parser,omitempty"`
+
+	// Info is a map of additional information for the model
+	Info map[string]any `json:"info,omitempty"`

 	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
@@ -448,8 +533,12 @@ type ShowResponse struct {
 	Parameters    string             `json:"parameters,omitempty"`
 	Template      string             `json:"template,omitempty"`
 	System        string             `json:"system,omitempty"`
+	Renderer      string             `json:"renderer,omitempty"`
+	Parser        string             `json:"parser,omitempty"`
 	Details       ModelDetails       `json:"details,omitempty"`
 	Messages      []Message          `json:"messages,omitempty"`
+	RemoteModel   string             `json:"remote_model,omitempty"`
+	RemoteHost    string             `json:"remote_host,omitempty"`
 	ModelInfo     map[string]any     `json:"model_info,omitempty"`
 	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
 	Tensors       []Tensor           `json:"tensors,omitempty"`
@@ -508,12 +597,14 @@ type ProcessResponse struct {

 // ListModelResponse is a single model description in [ListResponse].
 type ListModelResponse struct {
-	Name       string       `json:"name"`
-	Model      string       `json:"model"`
-	ModifiedAt time.Time    `json:"modified_at"`
-	Size       int64        `json:"size"`
-	Digest     string       `json:"digest"`
-	Details    ModelDetails `json:"details,omitempty"`
+	Name        string       `json:"name"`
+	Model       string       `json:"model"`
+	RemoteModel string       `json:"remote_model,omitempty"`
+	RemoteHost  string       `json:"remote_host,omitempty"`
+	ModifiedAt  time.Time    `json:"modified_at"`
+	Size        int64        `json:"size"`
+	Digest      string       `json:"digest"`
+	Details     ModelDetails `json:"details,omitempty"`
 }

 // ProcessModelResponse is a single model description in [ProcessResponse].
@@ -537,6 +628,12 @@ type GenerateResponse struct {
 	// Model is the model name that generated the response.
 	Model string `json:"model"`

+	// RemoteModel is the name of the upstream model that generated the response.
+	RemoteModel string `json:"remote_model,omitempty"`
+
+	// RemoteHost is the URL of the upstream Ollama host that generated the response.
+	RemoteHost string `json:"remote_host,omitempty"`
+
 	// CreatedAt is the timestamp of the response.
 	CreatedAt time.Time `json:"created_at"`

@@ -560,6 +657,8 @@ type GenerateResponse struct {
 	Metrics

 	ToolCalls []ToolCall `json:"tool_calls,omitempty"`
+
+	DebugInfo *DebugInfo `json:"_debug_info,omitempty"`
 }

 // ModelDetails provides details about a model.
@@ -572,6 +671,18 @@ type ModelDetails struct {
 	QuantizationLevel string   `json:"quantization_level"`
 }

+// UserResponse provides information about a user.
+type UserResponse struct {
+	ID        uuid.UUID `json:"id"`
+	Email     string    `json:"email"`
+	Name      string    `json:"name"`
+	Bio       string    `json:"bio,omitempty"`
+	AvatarURL string    `json:"avatarurl,omitempty"`
+	FirstName string    `json:"firstname,omitempty"`
+	LastName  string    `json:"lastname,omitempty"`
+	Plan      string    `json:"plan,omitempty"`
+}
+
 // Tensor describes the metadata for a given tensor.
 type Tensor struct {
 	Name  string   `json:"name"`
@@ -825,7 +936,7 @@ func (t *ThinkValue) UnmarshalJSON(data []byte) error {
 		return nil
 	}

-	return fmt.Errorf("think must be a boolean or string (\"high\", \"medium\", \"low\")")
+	return fmt.Errorf("think must be a boolean or string (\"high\", \"medium\", \"low\", true, or false)")
 }

 // MarshalJSON implements json.Marshaler
@@ -860,7 +971,7 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 		if t < 0 {
 			d.Duration = time.Duration(math.MaxInt64)
 		} else {
-			d.Duration = time.Duration(int(t) * int(time.Second))
+			d.Duration = time.Duration(t * float64(time.Second))
 		}
 	case string:
 		d.Duration, err = time.ParseDuration(t)
--- a/api/types_test.go
+++ b/api/types_test.go
@@ -17,6 +17,11 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 		req  string
 		exp  *Duration
 	}{
+		{
+			name: "Unset",
+			req:  `{ }`,
+			exp:  nil,
+		},
 		{
 			name: "Positive Integer",
 			req:  `{ "keep_alive": 42 }`,
@@ -25,7 +30,7 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 		{
 			name: "Positive Float",
 			req:  `{ "keep_alive": 42.5 }`,
-			exp:  &Duration{42 * time.Second},
+			exp:  &Duration{42500 * time.Millisecond},
 		},
 		{
 			name: "Positive Integer String",
@@ -436,3 +441,50 @@ func TestThinking_UnmarshalJSON(t *testing.T) {
 		})
 	}
 }
+
+func TestToolFunctionParameters_String(t *testing.T) {
+	tests := []struct {
+		name     string
+		params   ToolFunctionParameters
+		expected string
+	}{
+		{
+			name: "simple object with string property",
+			params: ToolFunctionParameters{
+				Type:     "object",
+				Required: []string{"name"},
+				Properties: map[string]ToolProperty{
+					"name": {
+						Type:        PropertyType{"string"},
+						Description: "The name of the person",
+					},
+				},
+			},
+			expected: `{"type":"object","required":["name"],"properties":{"name":{"type":"string","description":"The name of the person"}}}`,
+		},
+		{
+			name: "marshal failure returns empty string",
+			params: ToolFunctionParameters{
+				Type: "object",
+				Defs: func() any {
+					// Create a cycle that will cause json.Marshal to fail
+					type selfRef struct {
+						Self *selfRef
+					}
+					s := &selfRef{}
+					s.Self = s
+					return s
+				}(),
+				Properties: map[string]ToolProperty{},
+			},
+			expected: "",
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			result := test.params.String()
+			assert.Equal(t, test.expected, result)
+		})
+	}
+}
--- a/auth/auth.go
+++ b/auth/auth.go
@@ -18,21 +18,13 @@ import (

 const defaultPrivateKey = "id_ed25519"

-func keyPath() (string, error) {
+func GetPublicKey() (string, error) {
 	home, err := os.UserHomeDir()
 	if err != nil {
 		return "", err
 	}

-	return filepath.Join(home, ".ollama", defaultPrivateKey), nil
-}
-
-func GetPublicKey() (string, error) {
-	keyPath, err := keyPath()
-	if err != nil {
-		return "", err
-	}
-
+	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
 	privateKeyFile, err := os.ReadFile(keyPath)
 	if err != nil {
 		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
@@ -59,11 +51,12 @@ func NewNonce(r io.Reader, length int) (string, error) {
 }

 func Sign(ctx context.Context, bts []byte) (string, error) {
-	keyPath, err := keyPath()
+	home, err := os.UserHomeDir()
 	if err != nil {
 		return "", err
 	}

+	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
 	privateKeyFile, err := os.ReadFile(keyPath)
 	if err != nil {
 		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -47,6 +47,8 @@ import (
 	"github.com/ollama/ollama/version"
 )

+const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
+
 // ensureThinkingSupport emits a warning if the model does not advertise thinking support
 func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
 	if name == "" {
@@ -56,10 +58,8 @@ func ensureThinkingSupport(ctx context.Context, client *api.Client, name string)
 	if err != nil {
 		return
 	}
-	for _, cap := range resp.Capabilities {
-		if cap == model.CapabilityThinking {
-			return
-		}
+	if slices.Contains(resp.Capabilities, model.CapabilityThinking) {
+		return
 	}
 	fmt.Fprintf(os.Stderr, "warning: model %q does not support thinking output\n", name)
 }
@@ -288,7 +288,17 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 		Think: opts.Think,
 	}

-	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
+	return client.Generate(cmd.Context(), req, func(r api.GenerateResponse) error {
+		if r.RemoteModel != "" && opts.ShowConnect {
+			p.StopAndClear()
+			if strings.HasPrefix(r.RemoteHost, "https://ollama.com") {
+				fmt.Fprintf(os.Stderr, "Connecting to '%s' on 'ollama.com' ⚡\n", r.RemoteModel)
+			} else {
+				fmt.Fprintf(os.Stderr, "Connecting to '%s' on '%s'\n", r.RemoteModel, r.RemoteHost)
+			}
+		}
+		return nil
+	})
 }

 func StopHandler(cmd *cobra.Command, args []string) error {
@@ -309,9 +319,10 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	interactive := true

 	opts := runOptions{
-		Model:    args[0],
-		WordWrap: os.Getenv("TERM") == "xterm-256color",
-		Options:  map[string]any{},
+		Model:       args[0],
+		WordWrap:    os.Getenv("TERM") == "xterm-256color",
+		Options:     map[string]any{},
+		ShowConnect: true,
 	}

 	format, err := cmd.Flags().GetString("format")
@@ -369,6 +380,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		}

 		prompts = append([]string{string(in)}, prompts...)
+		opts.ShowConnect = false
 		opts.WordWrap = false
 		interactive = false
 	}
@@ -435,6 +447,15 @@ func RunHandler(cmd *cobra.Command, args []string) error {

 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
+			var sErr api.AuthorizationError
+			if errors.As(err, &sErr) && sErr.StatusCode == http.StatusUnauthorized {
+				fmt.Printf("You need to be signed in to Ollama to run Cloud models.\n\n")
+
+				if sErr.SigninURL != "" {
+					fmt.Printf(ConnectInstructions, sErr.SigninURL)
+				}
+				return nil
+			}
 			return err
 		}

@@ -455,6 +476,59 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	return generate(cmd, opts)
 }

+func SigninHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	user, err := client.Whoami(cmd.Context())
+	if err != nil {
+		var aErr api.AuthorizationError
+		if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+			fmt.Println("You need to be signed in to Ollama to run Cloud models.")
+			fmt.Println()
+
+			if aErr.SigninURL != "" {
+				fmt.Printf(ConnectInstructions, aErr.SigninURL)
+			}
+			return nil
+		}
+		return err
+	}
+
+	if user != nil && user.Name != "" {
+		fmt.Printf("You are already signed in as user '%s'\n", user.Name)
+		fmt.Println()
+		return nil
+	}
+
+	return nil
+}
+
+func SignoutHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	err = client.Signout(cmd.Context())
+	if err != nil {
+		var aErr api.AuthorizationError
+		if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+			fmt.Println("You are not signed in to ollama.com")
+			fmt.Println()
+			return nil
+		} else {
+			return err
+		}
+	}
+
+	fmt.Println("You have signed out of ollama.com")
+	fmt.Println()
+	return nil
+}
+
 func PushHandler(cmd *cobra.Command, args []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -466,6 +540,25 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}

+	n := model.ParseName(args[0])
+	if strings.HasSuffix(n.Host, ".ollama.ai") || strings.HasSuffix(n.Host, ".ollama.com") {
+		_, err := client.Whoami(cmd.Context())
+		if err != nil {
+			var aErr api.AuthorizationError
+			if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+				fmt.Println("You need to be signed in to push models to ollama.com.")
+				fmt.Println()
+
+				if aErr.SigninURL != "" {
+					fmt.Printf(ConnectInstructions, aErr.SigninURL)
+				}
+				return nil
+			}
+
+			return err
+		}
+	}
+
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()

@@ -502,12 +595,12 @@ func PushHandler(cmd *cobra.Command, args []string) error {

 	request := api.PushRequest{Name: args[0], Insecure: insecure}

-	n := model.ParseName(args[0])
 	if err := client.Push(cmd.Context(), &request, fn); err != nil {
 		if spinner != nil {
 			spinner.Stop()
 		}
-		if strings.Contains(err.Error(), "access denied") {
+		errStr := strings.ToLower(err.Error())
+		if strings.Contains(errStr, "access denied") || strings.Contains(errStr, "unauthorized") {
 			return errors.New("you are not authorized to push to this namespace, create the model under a namespace you own")
 		}
 		return err
@@ -541,7 +634,14 @@ func ListHandler(cmd *cobra.Command, args []string) error {

 	for _, m := range models.Models {
 		if len(args) == 0 || strings.HasPrefix(strings.ToLower(m.Name), strings.ToLower(args[0])) {
-			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), format.HumanTime(m.ModifiedAt, "Never")})
+			var size string
+			if m.RemoteModel != "" {
+				size = "-"
+			} else {
+				size = format.HumanBytes(m.Size)
+			}
+
+			data = append(data, []string{m.Name, m.Digest[:12], size, format.HumanTime(m.ModifiedAt, "Never")})
 		}
 	}

@@ -626,8 +726,8 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 		KeepAlive: &api.Duration{Duration: 0},
 	}
 	if err := loadOrUnloadModel(cmd, opts); err != nil {
-		if !strings.Contains(err.Error(), "not found") {
-			return fmt.Errorf("unable to stop existing running model \"%s\": %s", args[0], err)
+		if !strings.Contains(strings.ToLower(err.Error()), "not found") {
+			fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", args[0])
 		}
 	}

@@ -738,12 +838,36 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 	}

 	tableRender("Model", func() (rows [][]string) {
+		if resp.RemoteHost != "" {
+			rows = append(rows, []string{"", "Remote model", resp.RemoteModel})
+			rows = append(rows, []string{"", "Remote URL", resp.RemoteHost})
+		}
+
 		if resp.ModelInfo != nil {
 			arch := resp.ModelInfo["general.architecture"].(string)
 			rows = append(rows, []string{"", "architecture", arch})
-			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ModelInfo["general.parameter_count"].(float64)))})
-			rows = append(rows, []string{"", "context length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64), 'f', -1, 64)})
-			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64), 'f', -1, 64)})
+
+			var paramStr string
+			if resp.Details.ParameterSize != "" {
+				paramStr = resp.Details.ParameterSize
+			} else if v, ok := resp.ModelInfo["general.parameter_count"]; ok {
+				if f, ok := v.(float64); ok {
+					paramStr = format.HumanNumber(uint64(f))
+				}
+			}
+			rows = append(rows, []string{"", "parameters", paramStr})
+
+			if v, ok := resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)]; ok {
+				if f, ok := v.(float64); ok {
+					rows = append(rows, []string{"", "context length", strconv.FormatFloat(f, 'f', -1, 64)})
+				}
+			}
+
+			if v, ok := resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)]; ok {
+				if f, ok := v.(float64); ok {
+					rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(f, 'f', -1, 64)})
+				}
+			}
 		} else {
 			rows = append(rows, []string{"", "architecture", resp.Details.Family})
 			rows = append(rows, []string{"", "parameters", resp.Details.ParameterSize})
@@ -991,6 +1115,52 @@ type runOptions struct {
 	KeepAlive    *api.Duration
 	Think        *api.ThinkValue
 	HideThinking bool
+	ShowConnect  bool
+}
+
+func (r runOptions) Copy() runOptions {
+	var messages []api.Message
+	if r.Messages != nil {
+		messages = make([]api.Message, len(r.Messages))
+		copy(messages, r.Messages)
+	}
+
+	var images []api.ImageData
+	if r.Images != nil {
+		images = make([]api.ImageData, len(r.Images))
+		copy(images, r.Images)
+	}
+
+	var opts map[string]any
+	if r.Options != nil {
+		opts = make(map[string]any, len(r.Options))
+		for k, v := range r.Options {
+			opts[k] = v
+		}
+	}
+
+	var think *api.ThinkValue
+	if r.Think != nil {
+		cThink := *r.Think
+		think = &cThink
+	}
+
+	return runOptions{
+		Model:        r.Model,
+		ParentModel:  r.ParentModel,
+		Prompt:       r.Prompt,
+		Messages:     messages,
+		WordWrap:     r.WordWrap,
+		Format:       r.Format,
+		System:       r.System,
+		Images:       images,
+		Options:      opts,
+		MultiModal:   r.MultiModal,
+		KeepAlive:    r.KeepAlive,
+		Think:        think,
+		HideThinking: r.HideThinking,
+		ShowConnect:  r.ShowConnect,
+	}
 }

 type displayResponseState struct {
@@ -1546,6 +1716,22 @@ func NewCLI() *cobra.Command {

 	pushCmd.Flags().Bool("insecure", false, "Use an insecure registry")

+	signinCmd := &cobra.Command{
+		Use:     "signin",
+		Short:   "Sign in to ollama.com",
+		Args:    cobra.ExactArgs(0),
+		PreRunE: checkServerHeartbeat,
+		RunE:    SigninHandler,
+	}
+
+	signoutCmd := &cobra.Command{
+		Use:     "signout",
+		Short:   "Sign out from ollama.com",
+		Args:    cobra.ExactArgs(0),
+		PreRunE: checkServerHeartbeat,
+		RunE:    SignoutHandler,
+	}
+
 	listCmd := &cobra.Command{
 		Use:     "list",
 		Aliases: []string{"ls"},
@@ -1640,6 +1826,8 @@ func NewCLI() *cobra.Command {
 		stopCmd,
 		pullCmd,
 		pushCmd,
+		signinCmd,
+		signoutCmd,
 		listCmd,
 		psCmd,
 		copyCmd,
--- a/cmd/cmd_test.go
+++ b/cmd/cmd_test.go
@@ -3,10 +3,12 @@ package cmd
 import (
 	"bytes"
 	"encoding/json"
+	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
 	"os"
+	"reflect"
 	"strings"
 	"testing"
 	"time"
@@ -304,6 +306,8 @@ func TestDeleteHandler(t *testing.T) {
 				w.WriteHeader(http.StatusOK)
 			} else {
 				w.WriteHeader(http.StatusNotFound)
+				errPayload := `{"error":"model '%s' not found"}`
+				w.Write([]byte(fmt.Sprintf(errPayload, req.Name)))
 			}
 			return
 		}
@@ -346,7 +350,7 @@ func TestDeleteHandler(t *testing.T) {
 	}

 	err := DeleteHandler(cmd, []string{"test-model-not-found"})
-	if err == nil || !strings.Contains(err.Error(), "unable to stop existing running model \"test-model-not-found\"") {
+	if err == nil || !strings.Contains(err.Error(), "model 'test-model-not-found' not found") {
 		t.Fatalf("DeleteHandler failed: expected error about stopping non-existent model, got %v", err)
 	}
 }
@@ -488,9 +492,35 @@ func TestPushHandler(t *testing.T) {
 						w.(http.Flusher).Flush()
 					}
 				},
+				"/api/me": func(w http.ResponseWriter, r *http.Request) {
+					if r.Method != http.MethodPost {
+						t.Errorf("expected POST request, got %s", r.Method)
+					}
+				},
 			},
 			expectedOutput: "\nYou can find your model at:\n\n\thttps://ollama.com/test-model\n",
 		},
+		{
+			name:      "not signed in push",
+			modelName: "notsignedin-model",
+			serverResponse: map[string]func(w http.ResponseWriter, r *http.Request){
+				"/api/me": func(w http.ResponseWriter, r *http.Request) {
+					if r.Method != http.MethodPost {
+						t.Errorf("expected POST request, got %s", r.Method)
+					}
+					w.Header().Set("Content-Type", "application/json")
+					w.WriteHeader(http.StatusUnauthorized)
+					err := json.NewEncoder(w).Encode(map[string]string{
+						"error":      "unauthorized",
+						"signin_url": "https://somethingsomething",
+					})
+					if err != nil {
+						t.Fatal(err)
+					}
+				},
+			},
+			expectedOutput: "You need to be signed in to push",
+		},
 		{
 			name:      "unauthorized push",
 			modelName: "unauthorized-model",
@@ -499,12 +529,17 @@ func TestPushHandler(t *testing.T) {
 					w.Header().Set("Content-Type", "application/json")
 					w.WriteHeader(http.StatusUnauthorized)
 					err := json.NewEncoder(w).Encode(map[string]string{
-						"error": "access denied",
+						"error": "403: {\"errors\":[{\"code\":\"ACCESS DENIED\", \"message\":\"access denied\"}]}",
 					})
 					if err != nil {
 						t.Fatal(err)
 					}
 				},
+				"/api/me": func(w http.ResponseWriter, r *http.Request) {
+					if r.Method != http.MethodPost {
+						t.Errorf("expected POST request, got %s", r.Method)
+					}
+				},
 			},
 			expectedError: "you are not authorized to push to this namespace, create the model under a namespace you own",
 		},
@@ -522,6 +557,10 @@ func TestPushHandler(t *testing.T) {
 			defer mockServer.Close()

 			t.Setenv("OLLAMA_HOST", mockServer.URL)
+			tmpDir := t.TempDir()
+			t.Setenv("HOME", tmpDir)
+			t.Setenv("USERPROFILE", tmpDir)
+			initializeKeypair()

 			cmd := &cobra.Command{}
 			cmd.Flags().Bool("insecure", false, "")
@@ -557,7 +596,7 @@ func TestPushHandler(t *testing.T) {
 					t.Errorf("expected no error, got %v", err)
 				}
 				if tt.expectedOutput != "" {
-					if got := string(stdout); got != tt.expectedOutput {
+					if got := string(stdout); !strings.Contains(got, tt.expectedOutput) {
 						t.Errorf("expected output %q, got %q", tt.expectedOutput, got)
 					}
 				}
@@ -915,3 +954,286 @@ func TestNewCreateRequest(t *testing.T) {
 		})
 	}
 }
+
+func TestRunOptions_Copy(t *testing.T) {
+	// Setup test data
+	originalKeepAlive := &api.Duration{Duration: 5 * time.Minute}
+	originalThink := &api.ThinkValue{Value: "test reasoning"}
+
+	original := runOptions{
+		Model:       "test-model",
+		ParentModel: "parent-model",
+		Prompt:      "test prompt",
+		Messages: []api.Message{
+			{Role: "user", Content: "hello"},
+			{Role: "assistant", Content: "hi there"},
+		},
+		WordWrap: true,
+		Format:   "json",
+		System:   "system prompt",
+		Images: []api.ImageData{
+			[]byte("image1"),
+			[]byte("image2"),
+		},
+		Options: map[string]any{
+			"temperature": 0.7,
+			"max_tokens":  1000,
+			"top_p":       0.9,
+		},
+		MultiModal:   true,
+		KeepAlive:    originalKeepAlive,
+		Think:        originalThink,
+		HideThinking: false,
+		ShowConnect:  true,
+	}
+
+	// Test the copy
+	copied := original.Copy()
+
+	// Test 1: Verify the copy is not the same instance
+	if &copied == &original {
+		t.Error("Copy should return a different instance")
+	}
+
+	// Test 2: Verify all fields are copied correctly
+	tests := []struct {
+		name string
+		got  interface{}
+		want interface{}
+	}{
+		{"Model", copied.Model, original.Model},
+		{"ParentModel", copied.ParentModel, original.ParentModel},
+		{"Prompt", copied.Prompt, original.Prompt},
+		{"WordWrap", copied.WordWrap, original.WordWrap},
+		{"Format", copied.Format, original.Format},
+		{"System", copied.System, original.System},
+		{"MultiModal", copied.MultiModal, original.MultiModal},
+		{"HideThinking", copied.HideThinking, original.HideThinking},
+		{"ShowConnect", copied.ShowConnect, original.ShowConnect},
+	}
+
+	for _, tt := range tests {
+		if !reflect.DeepEqual(tt.got, tt.want) {
+			t.Errorf("%s mismatch: got %v, want %v", tt.name, tt.got, tt.want)
+		}
+	}
+
+	// Test 3: Verify Messages slice is deeply copied
+	if len(copied.Messages) != len(original.Messages) {
+		t.Errorf("Messages length mismatch: got %d, want %d", len(copied.Messages), len(original.Messages))
+	}
+
+	if len(copied.Messages) > 0 && &copied.Messages[0] == &original.Messages[0] {
+		t.Error("Messages should be different instances")
+	}
+
+	// Modify original to verify independence
+	if len(original.Messages) > 0 {
+		originalContent := original.Messages[0].Content
+		original.Messages[0].Content = "modified"
+		if len(copied.Messages) > 0 && copied.Messages[0].Content == "modified" {
+			t.Error("Messages should be independent after copy")
+		}
+		// Restore for other tests
+		original.Messages[0].Content = originalContent
+	}
+
+	// Test 4: Verify Images slice is deeply copied
+	if len(copied.Images) != len(original.Images) {
+		t.Errorf("Images length mismatch: got %d, want %d", len(copied.Images), len(original.Images))
+	}
+
+	if len(copied.Images) > 0 && &copied.Images[0] == &original.Images[0] {
+		t.Error("Images should be different instances")
+	}
+
+	// Modify original to verify independence
+	if len(original.Images) > 0 {
+		originalImage := original.Images[0]
+		original.Images[0] = []byte("modified")
+		if len(copied.Images) > 0 && string(copied.Images[0]) == "modified" {
+			t.Error("Images should be independent after copy")
+		}
+		// Restore for other tests
+		original.Images[0] = originalImage
+	}
+
+	// Test 5: Verify Options map is deeply copied
+	if len(copied.Options) != len(original.Options) {
+		t.Errorf("Options length mismatch: got %d, want %d", len(copied.Options), len(original.Options))
+	}
+
+	if len(copied.Options) > 0 && &copied.Options == &original.Options {
+		t.Error("Options map should be different instances")
+	}
+
+	// Modify original to verify independence
+	if len(original.Options) > 0 {
+		originalTemp := original.Options["temperature"]
+		original.Options["temperature"] = 0.9
+		if copied.Options["temperature"] == 0.9 {
+			t.Error("Options should be independent after copy")
+		}
+		// Restore for other tests
+		original.Options["temperature"] = originalTemp
+	}
+
+	// Test 6: Verify KeepAlive pointer is copied (shallow copy)
+	if copied.KeepAlive != original.KeepAlive {
+		t.Error("KeepAlive pointer should be the same (shallow copy)")
+	}
+
+	// Test 7: Verify Think pointer creates a new instance
+	if original.Think != nil && copied.Think == original.Think {
+		t.Error("Think should be a different instance")
+	}
+
+	if original.Think != nil && copied.Think != nil {
+		if !reflect.DeepEqual(copied.Think.Value, original.Think.Value) {
+			t.Errorf("Think.Value mismatch: got %v, want %v", copied.Think.Value, original.Think.Value)
+		}
+	}
+
+	// Test 8: Test with zero values
+	zeroOriginal := runOptions{}
+	zeroCopy := zeroOriginal.Copy()
+
+	if !reflect.DeepEqual(zeroCopy, zeroOriginal) {
+		fmt.Printf("orig: %#v\ncopy: %#v\n", zeroOriginal, zeroCopy)
+		t.Error("Copy of zero value should equal original zero value")
+	}
+}
+
+func TestRunOptions_Copy_EmptySlicesAndMaps(t *testing.T) {
+	// Test with empty slices and maps
+	original := runOptions{
+		Messages: []api.Message{},
+		Images:   []api.ImageData{},
+		Options:  map[string]any{},
+	}
+
+	copied := original.Copy()
+
+	if copied.Messages == nil {
+		t.Error("Empty Messages slice should remain empty, not nil")
+	}
+
+	if copied.Images == nil {
+		t.Error("Empty Images slice should remain empty, not nil")
+	}
+
+	if copied.Options == nil {
+		t.Error("Empty Options map should remain empty, not nil")
+	}
+
+	if len(copied.Messages) != 0 {
+		t.Error("Empty Messages slice should remain empty")
+	}
+
+	if len(copied.Images) != 0 {
+		t.Error("Empty Images slice should remain empty")
+	}
+
+	if len(copied.Options) != 0 {
+		t.Error("Empty Options map should remain empty")
+	}
+}
+
+func TestRunOptions_Copy_NilPointers(t *testing.T) {
+	// Test with nil pointers
+	original := runOptions{
+		KeepAlive: nil,
+		Think:     nil,
+	}
+
+	copied := original.Copy()
+
+	if copied.KeepAlive != nil {
+		t.Error("Nil KeepAlive should remain nil")
+	}
+
+	if copied.Think != nil {
+		t.Error("Nil Think should remain nil")
+	}
+}
+
+func TestRunOptions_Copy_ThinkValueVariants(t *testing.T) {
+	tests := []struct {
+		name  string
+		think *api.ThinkValue
+	}{
+		{"nil Think", nil},
+		{"bool true", &api.ThinkValue{Value: true}},
+		{"bool false", &api.ThinkValue{Value: false}},
+		{"string value", &api.ThinkValue{Value: "reasoning text"}},
+		{"int value", &api.ThinkValue{Value: 42}},
+		{"nil value", &api.ThinkValue{Value: nil}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			original := runOptions{Think: tt.think}
+			copied := original.Copy()
+
+			if tt.think == nil {
+				if copied.Think != nil {
+					t.Error("Nil Think should remain nil")
+				}
+				return
+			}
+
+			if copied.Think == nil {
+				t.Error("Non-nil Think should not become nil")
+				return
+			}
+
+			if copied.Think == original.Think {
+				t.Error("Think should be a different instance")
+			}
+
+			if !reflect.DeepEqual(copied.Think.Value, original.Think.Value) {
+				t.Errorf("Think.Value mismatch: got %v, want %v", copied.Think.Value, original.Think.Value)
+			}
+		})
+	}
+}
+
+func TestRunOptions_Copy_Independence(t *testing.T) {
+	// Test that modifications to original don't affect copy
+	originalThink := &api.ThinkValue{Value: "original"}
+	original := runOptions{
+		Model:    "original-model",
+		Messages: []api.Message{{Role: "user", Content: "original"}},
+		Options:  map[string]any{"key": "value"},
+		Think:    originalThink,
+	}
+
+	copied := original.Copy()
+
+	// Modify original
+	original.Model = "modified-model"
+	if len(original.Messages) > 0 {
+		original.Messages[0].Content = "modified"
+	}
+	original.Options["key"] = "modified"
+	if original.Think != nil {
+		original.Think.Value = "modified"
+	}
+
+	// Verify copy is unchanged
+	if copied.Model == "modified-model" {
+		t.Error("Copy Model should not be affected by original modification")
+	}
+
+	if len(copied.Messages) > 0 && copied.Messages[0].Content == "modified" {
+		t.Error("Copy Messages should not be affected by original modification")
+	}
+
+	if copied.Options["key"] == "modified" {
+		t.Error("Copy Options should not be affected by original modification")
+	}
+
+	if copied.Think != nil && copied.Think.Value == "modified" {
+		t.Error("Copy Think should not be affected by original modification")
+	}
+}
--- a/cmd/interactive.go
+++ b/cmd/interactive.go
@@ -195,16 +195,24 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				fmt.Println("Usage:\n  /load <modelname>")
 				continue
 			}
+			origOpts := opts.Copy()
+
 			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 {
+				if strings.Contains(err.Error(), "not found") {
+					fmt.Printf("Couldn't find model '%s'\n", opts.Model)
+					opts = origOpts.Copy()
+					continue
+				}
 				return err
 			}
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				if strings.Contains(err.Error(), "not found") {
-					fmt.Printf("error: %v\n", err)
+					fmt.Printf("Couldn't find model '%s'\n", opts.Model)
+					opts = origOpts.Copy()
 					continue
 				}
 				if strings.Contains(err.Error(), "does not support thinking") {
--- a/convert/convert_bert.go
+++ b/convert/convert_bert.go
@@ -28,6 +28,7 @@ type bertModel struct {
 	LayerNormEPS          float32 `json:"layer_norm_eps"`
 	LayerNormEpsilon      float32 `json:"layer_norm_epsilon"`
 	NormEpsilon           float32 `json:"norm_epsilon"`
+	normalizeEmbeddings   bool

 	PoolingType uint32
 }
@@ -54,9 +55,11 @@ func (p *bertModel) parseMore(fsys fs.FS) error {

 	var pooling string
 	for _, m := range modules {
-		if m.Type == "sentence_transformers.models.Pooling" {
+		switch m.Type {
+		case "sentence_transformers.models.Pooling":
 			pooling = m.Path
-			break
+		case "sentence_transformers.models.Normalize":
+			p.normalizeEmbeddings = true
 		}
 	}

@@ -90,6 +93,7 @@ func (p *bertModel) KV(t *Tokenizer) ggml.KV {
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false
 	kv["bert.pooling_type"] = p.PoolingType
+	kv["bert.normalize_embeddings"] = p.normalizeEmbeddings

 	kv["bert.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)

--- a/convert/convert_gptoss.go
+++ b/convert/convert_gptoss.go
@@ -15,19 +15,24 @@ import (

 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"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	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"`
+	LocalExperts          uint32  `json:"num_local_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"`
+	RopeScaling           struct {
+		Factor float32 `json:"factor"`
+	} `json:"rope_scaling"`
+	SlidingWindow uint32 `json:"sliding_window"`
 }

 var _ ModelConverter = (*gptossModel)(nil)
@@ -36,11 +41,11 @@ 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.context_length"] = cmp.Or(m.MaxPositionEmbeddings, 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_count"] = cmp.Or(m.Experts, m.LocalExperts)
 	kv["gptoss.expert_used_count"] = m.ExpertsPerToken
 	kv["gptoss.attention.head_count"] = m.AttentionHeads
 	kv["gptoss.attention.head_count_kv"] = m.KeyValueHeads
@@ -49,7 +54,7 @@ func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
 	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.factor"] = cmp.Or(m.RopeScalingFactor, m.RopeScaling.Factor)
 	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
@@ -92,6 +97,11 @@ func (m *gptossModel) Tensors(ts []Tensor) []*ggml.Tensor {

 	for name, mxfp4 := range mxfp4s {
 		dims := mxfp4.blocks.Shape()
+
+		if !strings.HasSuffix(name, ".weight") {
+			name += ".weight"
+		}
+
 		out = append(out, &ggml.Tensor{
 			Name:     name,
 			Kind:     uint32(ggml.TensorTypeMXFP4),
@@ -104,25 +114,47 @@ func (m *gptossModel) Tensors(ts []Tensor) []*ggml.Tensor {
 }

 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",
+	var replacements []string
+	if m.MaxPositionEmbeddings > 0 {
+		// hf flavored model
+		replacements = []string{
+			"lm_head", "output",
+			"model.embed_tokens", "token_embd",
+			"model.layers", "blk",
+			"input_layernorm", "attn_norm",
+			"self_attn.q_proj", "attn_q",
+			"self_attn.k_proj", "attn_k",
+			"self_attn.v_proj", "attn_v",
+			"self_attn.o_proj", "attn_out",
+			"self_attn.sinks", "attn_sinks",
+			"post_attention_layernorm", "ffn_norm",
+			"mlp.router", "ffn_gate_inp",
+			"mlp.experts.gate_up_proj_", "ffn_gate_up_exps.",
+			"mlp.experts.down_proj_", "ffn_down_exps.",
+			"model.norm", "output_norm",
+		}
+	} else {
+		replacements = []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",
+		}
 	}
+	return replacements
 }

 type mxfp4 struct {
@@ -140,7 +172,20 @@ func (m *mxfp4) WriteTo(w io.Writer) (int64, error) {
 		blocksDims[i] = int(d)
 	}

-	var blocks tensor.Tensor = tensor.New(tensor.WithShape(blocksDims...), tensor.WithBacking(b.Bytes()))
+	bts := b.Bytes()
+	var tmp [16]byte
+	for i := 0; i < b.Len(); i += 16 {
+		for j := range 8 {
+			// transform a1b2c3 ... x7y8z9 -> 71xa82yb93zc
+			a, b := bts[i+j], bts[i+j+8]
+			tmp[2*j+0] = (a & 0x0F) | (b << 4)
+			tmp[2*j+1] = (a >> 4) | (b & 0xF0)
+		}
+
+		copy(bts[i:i+16], tmp[:])
+	}
+
+	var blocks tensor.Tensor = tensor.New(tensor.WithShape(blocksDims...), tensor.WithBacking(bts))

 	var s bytes.Buffer
 	if _, err := m.scales.WriteTo(&s); err != nil {
@@ -174,5 +219,5 @@ func (m *mxfp4) WriteTo(w io.Writer) (int64, error) {
 		return 0, err
 	}

-	return 0, nil
+	return int64(len(u8s)), nil
 }
--- a/convert/reader.go
+++ b/convert/reader.go
@@ -33,8 +33,8 @@ func (t tensorBase) Shape() []uint64 {
 const (
 	tensorKindFP32 uint32 = iota
 	tensorKindFP16
-	tensorKindMXFP4 = 4
 	tensorKindBF16  = 30
+	tensorKindMXFP4 = 39
 )

 func (t tensorBase) Kind() uint32 {
--- a/convert/reader_safetensors.go
+++ b/convert/reader_safetensors.go
@@ -96,7 +96,7 @@ type safetensor struct {

 func (st safetensor) Kind() uint32 {
 	kind := st.tensorBase.Kind()
-	if st.dtype == "BF16" && kind != tensorKindFP32 {
+	if !strings.HasPrefix(st.name, "v.") && st.dtype == "BF16" && kind != tensorKindFP32 {
 		kind = tensorKindBF16
 	}

@@ -188,17 +188,17 @@ func (st safetensor) WriteTo(w io.Writer) (int64, error) {

 	switch st.Kind() {
 	case tensorKindFP32:
-		return 0, binary.Write(w, binary.LittleEndian, f32s)
+		return int64(len(f32s) * 4), binary.Write(w, binary.LittleEndian, f32s)
 	case tensorKindFP16:
 		f16s := make([]uint16, len(f32s))
 		for i := range f32s {
 			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
 		}

-		return 0, binary.Write(w, binary.LittleEndian, f16s)
+		return int64(len(f16s) * 2), binary.Write(w, binary.LittleEndian, f16s)
 	case tensorKindBF16:
 		u8s := bfloat16.EncodeFloat32(f32s)
-		return 0, binary.Write(w, binary.LittleEndian, u8s)
+		return int64(len(u8s)), 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
@@ -230,3 +230,65 @@ func TestSafetensors(t *testing.T) {
 		})
 	}
 }
+
+func TestSafetensorKind(t *testing.T) {
+	tests := []struct {
+		name     string
+		st       safetensor
+		expected uint32
+	}{
+		{
+			name: "BF16 dtype with non-v. prefix and non-FP32 base kind should return BF16",
+			st: safetensor{
+				tensorBase: &tensorBase{
+					name:  "weight.matrix",
+					shape: []uint64{10, 10}, // will default to FP16
+				},
+				dtype: "BF16",
+			},
+			expected: tensorKindBF16,
+		},
+		{
+			name: "BF16 dtype with v. prefix should return base kind",
+			st: safetensor{
+				tensorBase: &tensorBase{
+					name:  "v.weight.matrix",
+					shape: []uint64{10, 10}, // will default to FP16
+				},
+				dtype: "BF16",
+			},
+			expected: tensorKindFP16,
+		},
+		{
+			name: "BF16 dtype with FP32 base kind should return FP32",
+			st: safetensor{
+				tensorBase: &tensorBase{
+					name:  "weight.matrix",
+					shape: []uint64{10}, // will default to FP32
+				},
+				dtype: "BF16",
+			},
+			expected: tensorKindFP32,
+		},
+		{
+			name: "Non-BF16 dtype should return base kind",
+			st: safetensor{
+				tensorBase: &tensorBase{
+					name:  "weight.matrix",
+					shape: []uint64{10, 10}, // will default to FP16
+				},
+				dtype: "FP16",
+			},
+			expected: tensorKindFP16,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := tt.st.Kind()
+			if result != tt.expected {
+				t.Errorf("Kind() = %d, expected %d", result, tt.expected)
+			}
+		})
+	}
+}
--- a/discover/amd_common.go
+++ b/discover/amd_common.go
@@ -1,83 +0,0 @@
-//go:build linux || windows
-
-package discover
-
-import (
-	"errors"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"runtime"
-	"strings"
-)
-
-// Determine if the given ROCm lib directory is usable by checking for existence of some glob patterns
-func rocmLibUsable(libDir string) bool {
-	slog.Debug("evaluating potential rocm lib dir " + libDir)
-	for _, g := range ROCmLibGlobs {
-		res, _ := filepath.Glob(filepath.Join(libDir, g))
-		if len(res) == 0 {
-			return false
-		}
-	}
-	return true
-}
-
-func GetSupportedGFX(libDir string) ([]string, error) {
-	var ret []string
-	files, err := filepath.Glob(filepath.Join(libDir, "rocblas", "library", "TensileLibrary_lazy_gfx*.dat"))
-	if err != nil {
-		return nil, err
-	}
-	for _, file := range files {
-		ret = append(ret, strings.TrimSuffix(strings.TrimPrefix(filepath.Base(file), "TensileLibrary_lazy_"), ".dat"))
-	}
-	return ret, nil
-}
-
-func commonAMDValidateLibDir() (string, error) {
-	// Favor our bundled version
-
-	// Installer payload location if we're running the installed binary
-	rocmTargetDir := filepath.Join(LibOllamaPath, "rocm")
-	if rocmLibUsable(rocmTargetDir) {
-		slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-		return rocmTargetDir, nil
-	}
-
-	// Prefer explicit HIP env var
-	hipPath := os.Getenv("HIP_PATH")
-	if hipPath != "" {
-		hipLibDir := filepath.Join(hipPath, "bin")
-		if rocmLibUsable(hipLibDir) {
-			slog.Debug("detected ROCM via HIP_PATH=" + hipPath)
-			return hipLibDir, nil
-		}
-	}
-
-	// Scan the LD_LIBRARY_PATH or PATH
-	pathEnv := "LD_LIBRARY_PATH"
-	if runtime.GOOS == "windows" {
-		pathEnv = "PATH"
-	}
-
-	paths := os.Getenv(pathEnv)
-	for _, path := range filepath.SplitList(paths) {
-		d, err := filepath.Abs(path)
-		if err != nil {
-			continue
-		}
-		if rocmLibUsable(d) {
-			return d, nil
-		}
-	}
-
-	// Well known location(s)
-	for _, path := range RocmStandardLocations {
-		if rocmLibUsable(path) {
-			return path, nil
-		}
-	}
-
-	return "", errors.New("no suitable rocm found, falling back to CPU")
-}
--- a/discover/amd_hip_windows.go
+++ b/discover/amd_hip_windows.go
@@ -1,147 +0,0 @@
-package discover
-
-import (
-	"errors"
-	"fmt"
-	"log/slog"
-	"syscall"
-	"unsafe"
-
-	"golang.org/x/sys/windows"
-)
-
-const (
-	hipSuccess       = 0
-	hipErrorNoDevice = 100
-)
-
-type hipDevicePropMinimal struct {
-	Name        [256]byte
-	unused1     [140]byte
-	GcnArchName [256]byte // gfx####
-	iGPU        int       // Doesn't seem to actually report correctly
-	unused2     [128]byte
-}
-
-// Wrap the amdhip64.dll library for GPU discovery
-type HipLib struct {
-	dll                    windows.Handle
-	hipGetDeviceCount      uintptr
-	hipGetDeviceProperties uintptr
-	hipMemGetInfo          uintptr
-	hipSetDevice           uintptr
-	hipDriverGetVersion    uintptr
-}
-
-func NewHipLib() (*HipLib, error) {
-	// At runtime we depend on v6, so discover GPUs with the same library for a consistent set of GPUs
-	h, err := windows.LoadLibrary("amdhip64_6.dll")
-	if err != nil {
-		return nil, fmt.Errorf("unable to load amdhip64_6.dll, please make sure to upgrade to the latest amd driver: %w", err)
-	}
-	hl := &HipLib{}
-	hl.dll = h
-	hl.hipGetDeviceCount, err = windows.GetProcAddress(hl.dll, "hipGetDeviceCount")
-	if err != nil {
-		return nil, err
-	}
-	hl.hipGetDeviceProperties, err = windows.GetProcAddress(hl.dll, "hipGetDeviceProperties")
-	if err != nil {
-		return nil, err
-	}
-	hl.hipMemGetInfo, err = windows.GetProcAddress(hl.dll, "hipMemGetInfo")
-	if err != nil {
-		return nil, err
-	}
-	hl.hipSetDevice, err = windows.GetProcAddress(hl.dll, "hipSetDevice")
-	if err != nil {
-		return nil, err
-	}
-	hl.hipDriverGetVersion, err = windows.GetProcAddress(hl.dll, "hipDriverGetVersion")
-	if err != nil {
-		return nil, err
-	}
-	return hl, nil
-}
-
-// The hip library only evaluates the ROCR_VISIBLE_DEVICES variable at startup
-// so we have to unload/reset the library after we do our initial discovery
-// to make sure our updates to that variable are processed by llama.cpp
-func (hl *HipLib) Release() {
-	err := windows.FreeLibrary(hl.dll)
-	if err != nil {
-		slog.Warn("failed to unload amdhip64.dll", "error", err)
-	}
-	hl.dll = 0
-}
-
-func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
-	if hl.dll == 0 {
-		return 0, 0, errors.New("dll has been unloaded")
-	}
-	var version int
-	status, _, err := syscall.SyscallN(hl.hipDriverGetVersion, uintptr(unsafe.Pointer(&version)))
-	if status != hipSuccess {
-		return 0, 0, fmt.Errorf("failed call to hipDriverGetVersion: %d %s", status, err)
-	}
-
-	slog.Debug("hipDriverGetVersion", "version", version)
-	driverMajor = version / 10000000
-	driverMinor = (version - (driverMajor * 10000000)) / 100000
-
-	return driverMajor, driverMinor, nil
-}
-
-func (hl *HipLib) HipGetDeviceCount() int {
-	if hl.dll == 0 {
-		slog.Error("dll has been unloaded")
-		return 0
-	}
-	var count int
-	status, _, err := syscall.SyscallN(hl.hipGetDeviceCount, uintptr(unsafe.Pointer(&count)))
-	if status == hipErrorNoDevice {
-		slog.Info("AMD ROCm reports no devices found")
-		return 0
-	}
-	if status != hipSuccess {
-		slog.Warn("failed call to hipGetDeviceCount", "status", status, "error", err)
-	}
-	return count
-}
-
-func (hl *HipLib) HipSetDevice(device int) error {
-	if hl.dll == 0 {
-		return errors.New("dll has been unloaded")
-	}
-	status, _, err := syscall.SyscallN(hl.hipSetDevice, uintptr(device))
-	if status != hipSuccess {
-		return fmt.Errorf("failed call to hipSetDevice: %d %s", status, err)
-	}
-	return nil
-}
-
-func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, error) {
-	if hl.dll == 0 {
-		return nil, errors.New("dll has been unloaded")
-	}
-	var props hipDevicePropMinimal
-	status, _, err := syscall.SyscallN(hl.hipGetDeviceProperties, uintptr(unsafe.Pointer(&props)), uintptr(device))
-	if status != hipSuccess {
-		return nil, fmt.Errorf("failed call to hipGetDeviceProperties: %d %s", status, err)
-	}
-	return &props, nil
-}
-
-// free, total, err
-func (hl *HipLib) HipMemGetInfo() (uint64, uint64, error) {
-	if hl.dll == 0 {
-		return 0, 0, errors.New("dll has been unloaded")
-	}
-	var totalMemory uint64
-	var freeMemory uint64
-	status, _, err := syscall.SyscallN(hl.hipMemGetInfo, uintptr(unsafe.Pointer(&freeMemory)), uintptr(unsafe.Pointer(&totalMemory)))
-	if status != hipSuccess {
-		return 0, 0, fmt.Errorf("failed call to hipMemGetInfo: %d %s", status, err)
-	}
-	return freeMemory, totalMemory, nil
-}
--- a/discover/amd_linux.go
+++ b/discover/amd_linux.go
@@ -1,541 +0,0 @@
-package discover
-
-import (
-	"bufio"
-	"errors"
-	"fmt"
-	"io"
-	"io/fs"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"regexp"
-	"slices"
-	"sort"
-	"strconv"
-	"strings"
-
-	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/format"
-)
-
-// Discovery logic for AMD/ROCm GPUs
-
-const (
-	DriverVersionFile     = "/sys/module/amdgpu/version"
-	AMDNodesSysfsDir      = "/sys/class/kfd/kfd/topology/nodes/"
-	GPUPropertiesFileGlob = AMDNodesSysfsDir + "*/properties"
-
-	// Prefix with the node dir
-	GPUTotalMemoryFileGlob = "mem_banks/*/properties" // size_in_bytes line
-
-	// Direct Rendering Manager sysfs location
-	DRMDeviceDirGlob   = "/sys/class/drm/card*/device"
-	DRMTotalMemoryFile = "mem_info_vram_total"
-	DRMUsedMemoryFile  = "mem_info_vram_used"
-
-	// In hex; properties file is in decimal
-	DRMUniqueIDFile = "unique_id"
-	DRMVendorFile   = "vendor"
-	DRMDeviceFile   = "device"
-)
-
-var (
-	// Used to validate if the given ROCm lib is usable
-	ROCmLibGlobs          = []string{"libhipblas.so.2*", "rocblas"} // TODO - probably include more coverage of files here...
-	RocmStandardLocations = []string{"/opt/rocm/lib", "/usr/lib64"}
-)
-
-// Gather GPU information from the amdgpu driver if any supported GPUs are detected
-// Only called once during bootstrap
-func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
-	resp := []RocmGPUInfo{}
-	if !AMDDetected() {
-		return resp, fmt.Errorf("AMD GPUs not detected")
-	}
-
-	// Opportunistic logging of driver version to aid in troubleshooting
-	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)
-	}
-
-	// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
-	var visibleDevices []string
-	hipVD := envconfig.HipVisibleDevices()   // zero based index only
-	rocrVD := envconfig.RocrVisibleDevices() // zero based index or UUID
-	gpuDO := envconfig.GpuDeviceOrdinal()    // zero based index
-	switch {
-	case rocrVD != "":
-		visibleDevices = strings.Split(rocrVD, ",")
-	case hipVD != "":
-		visibleDevices = strings.Split(hipVD, ",")
-	case gpuDO != "":
-		visibleDevices = strings.Split(gpuDO, ",")
-	}
-
-	gfxOverride := envconfig.HsaOverrideGfxVersion()
-	var supported []string
-	var libDir string
-
-	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
-	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
-	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
-	sort.Slice(matches, func(i, j int) bool {
-		// /sys/class/kfd/kfd/topology/nodes/<number>/properties
-		a, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[i])), 10, 64)
-		if err != nil {
-			slog.Debug("parse err", "error", err, "match", matches[i])
-			return false
-		}
-		b, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[j])), 10, 64)
-		if err != nil {
-			slog.Debug("parse err", "error", err, "match", matches[i])
-			return false
-		}
-		return a < b
-	})
-	gpuCount := 0
-	gpuOrdinalID := 0
-	for _, match := range matches {
-		slog.Debug("evaluating amdgpu node " + match)
-		fp, err := os.Open(match)
-		if err != nil {
-			slog.Debug("failed to open sysfs node", "file", match, "error", err)
-			continue
-		}
-		defer fp.Close()
-
-		scanner := bufio.NewScanner(fp)
-		isCPU := false
-		var major, minor, patch uint64
-		var vendor, device, uniqueID uint64
-		for scanner.Scan() {
-			line := strings.TrimSpace(scanner.Text())
-			// Note: we could also use "cpu_cores_count X" where X is greater than zero to detect CPUs
-			if strings.HasPrefix(line, "gfx_target_version") {
-				ver := strings.Fields(line)
-
-				// Detect CPUs
-				if len(ver) == 2 && ver[1] == "0" {
-					slog.Debug("detected CPU " + match)
-					isCPU = true
-					break
-				}
-
-				if len(ver) != 2 || len(ver[1]) < 5 {
-					slog.Warn("malformed "+match, "gfx_target_version", line)
-					// If this winds up being a CPU, our offsets may be wrong
-					continue
-				}
-				l := len(ver[1])
-				var err1, err2, err3 error
-				patch, err1 = strconv.ParseUint(ver[1][l-2:l], 10, 32)
-				minor, err2 = strconv.ParseUint(ver[1][l-4:l-2], 10, 32)
-				major, err3 = strconv.ParseUint(ver[1][:l-4], 10, 32)
-				if err1 != nil || err2 != nil || err3 != nil {
-					slog.Debug("malformed int " + line)
-					continue
-				}
-			} else if strings.HasPrefix(line, "vendor_id") {
-				ver := strings.Fields(line)
-				if len(ver) != 2 {
-					slog.Debug("malformed", "vendor_id", line)
-					continue
-				}
-				vendor, err = strconv.ParseUint(ver[1], 10, 64)
-				if err != nil {
-					slog.Debug("malformed", "vendor_id", line, "error", err)
-				}
-			} else if strings.HasPrefix(line, "device_id") {
-				ver := strings.Fields(line)
-				if len(ver) != 2 {
-					slog.Debug("malformed", "device_id", line)
-					continue
-				}
-				device, err = strconv.ParseUint(ver[1], 10, 64)
-				if err != nil {
-					slog.Debug("malformed", "device_id", line, "error", err)
-				}
-			} else if strings.HasPrefix(line, "unique_id") {
-				ver := strings.Fields(line)
-				if len(ver) != 2 {
-					slog.Debug("malformed", "unique_id", line)
-					continue
-				}
-				uniqueID, err = strconv.ParseUint(ver[1], 10, 64)
-				if err != nil {
-					slog.Debug("malformed", "unique_id", line, "error", err)
-				}
-			}
-			// TODO - any other properties we want to extract and record?
-			// vendor_id + device_id -> pci lookup for "Name"
-			// Other metrics that may help us understand relative performance between multiple GPUs
-		}
-
-		// Note: while ./mem_banks/*/used_memory exists, it doesn't appear to take other VRAM consumers
-		// into consideration, so we instead map the device over to the DRM driver sysfs nodes which
-		// do reliably report VRAM usage.
-
-		if isCPU {
-			continue
-		}
-
-		// Skip over any GPUs that are masked
-		if major == 0 && minor == 0 && patch == 0 {
-			slog.Debug("skipping gpu with gfx000")
-			continue
-		}
-
-		// Look up the memory for the current node
-		totalMemory := uint64(0)
-		usedMemory := uint64(0)
-		var usedFile string
-		mapping := []struct {
-			id       uint64
-			filename string
-		}{
-			{vendor, DRMVendorFile},
-			{device, DRMDeviceFile},
-			{uniqueID, DRMUniqueIDFile}, // Not all devices will report this
-		}
-		slog.Debug("mapping amdgpu to drm sysfs nodes", "amdgpu", match, "vendor", vendor, "device", device, "unique_id", uniqueID)
-		// Map over to DRM location to find the total/free memory
-		drmMatches, _ := filepath.Glob(DRMDeviceDirGlob)
-		for _, devDir := range drmMatches {
-			matched := true
-			for _, m := range mapping {
-				if m.id == 0 {
-					// Null ID means it didn't populate, so we can't use it to match
-					continue
-				}
-				filename := filepath.Join(devDir, m.filename)
-				buf, err := os.ReadFile(filename)
-				if err != nil {
-					slog.Debug("failed to read sysfs node", "file", filename, "error", err)
-					matched = false
-					break
-				}
-				// values here are in hex, strip off the lead 0x and parse so we can compare the numeric (decimal) values in amdgpu
-				cmp, err := strconv.ParseUint(strings.TrimPrefix(strings.TrimSpace(string(buf)), "0x"), 16, 64)
-				if err != nil {
-					slog.Debug("failed to parse sysfs node", "file", filename, "error", err)
-					matched = false
-					break
-				}
-				if cmp != m.id {
-					matched = false
-					break
-				}
-			}
-			if !matched {
-				continue
-			}
-
-			// Found the matching DRM directory
-			slog.Debug("matched", "amdgpu", match, "drm", devDir)
-			totalFile := filepath.Join(devDir, DRMTotalMemoryFile)
-			buf, err := os.ReadFile(totalFile)
-			if err != nil {
-				slog.Debug("failed to read sysfs node", "file", totalFile, "error", err)
-				break
-			}
-			totalMemory, err = strconv.ParseUint(strings.TrimSpace(string(buf)), 10, 64)
-			if err != nil {
-				slog.Debug("failed to parse sysfs node", "file", totalFile, "error", err)
-				break
-			}
-
-			usedFile = filepath.Join(devDir, DRMUsedMemoryFile)
-			usedMemory, err = getFreeMemory(usedFile)
-			if err != nil {
-				slog.Debug("failed to update used memory", "error", err)
-			}
-			break
-		}
-
-		var name string
-		// TODO - PCI ID lookup
-		if vendor > 0 && device > 0 {
-			name = fmt.Sprintf("%04x:%04x", vendor, device)
-		}
-
-		// Favor UUIDs if available to reduce possibility of getting the numeric IDs wrong
-		var ID string
-		if uniqueID != 0 {
-			ID = fmt.Sprintf("GPU-%016x", uniqueID)
-		} else {
-			ID = strconv.Itoa(gpuOrdinalID)
-		}
-
-		gpuInfo := RocmGPUInfo{
-			GpuInfo: GpuInfo{
-				Library: "rocm",
-				memInfo: memInfo{
-					TotalMemory: totalMemory,
-					FreeMemory:  (totalMemory - usedMemory),
-				},
-				ID:            ID,
-				Name:          name,
-				Compute:       fmt.Sprintf("gfx%d%x%x", major, minor, patch),
-				MinimumMemory: rocmMinimumMemory,
-				DriverMajor:   driverMajor,
-				DriverMinor:   driverMinor,
-			},
-			usedFilepath: usedFile,
-			index:        gpuCount,
-		}
-
-		// 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))
-			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-				GpuInfo: gpuInfo.GpuInfo,
-				Reason:  reason,
-			})
-			continue
-		}
-		minVer, err := strconv.Atoi(RocmComputeMajorMin)
-		if err != nil {
-			slog.Error("invalid RocmComputeMajorMin setting", "value", RocmComputeMajorMin, "error", err)
-		}
-		if int(major) < minVer {
-			reason := fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch)
-			slog.Warn(reason, "gpu", gpuInfo.ID)
-			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-				GpuInfo: gpuInfo.GpuInfo,
-				Reason:  reason,
-			})
-
-			continue
-		}
-
-		slog.Debug("amdgpu memory", "gpu", gpuInfo.ID, "total", format.HumanBytes2(totalMemory))
-		slog.Debug("amdgpu memory", "gpu", gpuInfo.ID, "available", format.HumanBytes2(totalMemory-usedMemory))
-
-		// 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
-		if libDir == "" {
-			libDir, err = AMDValidateLibDir()
-			if err != nil {
-				err = fmt.Errorf("unable to verify rocm library: %w", err)
-				slog.Warn(err.Error())
-				unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-					GpuInfo: gpuInfo.GpuInfo,
-					Reason:  err.Error(),
-				})
-				return nil, err
-			}
-		}
-		gpuInfo.DependencyPath = []string{libDir}
-
-		if gfxOverride == "" {
-			// Only load supported list once
-			if len(supported) == 0 {
-				supported, err = GetSupportedGFX(libDir)
-				if err != nil {
-					err = fmt.Errorf("failed to lookup supported GFX types: %w", err)
-					slog.Warn(err.Error())
-					unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-						GpuInfo: gpuInfo.GpuInfo,
-						Reason:  err.Error(),
-					})
-					return nil, err
-				}
-				slog.Debug("rocm supported GPUs", "types", supported)
-			}
-			gfx := gpuInfo.Compute
-			if !slices.Contains[[]string, string](supported, gfx) {
-				reason := fmt.Sprintf("amdgpu is not supported (supported types:%s)", supported)
-				slog.Warn(reason, "gpu_type", gfx, "gpu", gpuInfo.ID, "library", libDir)
-				unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-					GpuInfo: gpuInfo.GpuInfo,
-					Reason:  reason,
-				})
-
-				// TODO - consider discrete markdown just for ROCM troubleshooting?
-				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/gpu.md#overrides for HSA_OVERRIDE_GFX_VERSION usage")
-				continue
-			} else {
-				slog.Info("amdgpu is supported", "gpu", gpuInfo.ID, "gpu_type", gfx)
-			}
-		} else {
-			slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
-		}
-
-		// Check for env var workarounds
-		if name == "1002:687f" { // Vega RX 56
-			gpuInfo.EnvWorkarounds = append(gpuInfo.EnvWorkarounds, [2]string{"HSA_ENABLE_SDMA", "0"})
-		}
-
-		// The GPU has passed all the verification steps and is supported
-		resp = append(resp, gpuInfo)
-	}
-	if len(resp) == 0 {
-		err := fmt.Errorf("no compatible amdgpu devices detected")
-		slog.Info(err.Error())
-		return nil, err
-	}
-	if err := verifyKFDDriverAccess(); err != nil {
-		err = fmt.Errorf("amdgpu devices detected but permission problems block access: %w", err)
-		slog.Error(err.Error())
-		return nil, err
-	}
-	return resp, nil
-}
-
-// Quick check for AMD driver so we can skip amdgpu discovery if not present
-func AMDDetected() bool {
-	// Some driver versions (older?) don't have a version file, so just lookup the parent dir
-	sysfsDir := filepath.Dir(DriverVersionFile)
-	_, err := os.Stat(sysfsDir)
-	if errors.Is(err, os.ErrNotExist) {
-		slog.Debug("amdgpu driver not detected " + sysfsDir)
-		return false
-	} else if err != nil {
-		slog.Debug("error looking up amd driver", "path", sysfsDir, "error", err)
-		return false
-	}
-	return true
-}
-
-// Prefer to use host installed ROCm, as long as it meets our minimum requirements
-// failing that, tell the user how to download it on their own
-func AMDValidateLibDir() (string, error) {
-	libDir, err := commonAMDValidateLibDir()
-	if err == nil {
-		return libDir, nil
-	}
-
-	// Well known ollama installer path
-	installedRocmDir := "/usr/share/ollama/lib/rocm"
-	if rocmLibUsable(installedRocmDir) {
-		return installedRocmDir, nil
-	}
-
-	// If we still haven't found a usable rocm, the user will have to install it on their own
-	slog.Warn("amdgpu detected, but no compatible rocm library found.  Either install rocm v6, or follow manual install instructions at https://github.com/ollama/ollama/blob/main/docs/linux.md#manual-install")
-	return "", errors.New("no suitable rocm found, falling back to CPU")
-}
-
-func AMDDriverVersion() (driverMajor, driverMinor int, err error) {
-	_, err = os.Stat(DriverVersionFile)
-	if err != nil {
-		return 0, 0, fmt.Errorf("amdgpu version file missing: %s %w", DriverVersionFile, err)
-	}
-	fp, err := os.Open(DriverVersionFile)
-	if err != nil {
-		return 0, 0, err
-	}
-	defer fp.Close()
-	verString, err := io.ReadAll(fp)
-	if err != nil {
-		return 0, 0, err
-	}
-
-	pattern := `\A(\d+)\.(\d+).*`
-	regex := regexp.MustCompile(pattern)
-	match := regex.FindStringSubmatch(string(verString))
-	if len(match) < 2 {
-		return 0, 0, fmt.Errorf("malformed version string %s", string(verString))
-	}
-	driverMajor, err = strconv.Atoi(match[1])
-	if err != nil {
-		return 0, 0, err
-	}
-	driverMinor, err = strconv.Atoi(match[2])
-	if err != nil {
-		return 0, 0, err
-	}
-	return driverMajor, driverMinor, nil
-}
-
-func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
-	if len(gpus) == 0 {
-		return nil
-	}
-	for i := range gpus {
-		usedMemory, err := getFreeMemory(gpus[i].usedFilepath)
-		if err != nil {
-			return err
-		}
-		slog.Debug("updating rocm free memory", "gpu", gpus[i].ID, "name", gpus[i].Name, "before", format.HumanBytes2(gpus[i].FreeMemory), "now", format.HumanBytes2(gpus[i].TotalMemory-usedMemory))
-		gpus[i].FreeMemory = gpus[i].TotalMemory - usedMemory
-	}
-	return nil
-}
-
-func getFreeMemory(usedFile string) (uint64, error) {
-	buf, err := os.ReadFile(usedFile)
-	if err != nil {
-		return 0, fmt.Errorf("failed to read sysfs node %s %w", usedFile, err)
-	}
-	usedMemory, err := strconv.ParseUint(strings.TrimSpace(string(buf)), 10, 64)
-	if err != nil {
-		slog.Debug("failed to parse sysfs node", "file", usedFile, "error", err)
-		return 0, fmt.Errorf("failed to parse sysfs node %s %w", usedFile, err)
-	}
-	return usedMemory, nil
-}
-
-func verifyKFDDriverAccess() error {
-	// Verify we have permissions - either running as root, or we have group access to the driver
-	fd, err := os.OpenFile("/dev/kfd", os.O_RDWR, 0o666)
-	if err != nil {
-		if errors.Is(err, fs.ErrPermission) {
-			return fmt.Errorf("permissions not set up properly.  Either run ollama as root, or add you user account to the render group. %w", err)
-		} else if errors.Is(err, fs.ErrNotExist) {
-			// Container runtime failure?
-			return fmt.Errorf("kfd driver not loaded.  If running in a container, remember to include '--device /dev/kfd --device /dev/dri'")
-		}
-		return fmt.Errorf("failed to check permission on /dev/kfd: %w", err)
-	}
-	fd.Close()
-	return nil
-}
-
-func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
-	ids := []string{}
-	for _, info := range gpuInfo {
-		if info.Library != "rocm" {
-			// TODO shouldn't happen if things are wired correctly...
-			slog.Debug("rocmGetVisibleDevicesEnv skipping over non-rocm device", "library", info.Library)
-			continue
-		}
-		ids = append(ids, info.ID)
-	}
-	// There are 3 potential env vars to use to select GPUs.
-	// ROCR_VISIBLE_DEVICES supports UUID or numeric so is our preferred on linux
-	// GPU_DEVICE_ORDINAL supports numeric IDs only
-	// HIP_VISIBLE_DEVICES supports numeric IDs only
-	return "ROCR_VISIBLE_DEVICES", strings.Join(ids, ",")
-}
--- a/discover/amd_windows.go
+++ b/discover/amd_windows.go
@@ -1,218 +0,0 @@
-package discover
-
-import (
-	"bytes"
-	"errors"
-	"fmt"
-	"log/slog"
-	"path/filepath"
-	"slices"
-	"strconv"
-	"strings"
-
-	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/format"
-)
-
-const (
-
-	// TODO  We're lookinng for this exact name to detect iGPUs since hipGetDeviceProperties never reports integrated==true
-	iGPUName = "AMD Radeon(TM) Graphics"
-)
-
-var (
-	// Used to validate if the given ROCm lib is usable
-	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // This is not sufficient to discern v5 vs v6
-	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\6.1\\bin"} // TODO glob?
-)
-
-// Only called once during bootstrap
-func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
-	resp := []RocmGPUInfo{}
-	hl, err := NewHipLib()
-	if err != nil {
-		slog.Debug(err.Error())
-		return nil, err
-	}
-	defer hl.Release()
-
-	driverMajor, driverMinor, err := hl.AMDDriverVersion()
-	if err != nil {
-		// For now this is benign, but we may eventually need to fail compatibility checks
-		slog.Debug("error looking up amd driver version", "error", err)
-	}
-
-	// Note: the HIP library automatically handles subsetting to any *_VISIBLE_DEVICES the user specified
-	count := hl.HipGetDeviceCount()
-	if count == 0 {
-		err := fmt.Errorf("no compatible amdgpu devices detected")
-		slog.Info(err.Error())
-		return nil, err
-	}
-
-	libDir, err := AMDValidateLibDir()
-	if err != nil {
-		err = fmt.Errorf("unable to verify rocm library: %w", err)
-		slog.Warn(err.Error())
-		return nil, err
-	}
-
-	var supported []string
-	gfxOverride := envconfig.HsaOverrideGfxVersion()
-	if gfxOverride == "" {
-		supported, err = GetSupportedGFX(libDir)
-		if err != nil {
-			err = fmt.Errorf("failed to lookup supported GFX types: %w", err)
-			slog.Warn(err.Error())
-			return nil, err
-		}
-	} else {
-		slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
-	}
-
-	slog.Debug("detected hip devices", "count", count)
-	// TODO how to determine the underlying device ID when visible devices is causing this to subset?
-	for i := range count {
-		err = hl.HipSetDevice(i)
-		if err != nil {
-			slog.Warn("set device", "id", i, "error", err)
-			continue
-		}
-
-		props, err := hl.HipGetDeviceProperties(i)
-		if err != nil {
-			slog.Warn("get properties", "id", i, "error", err)
-			continue
-		}
-		n := bytes.IndexByte(props.Name[:], 0)
-		name := string(props.Name[:n])
-		// TODO is UUID actually populated on windows?
-		// Can luid be used on windows for setting visible devices (and is it actually set?)
-		n = bytes.IndexByte(props.GcnArchName[:], 0)
-		gfx := string(props.GcnArchName[:n])
-		slog.Debug("hip device", "id", i, "name", name, "gfx", gfx)
-		// slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
-		// TODO  Why isn't props.iGPU accurate!?
-
-		freeMemory, totalMemory, err := hl.HipMemGetInfo()
-		if err != nil {
-			slog.Warn("get mem info", "id", i, "error", err)
-			continue
-		}
-
-		gpuInfo := RocmGPUInfo{
-			GpuInfo: GpuInfo{
-				Library: "rocm",
-				memInfo: memInfo{
-					TotalMemory: totalMemory,
-					FreeMemory:  freeMemory,
-				},
-				// Free memory reporting on Windows is not reliable until we bump to ROCm v6.2
-				UnreliableFreeMemory: true,
-
-				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
-				DependencyPath: []string{libDir},
-				MinimumMemory:  rocmMinimumMemory,
-				Name:           name,
-				Compute:        gfx,
-				DriverMajor:    driverMajor,
-				DriverMinor:    driverMinor,
-			},
-			index: i,
-		}
-
-		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
-		if strings.EqualFold(name, iGPUName) || totalMemory < IGPUMemLimit {
-			reason := "unsupported Radeon iGPU detected skipping"
-			slog.Info(reason, "id", gpuInfo.ID, "total", format.HumanBytes2(totalMemory))
-			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-				GpuInfo: gpuInfo.GpuInfo,
-				Reason:  reason,
-			})
-			continue
-		}
-
-		// Strip off Target Features when comparing
-		if !slices.Contains[[]string, string](supported, strings.Split(gfx, ":")[0]) {
-			reason := fmt.Sprintf("amdgpu is not supported (supported types:%s)", supported)
-			slog.Warn(reason, "gpu_type", gfx, "gpu", gpuInfo.ID, "library", libDir)
-			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-				GpuInfo: gpuInfo.GpuInfo,
-				Reason:  reason,
-			})
-			// HSA_OVERRIDE_GFX_VERSION not supported on windows
-			continue
-		} else {
-			slog.Debug("amdgpu is supported", "gpu", i, "gpu_type", gfx)
-		}
-
-		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
-		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
-
-		resp = append(resp, gpuInfo)
-	}
-
-	return resp, nil
-}
-
-func AMDValidateLibDir() (string, error) {
-	libDir, err := commonAMDValidateLibDir()
-	if err == nil {
-		return libDir, nil
-	}
-
-	// Installer payload (if we're running from some other location)
-	rocmTargetDir := filepath.Join(LibOllamaPath, "rocm")
-	if rocmLibUsable(rocmTargetDir) {
-		slog.Debug("detected ollama installed ROCm at " + rocmTargetDir)
-		return rocmTargetDir, nil
-	}
-
-	// Should not happen on windows since we include it in the installer, but stand-alone binary might hit this
-	slog.Warn("amdgpu detected, but no compatible rocm library found.  Please install ROCm")
-	return "", errors.New("no suitable rocm found, falling back to CPU")
-}
-
-func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
-	if len(gpus) == 0 {
-		return nil
-	}
-	hl, err := NewHipLib()
-	if err != nil {
-		slog.Debug(err.Error())
-		return err
-	}
-	defer hl.Release()
-
-	for i := range gpus {
-		err := hl.HipSetDevice(gpus[i].index)
-		if err != nil {
-			return err
-		}
-		freeMemory, _, err := hl.HipMemGetInfo()
-		if err != nil {
-			slog.Warn("get mem info", "id", i, "error", err)
-			continue
-		}
-		slog.Debug("updating rocm free memory", "gpu", gpus[i].ID, "name", gpus[i].Name, "before", format.HumanBytes2(gpus[i].FreeMemory), "now", format.HumanBytes2(freeMemory))
-		gpus[i].FreeMemory = freeMemory
-	}
-	return nil
-}
-
-func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
-	ids := []string{}
-	for _, info := range gpuInfo {
-		if info.Library != "rocm" {
-			// TODO shouldn't happen if things are wired correctly...
-			slog.Debug("rocmGetVisibleDevicesEnv skipping over non-rocm device", "library", info.Library)
-			continue
-		}
-		ids = append(ids, info.ID)
-	}
-	// There are 3 potential env vars to use to select GPUs.
-	// ROCR_VISIBLE_DEVICES supports UUID or numeric but does not work on Windows
-	// HIP_VISIBLE_DEVICES supports numeric IDs only
-	// GPU_DEVICE_ORDINAL supports numeric IDs only
-	return "HIP_VISIBLE_DEVICES", strings.Join(ids, ",")
-}
--- a/discover/cpu_common.go
+++ b/discover/cpu_common.go
@@ -1,24 +0,0 @@
-package discover
-
-import (
-	"os"
-	"path/filepath"
-	"runtime"
-	"strings"
-)
-
-func IsNUMA() bool {
-	if runtime.GOOS != "linux" {
-		// numa support in llama.cpp is linux only
-		return false
-	}
-	ids := map[string]any{}
-	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
-	for _, packageId := range packageIds {
-		id, err := os.ReadFile(packageId)
-		if err == nil {
-			ids[strings.TrimSpace(string(id))] = struct{}{}
-		}
-	}
-	return len(ids) > 1
-}
--- a/discover/cpu_linux.go
+++ b/discover/cpu_linux.go
@@ -4,7 +4,9 @@ import (
 	"bufio"
 	"fmt"
 	"io"
+	"log/slog"
 	"os"
+	"path/filepath"
 	"reflect"
 	"regexp"
 	"sort"
@@ -13,47 +15,6 @@ import (
 	"github.com/ollama/ollama/format"
 )

-var CudartGlobs = []string{
-	"/usr/local/cuda/lib64/libcudart.so*",
-	"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
-	"/usr/lib/x86_64-linux-gnu/libcudart.so*",
-	"/usr/lib/wsl/lib/libcudart.so*",
-	"/usr/lib/wsl/drivers/*/libcudart.so*",
-	"/opt/cuda/lib64/libcudart.so*",
-	"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
-	"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
-	"/usr/lib/aarch64-linux-gnu/libcudart.so*",
-	"/usr/local/cuda/lib*/libcudart.so*",
-	"/usr/lib*/libcudart.so*",
-	"/usr/local/lib*/libcudart.so*",
-}
-
-var NvmlGlobs = []string{}
-
-var NvcudaGlobs = []string{
-	"/usr/local/cuda*/targets/*/lib/libcuda.so*",
-	"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
-	"/usr/lib/*-linux-gnu/libcuda.so*",
-	"/usr/lib/wsl/lib/libcuda.so*",
-	"/usr/lib/wsl/drivers/*/libcuda.so*",
-	"/opt/cuda/lib*/libcuda.so*",
-	"/usr/local/cuda/lib*/libcuda.so*",
-	"/usr/lib*/libcuda.so*",
-	"/usr/local/lib*/libcuda.so*",
-}
-
-var OneapiGlobs = []string{
-	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
-	"/usr/lib*/libze_intel_gpu.so*",
-}
-
-var (
-	CudartMgmtName = "libcudart.so*"
-	NvcudaMgmtName = "libcuda.so*"
-	NvmlMgmtName   = "" // not currently wired on linux
-	OneapiMgmtName = "libze_intel_gpu.so*"
-)
-
 func GetCPUMem() (memInfo, error) {
 	var mem memInfo
 	var total, available, free, buffers, cached, freeSwap uint64
@@ -106,16 +67,17 @@ type linuxCpuInfo struct {
 	CoreID     string `cpuinfo:"core id"`
 }

-func GetCPUDetails() ([]CPU, error) {
+func GetCPUDetails() []CPU {
 	file, err := os.Open(CpuInfoFilename)
 	if err != nil {
-		return nil, err
+		slog.Warn("failed to get CPU details", "error", err)
+		return nil
 	}
 	defer file.Close()
 	return linuxCPUDetails(file)
 }

-func linuxCPUDetails(file io.Reader) ([]CPU, error) {
+func linuxCPUDetails(file io.Reader) []CPU {
 	reColumns := regexp.MustCompile("\t+: ")
 	scanner := bufio.NewScanner(file)
 	cpuInfos := []linuxCpuInfo{}
@@ -194,5 +156,17 @@ func linuxCPUDetails(file io.Reader) ([]CPU, error) {
 	for _, k := range keys {
 		result = append(result, *socketByID[k])
 	}
-	return result, nil
+	return result
+}
+
+func IsNUMA() bool {
+	ids := map[string]any{}
+	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
+	for _, packageId := range packageIds {
+		id, err := os.ReadFile(packageId)
+		if err == nil {
+			ids[strings.TrimSpace(string(id))] = struct{}{}
+		}
+	}
+	return len(ids) > 1
 }
--- a/discover/cpu_linux_test.go
+++ b/discover/cpu_linux_test.go
@@ -2062,10 +2062,7 @@ power management:
 	for k, v := range testCases {
 		t.Run(k, func(t *testing.T) {
 			buf := bytes.NewBufferString(v.input)
-			cpus, err := linuxCPUDetails(buf)
-			if err != nil {
-				t.Fatal(err)
-			}
+			cpus := linuxCPUDetails(buf)

 			slog.Info("example", "scenario", k, "cpus", cpus)
 			si := SystemInfo{
--- a/discover/cpu_windows.go
+++ b/discover/cpu_windows.go
@@ -26,29 +26,6 @@ var (
 	GetLogicalProcessorInformationEx = k32.NewProc("GetLogicalProcessorInformationEx")
 )

-var CudartGlobs = []string{
-	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
-}
-
-var NvmlGlobs = []string{
-	"c:\\Windows\\System32\\nvml.dll",
-}
-
-var NvcudaGlobs = []string{
-	"c:\\windows\\system*\\nvcuda.dll",
-}
-
-var OneapiGlobs = []string{
-	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
-}
-
-var (
-	CudartMgmtName = "cudart64_*.dll"
-	NvcudaMgmtName = "nvcuda.dll"
-	NvmlMgmtName   = "nvml.dll"
-	OneapiMgmtName = "ze_intel_gpu64.dll"
-)
-
 func GetCPUMem() (memInfo, error) {
 	memStatus := MEMORYSTATUSEX{length: sizeofMemoryStatusEx}
 	r1, _, err := globalMemoryStatusExProc.Call(uintptr(unsafe.Pointer(&memStatus)))
@@ -122,27 +99,22 @@ func (pkg *winPackage) IsMember(target *GROUP_AFFINITY) bool {
 }

 func getLogicalProcessorInformationEx() ([]byte, error) {
-	buf := make([]byte, 1)
+	buf := make([]byte, 1024)
 	bufSize := len(buf)
-	ret, _, err := GetLogicalProcessorInformationEx.Call(
-		uintptr(RelationAll),
-		uintptr(unsafe.Pointer(&buf[0])),
-		uintptr(unsafe.Pointer(&bufSize)),
-	)
-	if ret != 0 {
-		return nil, fmt.Errorf("failed to determine size info ret:%d %w", ret, err)
+	var err error
+	for range 3 {
+		var ret uintptr
+		ret, _, err = GetLogicalProcessorInformationEx.Call(
+			uintptr(RelationAll),
+			uintptr(unsafe.Pointer(&buf[0])),
+			uintptr(unsafe.Pointer(&bufSize)),
+		)
+		if ret == 1 && bufSize <= len(buf) {
+			return buf, nil
+		}
+		buf = make([]byte, bufSize)
 	}
-
-	buf = make([]byte, bufSize)
-	ret, _, err = GetLogicalProcessorInformationEx.Call(
-		uintptr(RelationAll),
-		uintptr(unsafe.Pointer(&buf[0])),
-		uintptr(unsafe.Pointer(&bufSize)),
-	)
-	if ret == 0 {
-		return nil, fmt.Errorf("failed to gather processor information ret:%d buflen:%d %w", ret, bufSize, err)
-	}
-	return buf, nil
+	return nil, fmt.Errorf("unable to determine CPU details: %w", err)
 }

 func processSystemLogicalProcessorInforationList(buf []byte) []*winPackage {
@@ -217,10 +189,11 @@ func processSystemLogicalProcessorInforationList(buf []byte) []*winPackage {
 	return packages
 }

-func GetCPUDetails() ([]CPU, error) {
+func GetCPUDetails() []CPU {
 	buf, err := getLogicalProcessorInformationEx()
 	if err != nil {
-		return nil, err
+		slog.Warn("failed to get CPU details", "error", err)
+		return nil
 	}
 	packages := processSystemLogicalProcessorInforationList(buf)
 	cpus := make([]CPU, len(packages))
@@ -230,5 +203,10 @@ func GetCPUDetails() ([]CPU, error) {
 		cpus[i].EfficiencyCoreCount = pkg.efficiencyCoreCount
 		cpus[i].ThreadCount = pkg.threadCount
 	}
-	return cpus, nil
+	return cpus
+}
+
+func IsNUMA() bool {
+	// numa support in ggml is linux only
+	return false
 }
--- a/discover/cpu_windows_test.go
+++ b/discover/cpu_windows_test.go
--- a/discover/cuda_common.go
+++ b/discover/cuda_common.go
@@ -1,69 +0,0 @@
-//go:build linux || windows
-
-package discover
-
-import (
-	"fmt"
-	"log/slog"
-	"os"
-	"regexp"
-	"runtime"
-	"strconv"
-	"strings"
-)
-
-// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
-// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
-var CudaTegra string = os.Getenv("JETSON_JETPACK")
-
-func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
-	ids := []string{}
-	for _, info := range gpuInfo {
-		if info.Library != "cuda" {
-			// TODO shouldn't happen if things are wired correctly...
-			slog.Debug("cudaGetVisibleDevicesEnv skipping over non-cuda device", "library", info.Library)
-			continue
-		}
-		ids = append(ids, info.ID)
-	}
-	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
-}
-
-func cudaVariant(gpuInfo CudaGPUInfo) string {
-	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
-		if CudaTegra != "" {
-			ver := strings.Split(CudaTegra, ".")
-			if len(ver) > 0 {
-				return "jetpack" + ver[0]
-			}
-		} else if data, err := os.ReadFile("/etc/nv_tegra_release"); err == nil {
-			r := regexp.MustCompile(` R(\d+) `)
-			m := r.FindSubmatch(data)
-			if len(m) != 2 {
-				slog.Info("Unexpected format for /etc/nv_tegra_release.  Set JETSON_JETPACK to select version")
-			} else {
-				if l4t, err := strconv.Atoi(string(m[1])); err == nil {
-					// Note: mapping from L4t -> JP is inconsistent (can't just subtract 30)
-					// https://developer.nvidia.com/embedded/jetpack-archive
-					switch l4t {
-					case 35:
-						return "jetpack5"
-					case 36:
-						return "jetpack6"
-					default:
-						slog.Info("unsupported L4T version", "nv_tegra_release", string(data))
-					}
-				}
-			}
-		}
-		return "sbsa"
-	}
-
-	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
-	if gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
-		// The detected driver is older than Feb 2023
-		slog.Warn("old CUDA driver detected - please upgrade to a newer driver", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
-		return "v11"
-	}
-	return "v12"
-}
--- a/discover/gpu.go
+++ b/discover/gpu.go
@@ -1,720 +1,178 @@
-//go:build linux || windows
-
 package discover

-/*
-#cgo linux LDFLAGS: -lrt -lpthread -ldl -lstdc++ -lm
-#cgo windows LDFLAGS: -lpthread
-
-#include "gpu_info.h"
-*/
-import "C"
-
 import (
-	"fmt"
+	"context"
 	"log/slog"
 	"os"
 	"path/filepath"
+	"regexp"
 	"runtime"
 	"strconv"
 	"strings"
-	"sync"
-	"unsafe"

-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/ml"
 )

-type cudaHandles struct {
-	deviceCount int
-	cudart      *C.cudart_handle_t
-	nvcuda      *C.nvcuda_handle_t
-	nvml        *C.nvml_handle_t
+// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
+// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
+var CudaTegra string = os.Getenv("JETSON_JETPACK")
+
+func GetCPUInfo() GpuInfo {
+	mem, err := GetCPUMem()
+	if err != nil {
+		slog.Warn("error looking up system memory", "error", err)
+	}
+
+	return GpuInfo{
+		memInfo: mem,
+		DeviceID: ml.DeviceID{
+			Library: "cpu",
+			ID:      "0",
+		},
+	}
 }

-type oneapiHandles struct {
-	oneapi      *C.oneapi_handle_t
-	deviceCount int
+func GetGPUInfo(ctx context.Context, runners []FilteredRunnerDiscovery) GpuInfoList {
+	devs := GPUDevices(ctx, runners)
+	return devInfoToInfoList(devs)
 }

-const (
-	cudaMinimumMemory = 457 * format.MebiByte
-	rocmMinimumMemory = 457 * format.MebiByte
-	// TODO OneAPI minimum memory
-)
-
-var (
-	gpuMutex      sync.Mutex
-	bootstrapped  bool
-	cpus          []CPUInfo
-	cudaGPUs      []CudaGPUInfo
-	nvcudaLibPath string
-	cudartLibPath string
-	oneapiLibPath string
-	nvmlLibPath   string
-	rocmGPUs      []RocmGPUInfo
-	oneapiGPUs    []OneapiGPUInfo
-
-	// If any discovered GPUs are incompatible, report why
-	unsupportedGPUs []UnsupportedGPUInfo
-
-	// Keep track of errors during bootstrapping so that if GPUs are missing
-	// they expected to be present this may explain why
-	bootstrapErrors []error
-)
-
-// With our current CUDA compile flags, older than 5.0 will not work properly
-// (string values used to allow ldflags overrides at build time)
-var (
-	CudaComputeMajorMin = "5"
-	CudaComputeMinorMin = "0"
-)
-
-var RocmComputeMajorMin = "9"
-
-// TODO find a better way to detect iGPU instead of minimum memory
-const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
-
-// Note: gpuMutex must already be held
-func initCudaHandles() *cudaHandles {
-	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
-
-	cHandles := &cudaHandles{}
-	// Short Circuit if we already know which library to use
-	// ignore bootstrap errors in this case since we already recorded them
-	if nvmlLibPath != "" {
-		cHandles.nvml, _, _ = loadNVMLMgmt([]string{nvmlLibPath})
-		return cHandles
-	}
-	if nvcudaLibPath != "" {
-		cHandles.deviceCount, cHandles.nvcuda, _, _ = loadNVCUDAMgmt([]string{nvcudaLibPath})
-		return cHandles
-	}
-	if cudartLibPath != "" {
-		cHandles.deviceCount, cHandles.cudart, _, _ = loadCUDARTMgmt([]string{cudartLibPath})
-		return cHandles
-	}
-
-	slog.Debug("searching for GPU discovery libraries for NVIDIA")
-	var cudartMgmtPatterns []string
-
-	// Aligned with driver, we can't carry as payloads
-	nvcudaMgmtPatterns := NvcudaGlobs
-	cudartMgmtPatterns = append(cudartMgmtPatterns, filepath.Join(LibOllamaPath, "cuda_v*", CudartMgmtName))
-	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)
-
-	if len(NvmlGlobs) > 0 {
-		nvmlLibPaths := FindGPULibs(NvmlMgmtName, NvmlGlobs)
-		if len(nvmlLibPaths) > 0 {
-			nvml, libPath, err := loadNVMLMgmt(nvmlLibPaths)
-			if nvml != nil {
-				slog.Debug("nvidia-ml loaded", "library", libPath)
-				cHandles.nvml = nvml
-				nvmlLibPath = libPath
-			}
-			if err != nil {
-				bootstrapErrors = append(bootstrapErrors, err)
-			}
-		}
-	}
-
-	nvcudaLibPaths := FindGPULibs(NvcudaMgmtName, nvcudaMgmtPatterns)
-	if len(nvcudaLibPaths) > 0 {
-		deviceCount, nvcuda, libPath, err := loadNVCUDAMgmt(nvcudaLibPaths)
-		if nvcuda != nil {
-			slog.Debug("detected GPUs", "count", deviceCount, "library", libPath)
-			cHandles.nvcuda = nvcuda
-			cHandles.deviceCount = deviceCount
-			nvcudaLibPath = libPath
-			return cHandles
-		}
-		if err != nil {
-			bootstrapErrors = append(bootstrapErrors, err)
-		}
-	}
-
-	cudartLibPaths := FindGPULibs(CudartMgmtName, cudartMgmtPatterns)
-	if len(cudartLibPaths) > 0 {
-		deviceCount, cudart, libPath, err := loadCUDARTMgmt(cudartLibPaths)
-		if cudart != nil {
-			slog.Debug("detected GPUs", "library", libPath, "count", deviceCount)
-			cHandles.cudart = cudart
-			cHandles.deviceCount = deviceCount
-			cudartLibPath = libPath
-			return cHandles
-		}
-		if err != nil {
-			bootstrapErrors = append(bootstrapErrors, err)
-		}
-	}
-
-	return cHandles
-}
-
-// Note: gpuMutex must already be held
-func initOneAPIHandles() *oneapiHandles {
-	oHandles := &oneapiHandles{}
-
-	// Short Circuit if we already know which library to use
-	// ignore bootstrap errors in this case since we already recorded them
-	if oneapiLibPath != "" {
-		oHandles.deviceCount, oHandles.oneapi, _, _ = loadOneapiMgmt([]string{oneapiLibPath})
-		return oHandles
-	}
-
-	oneapiLibPaths := FindGPULibs(OneapiMgmtName, OneapiGlobs)
-	if len(oneapiLibPaths) > 0 {
-		var err error
-		oHandles.deviceCount, oHandles.oneapi, oneapiLibPath, err = loadOneapiMgmt(oneapiLibPaths)
-		if err != nil {
-			bootstrapErrors = append(bootstrapErrors, err)
-		}
-	}
-
-	return oHandles
-}
-
-func GetCPUInfo() GpuInfoList {
-	gpuMutex.Lock()
-	if !bootstrapped {
-		gpuMutex.Unlock()
-		GetGPUInfo()
-	} else {
-		gpuMutex.Unlock()
-	}
-	return GpuInfoList{cpus[0].GpuInfo}
-}
-
-func GetGPUInfo() GpuInfoList {
-	// TODO - consider exploring lspci (and equivalent on windows) to check for
-	// GPUs so we can report warnings if we see Nvidia/AMD but fail to load the libraries
-	gpuMutex.Lock()
-	defer gpuMutex.Unlock()
-	needRefresh := true
-	var cHandles *cudaHandles
-	var oHandles *oneapiHandles
-	defer func() {
-		if cHandles != nil {
-			if cHandles.cudart != nil {
-				C.cudart_release(*cHandles.cudart)
-			}
-			if cHandles.nvcuda != nil {
-				C.nvcuda_release(*cHandles.nvcuda)
-			}
-			if cHandles.nvml != nil {
-				C.nvml_release(*cHandles.nvml)
-			}
-		}
-		if oHandles != nil {
-			if oHandles.oneapi != nil {
-				// TODO - is this needed?
-				C.oneapi_release(*oHandles.oneapi)
-			}
-		}
-	}()
-
-	if !bootstrapped {
-		slog.Info("looking for compatible GPUs")
-		cudaComputeMajorMin, err := strconv.Atoi(CudaComputeMajorMin)
-		if err != nil {
-			slog.Error("invalid CudaComputeMajorMin setting", "value", CudaComputeMajorMin, "error", err)
-		}
-		cudaComputeMinorMin, err := strconv.Atoi(CudaComputeMinorMin)
-		if err != nil {
-			slog.Error("invalid CudaComputeMinorMin setting", "value", CudaComputeMinorMin, "error", err)
-		}
-		bootstrapErrors = []error{}
-		needRefresh = false
-		var memInfo C.mem_info_t
-
-		mem, err := GetCPUMem()
-		if err != nil {
-			slog.Warn("error looking up system memory", "error", err)
-		}
-
-		details, err := GetCPUDetails()
-		if err != nil {
-			slog.Warn("failed to lookup CPU details", "error", err)
-		}
-		cpus = []CPUInfo{
-			{
-				GpuInfo: GpuInfo{
-					memInfo: mem,
-					Library: "cpu",
-					ID:      "0",
-				},
-				CPUs: details,
-			},
-		}
-
-		// Load ALL libraries
-		cHandles = initCudaHandles()
-
-		// NVIDIA
-		for i := range cHandles.deviceCount {
-			if cHandles.cudart != nil || cHandles.nvcuda != nil {
-				gpuInfo := CudaGPUInfo{
-					GpuInfo: GpuInfo{
-						Library: "cuda",
-					},
-					index: i,
-				}
-				var driverMajor int
-				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)
-					driverMinor = int(cHandles.nvcuda.driver_minor)
-				}
-				if memInfo.err != nil {
-					slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
-					C.free(unsafe.Pointer(memInfo.err))
-					continue
-				}
-				gpuInfo.TotalMemory = uint64(memInfo.total)
-				gpuInfo.FreeMemory = uint64(memInfo.free)
-				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-				gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
-				gpuInfo.computeMajor = int(memInfo.major)
-				gpuInfo.computeMinor = int(memInfo.minor)
-				gpuInfo.MinimumMemory = cudaMinimumMemory
-				gpuInfo.DriverMajor = driverMajor
-				gpuInfo.DriverMinor = driverMinor
-				variant := cudaVariant(gpuInfo)
-
-				// Start with our bundled libraries
-				if variant != "" {
-					variantPath := filepath.Join(LibOllamaPath, "cuda_"+variant)
-					if _, err := os.Stat(variantPath); err == nil {
-						// Put the variant directory first in the search path to avoid runtime linking to the wrong library
-						gpuInfo.DependencyPath = append([]string{variantPath}, gpuInfo.DependencyPath...)
-					}
-				}
-				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-				gpuInfo.Variant = variant
-
-				if int(memInfo.major) < cudaComputeMajorMin || (int(memInfo.major) == cudaComputeMajorMin && int(memInfo.minor) < cudaComputeMinorMin) {
-					unsupportedGPUs = append(unsupportedGPUs,
-						UnsupportedGPUInfo{
-							GpuInfo: gpuInfo.GpuInfo,
-						})
-					slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
-					continue
-				}
-
-				// query the management library as well so we can record any skew between the two
-				// which represents overhead on the GPU we must set aside on subsequent updates
-				if cHandles.nvml != nil {
-					uuid := C.CString(gpuInfo.ID)
-					defer C.free(unsafe.Pointer(uuid))
-					C.nvml_get_free(*cHandles.nvml, uuid, &memInfo.free, &memInfo.total, &memInfo.used)
-					if memInfo.err != nil {
-						slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
-						C.free(unsafe.Pointer(memInfo.err))
-					} else {
-						if memInfo.free != 0 && uint64(memInfo.free) > gpuInfo.FreeMemory {
-							gpuInfo.OSOverhead = uint64(memInfo.free) - gpuInfo.FreeMemory
-							slog.Info("detected OS VRAM overhead",
-								"id", gpuInfo.ID,
-								"library", gpuInfo.Library,
-								"compute", gpuInfo.Compute,
-								"driver", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor),
-								"name", gpuInfo.Name,
-								"overhead", format.HumanBytes2(gpuInfo.OSOverhead),
-							)
-						}
-					}
-				}
-
-				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
-				cudaGPUs = append(cudaGPUs, gpuInfo)
-			}
-		}
-
-		// Intel
-		if envconfig.IntelGPU() {
-			oHandles = initOneAPIHandles()
-			if oHandles != nil && oHandles.oneapi != nil {
-				for d := range oHandles.oneapi.num_drivers {
-					if oHandles.oneapi == nil {
-						// shouldn't happen
-						slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
-						continue
-					}
-					devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
-					for i := range devCount {
-						gpuInfo := OneapiGPUInfo{
-							GpuInfo: GpuInfo{
-								Library: "oneapi",
-							},
-							driverIndex: int(d),
-							gpuIndex:    int(i),
-						}
-						// TODO - split bootstrapping from updating free memory
-						C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
-						// TODO - convert this to MinimumMemory based on testing...
-						var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
-						memInfo.free = C.uint64_t(totalFreeMem)
-						gpuInfo.TotalMemory = uint64(memInfo.total)
-						gpuInfo.FreeMemory = uint64(memInfo.free)
-						gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-						gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-						gpuInfo.DependencyPath = []string{LibOllamaPath}
-						oneapiGPUs = append(oneapiGPUs, gpuInfo)
-					}
-				}
-			}
-		}
-
-		rocmGPUs, err = AMDGetGPUInfo()
-		if err != nil {
-			bootstrapErrors = append(bootstrapErrors, err)
-		}
-		bootstrapped = true
-		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
-			slog.Info("no compatible GPUs were discovered")
-		}
-
-		// TODO verify we have runners for the discovered GPUs, filter out any that aren't supported with good error messages
-	}
-
-	// For detected GPUs, load library if not loaded
-
-	// Refresh free memory usage
-	if needRefresh {
-		mem, err := GetCPUMem()
-		if err != nil {
-			slog.Warn("error looking up system memory", "error", err)
-		} else {
-			slog.Debug("updating system memory data",
-				slog.Group(
-					"before",
-					"total", format.HumanBytes2(cpus[0].TotalMemory),
-					"free", format.HumanBytes2(cpus[0].FreeMemory),
-					"free_swap", format.HumanBytes2(cpus[0].FreeSwap),
-				),
-				slog.Group(
-					"now",
-					"total", format.HumanBytes2(mem.TotalMemory),
-					"free", format.HumanBytes2(mem.FreeMemory),
-					"free_swap", format.HumanBytes2(mem.FreeSwap),
-				),
-			)
-			cpus[0].FreeMemory = mem.FreeMemory
-			cpus[0].FreeSwap = mem.FreeSwap
-		}
-
-		var memInfo C.mem_info_t
-		if cHandles == nil && len(cudaGPUs) > 0 {
-			cHandles = initCudaHandles()
-		}
-		for i, gpu := range cudaGPUs {
-			if cHandles.nvml != nil {
-				uuid := C.CString(gpu.ID)
-				defer C.free(unsafe.Pointer(uuid))
-				C.nvml_get_free(*cHandles.nvml, uuid, &memInfo.free, &memInfo.total, &memInfo.used)
-			} else if cHandles.cudart != nil {
-				C.cudart_bootstrap(*cHandles.cudart, C.int(gpu.index), &memInfo)
-			} else if cHandles.nvcuda != nil {
-				C.nvcuda_get_free(*cHandles.nvcuda, C.int(gpu.index), &memInfo.free, &memInfo.total)
-				memInfo.used = memInfo.total - memInfo.free
-			} else {
-				// shouldn't happen
-				slog.Warn("no valid cuda library loaded to refresh vram usage")
-				break
-			}
-			if memInfo.err != nil {
-				slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
-				C.free(unsafe.Pointer(memInfo.err))
-				continue
-			}
-			if memInfo.free == 0 {
-				slog.Warn("error looking up nvidia GPU memory")
-				continue
-			}
-			if cHandles.nvml != nil && gpu.OSOverhead > 0 {
-				// When using the management library update based on recorded overhead
-				memInfo.free -= C.uint64_t(gpu.OSOverhead)
-			}
-			slog.Debug("updating cuda memory data",
-				"gpu", gpu.ID,
-				"name", gpu.Name,
-				"overhead", format.HumanBytes2(gpu.OSOverhead),
-				slog.Group(
-					"before",
-					"total", format.HumanBytes2(gpu.TotalMemory),
-					"free", format.HumanBytes2(gpu.FreeMemory),
-				),
-				slog.Group(
-					"now",
-					"total", format.HumanBytes2(uint64(memInfo.total)),
-					"free", format.HumanBytes2(uint64(memInfo.free)),
-					"used", format.HumanBytes2(uint64(memInfo.used)),
-				),
-			)
-			cudaGPUs[i].FreeMemory = uint64(memInfo.free)
-		}
-
-		if oHandles == nil && len(oneapiGPUs) > 0 {
-			oHandles = initOneAPIHandles()
-		}
-		for i, gpu := range oneapiGPUs {
-			if oHandles.oneapi == nil {
-				// shouldn't happen
-				slog.Warn("nil oneapi handle with device count", "count", oHandles.deviceCount)
-				continue
-			}
-			C.oneapi_check_vram(*oHandles.oneapi, C.int(gpu.driverIndex), C.int(gpu.gpuIndex), &memInfo)
-			// TODO - convert this to MinimumMemory based on testing...
-			var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
-			memInfo.free = C.uint64_t(totalFreeMem)
-			oneapiGPUs[i].FreeMemory = uint64(memInfo.free)
-		}
-
-		err = RocmGPUInfoList(rocmGPUs).RefreshFreeMemory()
-		if err != nil {
-			slog.Debug("problem refreshing ROCm free memory", "error", err)
-		}
-	}
-
+func devInfoToInfoList(devs []ml.DeviceInfo) GpuInfoList {
 	resp := []GpuInfo{}
-	for _, gpu := range cudaGPUs {
-		resp = append(resp, gpu.GpuInfo)
+	// Our current packaging model places ggml-hip in the main directory
+	// but keeps rocm in an isolated directory.  We have to add it to
+	// the [LD_LIBRARY_]PATH so ggml-hip will load properly
+	rocmDir := filepath.Join(LibOllamaPath, "rocm")
+	if _, err := os.Stat(rocmDir); err != nil {
+		rocmDir = ""
 	}
-	for _, gpu := range rocmGPUs {
-		resp = append(resp, gpu.GpuInfo)
-	}
-	for _, gpu := range oneapiGPUs {
-		resp = append(resp, gpu.GpuInfo)
+
+	for _, dev := range devs {
+		info := GpuInfo{
+			DeviceID: dev.DeviceID,
+			filterID: dev.FilteredID,
+			Name:     dev.Description,
+			memInfo: memInfo{
+				TotalMemory: dev.TotalMemory,
+				FreeMemory:  dev.FreeMemory,
+			},
+			// TODO can we avoid variant
+			DependencyPath: dev.LibraryPath,
+			DriverMajor:    dev.DriverMajor,
+			DriverMinor:    dev.DriverMinor,
+			ComputeMajor:   dev.ComputeMajor,
+			ComputeMinor:   dev.ComputeMinor,
+		}
+		if dev.Library == "CUDA" || dev.Library == "ROCm" {
+			info.MinimumMemory = 457 * format.MebiByte
+		}
+		if dev.Library == "ROCm" && rocmDir != "" {
+			info.DependencyPath = append(info.DependencyPath, rocmDir)
+		}
+		resp = append(resp, info)
 	}
 	if len(resp) == 0 {
-		resp = append(resp, cpus[0].GpuInfo)
+		mem, err := GetCPUMem()
+		if err != nil {
+			slog.Warn("error looking up system memory", "error", err)
+		}
+
+		resp = append(resp, GpuInfo{
+			memInfo: mem,
+			DeviceID: ml.DeviceID{
+				Library: "cpu",
+				ID:      "0",
+			},
+		})
 	}
 	return resp
 }

-func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
-	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
-	gpuLibPaths := []string{}
-	slog.Debug("Searching for GPU library", "name", baseLibName)
-
-	// search our bundled libraries first
-	patterns := []string{filepath.Join(LibOllamaPath, baseLibName)}
-
-	var ldPaths []string
-	switch runtime.GOOS {
-	case "windows":
-		ldPaths = strings.Split(os.Getenv("PATH"), string(os.PathListSeparator))
-	case "linux":
-		ldPaths = strings.Split(os.Getenv("LD_LIBRARY_PATH"), string(os.PathListSeparator))
-	}
-
-	// then search the system's LD_LIBRARY_PATH
-	for _, p := range ldPaths {
-		p, err := filepath.Abs(p)
-		if err != nil {
-			continue
-		}
-		patterns = append(patterns, filepath.Join(p, baseLibName))
-	}
-
-	// finally, search the default patterns provided by the caller
-	patterns = append(patterns, defaultPatterns...)
-	slog.Debug("gpu library search", "globs", patterns)
-	for _, pattern := range patterns {
-		// Nvidia PhysX known to return bogus results
-		if strings.Contains(pattern, "PhysX") {
-			slog.Debug("skipping PhysX cuda library path", "path", pattern)
-			continue
-		}
-		// Ignore glob discovery errors
-		matches, _ := filepath.Glob(pattern)
-		for _, match := range matches {
-			// Resolve any links so we don't try the same lib multiple times
-			// and weed out any dups across globs
-			libPath := match
-			tmp := match
-			var err error
-			for ; err == nil; tmp, err = os.Readlink(libPath) {
-				if !filepath.IsAbs(tmp) {
-					tmp = filepath.Join(filepath.Dir(libPath), tmp)
-				}
-				libPath = tmp
-			}
-			new := true
-			for _, cmp := range gpuLibPaths {
-				if cmp == libPath {
-					new = false
-					break
-				}
-			}
-			if new {
-				gpuLibPaths = append(gpuLibPaths, libPath)
-			}
-		}
-	}
-	slog.Debug("discovered GPU libraries", "paths", gpuLibPaths)
-	return gpuLibPaths
-}
-
-// Bootstrap the runtime library
-// Returns: num devices, handle, libPath, error
-func loadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string, error) {
-	var resp C.cudart_init_resp_t
-	resp.ch.verbose = getVerboseState()
-	var err error
-	for _, libPath := range cudartLibPaths {
-		lib := C.CString(libPath)
-		defer C.free(unsafe.Pointer(lib))
-		C.cudart_init(lib, &resp)
-		if resp.err != nil {
-			err = fmt.Errorf("Unable to load cudart library %s: %s", libPath, C.GoString(resp.err))
-			slog.Debug(err.Error())
-			C.free(unsafe.Pointer(resp.err))
-		} else {
-			err = nil
-			return int(resp.num_devices), &resp.ch, libPath, err
-		}
-	}
-	return 0, nil, "", err
-}
-
-// Bootstrap the driver library
-// Returns: num devices, handle, libPath, error
-func loadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string, error) {
-	var resp C.nvcuda_init_resp_t
-	resp.ch.verbose = getVerboseState()
-	var err error
-	for _, libPath := range nvcudaLibPaths {
-		lib := C.CString(libPath)
-		defer C.free(unsafe.Pointer(lib))
-		C.nvcuda_init(lib, &resp)
-		if resp.err != nil {
-			// Decide what log level based on the type of error message to help users understand why
-			switch resp.cudaErr {
-			case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
-				err = fmt.Errorf("version mismatch between driver and cuda driver library - reboot or upgrade may be required: library %s", libPath)
-				slog.Warn(err.Error())
-			case C.CUDA_ERROR_NO_DEVICE:
-				err = fmt.Errorf("no nvidia devices detected by library %s", libPath)
-				slog.Info(err.Error())
-			case C.CUDA_ERROR_UNKNOWN:
-				err = fmt.Errorf("unknown error initializing cuda driver library %s: %s. see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information", libPath, C.GoString(resp.err))
-				slog.Warn(err.Error())
-			default:
-				msg := C.GoString(resp.err)
-				if strings.Contains(msg, "wrong ELF class") {
-					slog.Debug("skipping 32bit library", "library", libPath)
-				} else {
-					err = fmt.Errorf("Unable to load cudart library %s: %s", libPath, C.GoString(resp.err))
-					slog.Info(err.Error())
-				}
-			}
-			C.free(unsafe.Pointer(resp.err))
-		} else {
-			err = nil
-			return int(resp.num_devices), &resp.ch, libPath, err
-		}
-	}
-	return 0, nil, "", err
-}
-
-// Bootstrap the management library
-// Returns: handle, libPath, error
-func loadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string, error) {
-	var resp C.nvml_init_resp_t
-	resp.ch.verbose = getVerboseState()
-	var err error
-	for _, libPath := range nvmlLibPaths {
-		lib := C.CString(libPath)
-		defer C.free(unsafe.Pointer(lib))
-		C.nvml_init(lib, &resp)
-		if resp.err != nil {
-			err = fmt.Errorf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err))
-			slog.Info(err.Error())
-			C.free(unsafe.Pointer(resp.err))
-		} else {
-			err = nil
-			return &resp.ch, libPath, err
-		}
-	}
-	return nil, "", err
-}
-
-// bootstrap the Intel GPU library
-// Returns: num devices, handle, libPath, error
-func loadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string, error) {
-	var resp C.oneapi_init_resp_t
-	num_devices := 0
-	resp.oh.verbose = getVerboseState()
-	var err error
-	for _, libPath := range oneapiLibPaths {
-		lib := C.CString(libPath)
-		defer C.free(unsafe.Pointer(lib))
-		C.oneapi_init(lib, &resp)
-		if resp.err != nil {
-			err = fmt.Errorf("Unable to load oneAPI management library %s: %s", libPath, C.GoString(resp.err))
-			slog.Debug(err.Error())
-			C.free(unsafe.Pointer(resp.err))
-		} else {
-			err = nil
-			for i := range resp.oh.num_drivers {
-				num_devices += int(C.oneapi_get_device_count(resp.oh, C.int(i)))
-			}
-			return num_devices, &resp.oh, libPath, err
-		}
-	}
-	return 0, nil, "", err
-}
-
-func getVerboseState() C.uint16_t {
-	if envconfig.LogLevel() < slog.LevelInfo {
-		return C.uint16_t(1)
-	}
-	return C.uint16_t(0)
-}
-
 // Given the list of GPUs this instantiation is targeted for,
 // figure out the visible devices environment variable
 //
 // If different libraries are detected, the first one is what we use
-func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
+func (l GpuInfoList) GetVisibleDevicesEnv() []string {
 	if len(l) == 0 {
-		return "", ""
-	}
-	switch l[0].Library {
-	case "cuda":
-		return cudaGetVisibleDevicesEnv(l)
-	case "rocm":
-		return rocmGetVisibleDevicesEnv(l)
-	case "oneapi":
-		return oneapiGetVisibleDevicesEnv(l)
-	default:
-		slog.Debug("no filter required for library " + l[0].Library)
-		return "", ""
+		return nil
 	}
+	return []string{rocmGetVisibleDevicesEnv(l)}
 }

-func GetSystemInfo() SystemInfo {
-	gpus := GetGPUInfo()
-	gpuMutex.Lock()
-	defer gpuMutex.Unlock()
-	discoveryErrors := []string{}
-	for _, err := range bootstrapErrors {
-		discoveryErrors = append(discoveryErrors, err.Error())
+func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) string {
+	ids := []string{}
+	for _, info := range gpuInfo {
+		if info.Library != "ROCm" {
+			continue
+		}
+		// If the devices requires a numeric ID, for filtering purposes, we use the unfiltered ID number
+		if info.filterID != "" {
+			ids = append(ids, info.filterID)
+		} else {
+			ids = append(ids, info.ID)
+		}
 	}
+	if len(ids) == 0 {
+		return ""
+	}
+	envVar := "ROCR_VISIBLE_DEVICES="
+	if runtime.GOOS != "linux" {
+		envVar = "HIP_VISIBLE_DEVICES="
+	}
+	// There are 3 potential env vars to use to select GPUs.
+	// ROCR_VISIBLE_DEVICES supports UUID or numeric but does not work on Windows
+	// HIP_VISIBLE_DEVICES supports numeric IDs only
+	// GPU_DEVICE_ORDINAL supports numeric IDs only
+	return envVar + strings.Join(ids, ",")
+}
+
+// GetSystemInfo returns the last cached state of the GPUs on the system
+func GetSystemInfo() SystemInfo {
+	deviceMu.Lock()
+	defer deviceMu.Unlock()
+	gpus := devInfoToInfoList(devices)
 	if len(gpus) == 1 && gpus[0].Library == "cpu" {
 		gpus = []GpuInfo{}
 	}

 	return SystemInfo{
-		System:          cpus[0],
-		GPUs:            gpus,
-		UnsupportedGPUs: unsupportedGPUs,
-		DiscoveryErrors: discoveryErrors,
+		System: CPUInfo{
+			CPUs:    GetCPUDetails(),
+			GpuInfo: GetCPUInfo(),
+		},
+		GPUs: gpus,
 	}
 }
+
+func cudaJetpack() string {
+	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
+		if CudaTegra != "" {
+			ver := strings.Split(CudaTegra, ".")
+			if len(ver) > 0 {
+				return "jetpack" + ver[0]
+			}
+		} else if data, err := os.ReadFile("/etc/nv_tegra_release"); err == nil {
+			r := regexp.MustCompile(` R(\d+) `)
+			m := r.FindSubmatch(data)
+			if len(m) != 2 {
+				slog.Info("Unexpected format for /etc/nv_tegra_release.  Set JETSON_JETPACK to select version")
+			} else {
+				if l4t, err := strconv.Atoi(string(m[1])); err == nil {
+					// Note: mapping from L4t -> JP is inconsistent (can't just subtract 30)
+					// https://developer.nvidia.com/embedded/jetpack-archive
+					switch l4t {
+					case 35:
+						return "jetpack5"
+					case 36:
+						return "jetpack6"
+					default:
+						// Newer Jetson systems use the SBSU runtime
+						slog.Debug("unrecognized L4T version", "nv_tegra_release", string(data))
+					}
+				}
+			}
+		}
+	}
+	return ""
+}
--- a/discover/gpu_darwin.go
+++ b/discover/gpu_darwin.go
@@ -1,5 +1,3 @@
-//go:build darwin
-
 package discover

 /*
@@ -11,7 +9,6 @@ import "C"

 import (
 	"log/slog"
-	"runtime"
 	"syscall"

 	"github.com/ollama/ollama/format"
@@ -21,39 +18,6 @@ const (
 	metalMinimumMemory = 512 * format.MebiByte
 )

-func GetGPUInfo() GpuInfoList {
-	mem, _ := GetCPUMem()
-	if runtime.GOARCH == "amd64" {
-		return []GpuInfo{
-			{
-				Library: "cpu",
-				memInfo: mem,
-			},
-		}
-	}
-	info := GpuInfo{
-		Library: "metal",
-		ID:      "0",
-	}
-	info.TotalMemory = uint64(C.getRecommendedMaxVRAM())
-
-	// TODO is there a way to gather actual allocated video memory? (currentAllocatedSize doesn't work)
-	info.FreeMemory = info.TotalMemory
-
-	info.MinimumMemory = metalMinimumMemory
-	return []GpuInfo{info}
-}
-
-func GetCPUInfo() GpuInfoList {
-	mem, _ := GetCPUMem()
-	return []GpuInfo{
-		{
-			Library: "cpu",
-			memInfo: mem,
-		},
-	}
-}
-
 func GetCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
@@ -62,13 +26,7 @@ func GetCPUMem() (memInfo, error) {
 	}, nil
 }

-func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
-	// No-op on darwin
-	return "", ""
-}
-
-func GetSystemInfo() SystemInfo {
-	mem, _ := GetCPUMem()
+func GetCPUDetails() []CPU {
 	query := "hw.perflevel0.physicalcpu"
 	perfCores, err := syscall.SysctlUint32(query)
 	if err != nil {
@@ -81,19 +39,16 @@ func GetSystemInfo() SystemInfo {
 	query = "hw.logicalcpu"
 	logicalCores, _ := syscall.SysctlUint32(query)

-	return SystemInfo{
-		System: CPUInfo{
-			GpuInfo: GpuInfo{
-				memInfo: mem,
-			},
-			CPUs: []CPU{
-				{
-					CoreCount:           int(perfCores + efficiencyCores),
-					EfficiencyCoreCount: int(efficiencyCores),
-					ThreadCount:         int(logicalCores),
-				},
-			},
+	return []CPU{
+		{
+			CoreCount:           int(perfCores + efficiencyCores),
+			EfficiencyCoreCount: int(efficiencyCores),
+			ThreadCount:         int(logicalCores),
 		},
-		GPUs: GetGPUInfo(),
 	}
 }
+
+func IsNUMA() bool {
+	// numa support in ggml is linux only
+	return false
+}
--- a/discover/gpu_info.h
+++ b/discover/gpu_info.h
@@ -1,72 +0,0 @@
-#ifndef __APPLE__
-#ifndef __GPU_INFO_H__
-#define __GPU_INFO_H__
-#include <stdint.h>
-#include <stdio.h>
-#include <stdlib.h>
-
-#ifndef _WIN32
-#include <dlfcn.h>
-#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags)
-#define LOAD_SYMBOL(handle, sym) dlsym(handle, sym)
-#define LOAD_ERR() strdup(dlerror())
-#define UNLOAD_LIBRARY(handle) dlclose(handle)
-#else
-#include <windows.h>
-#define LOAD_LIBRARY(lib, flags) LoadLibrary(lib)
-#define LOAD_SYMBOL(handle, sym) GetProcAddress(handle, sym)
-#define UNLOAD_LIBRARY(handle) FreeLibrary(handle)
-#define LOAD_ERR() ({\
-  LPSTR messageBuffer = NULL; \
-  size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, \
-                                 NULL, GetLastError(), MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&messageBuffer, 0, NULL); \
-  char *resp = strdup(messageBuffer); \
-  LocalFree(messageBuffer); \
-  resp; \
-})
-
-#endif
-
-#ifndef LOG
-#define LOG(verbose, ...) \
-  do { \
-    if (verbose) { \
-      fprintf(stderr, __VA_ARGS__); \
-    } \
-  } while (0)
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define GPU_ID_LEN 64
-#define GPU_NAME_LEN 96
-
-typedef struct mem_info {
-  char *err;  // If non-nill, caller responsible for freeing
-  char gpu_id[GPU_ID_LEN];
-  char gpu_name[GPU_NAME_LEN];
-  uint64_t total;
-  uint64_t free;
-  uint64_t used;
-
-  // Compute Capability
-  int major; 
-  int minor;
-  int patch;
-} mem_info_t;
-
-void cpu_check_ram(mem_info_t *resp);
-
-#ifdef __cplusplus
-}
-#endif
-
-#include "gpu_info_cudart.h"
-#include "gpu_info_nvcuda.h"
-#include "gpu_info_nvml.h"
-#include "gpu_info_oneapi.h"
-
-#endif  // __GPU_INFO_H__
-#endif  // __APPLE__
--- a/discover/gpu_info_cudart.c
+++ b/discover/gpu_info_cudart.c
@@ -1,181 +0,0 @@
-#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
-
-#include <string.h>
-#include <inttypes.h>
-#include "gpu_info_cudart.h"
-
-void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
-  cudartReturn_t ret;
-  resp->err = NULL;
-  resp->num_devices = 0;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i;
-
-  struct lookup {
-    char *s;
-    void **p;
-  } l[] = {
-      {"cudaSetDevice", (void *)&resp->ch.cudaSetDevice},
-      {"cudaDeviceSynchronize", (void *)&resp->ch.cudaDeviceSynchronize},
-      {"cudaDeviceReset", (void *)&resp->ch.cudaDeviceReset},
-      {"cudaMemGetInfo", (void *)&resp->ch.cudaMemGetInfo},
-      {"cudaGetDeviceCount", (void *)&resp->ch.cudaGetDeviceCount},
-      {"cudaDeviceGetAttribute", (void *)&resp->ch.cudaDeviceGetAttribute},
-      {"cudaDriverGetVersion", (void *)&resp->ch.cudaDriverGetVersion},
-      {"cudaGetDeviceProperties", (void *)&resp->ch.cudaGetDeviceProperties},
-      {NULL, NULL},
-  };
-
-  resp->ch.handle = LOAD_LIBRARY(cudart_lib_path, RTLD_LAZY);
-  if (!resp->ch.handle) {
-    char *msg = LOAD_ERR();
-    LOG(resp->ch.verbose, "library %s load err: %s\n", cudart_lib_path, msg);
-    snprintf(buf, buflen,
-            "Unable to load %s library to query for Nvidia GPUs: %s",
-            cudart_lib_path, msg);
-    free(msg);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  for (i = 0; l[i].s != NULL; i++) {
-    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!*(l[i].p)) {
-      char *msg = LOAD_ERR();
-      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
-      UNLOAD_LIBRARY(resp->ch.handle);
-      resp->ch.handle = NULL;
-      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
-              msg);
-      free(msg);
-      resp->err = strdup(buf);
-      return;
-    }
-  }
-
-  ret = (*resp->ch.cudaSetDevice)(0);
-  if (ret != CUDART_SUCCESS) {
-    LOG(resp->ch.verbose, "cudaSetDevice err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->ch.handle);
-    resp->ch.handle = NULL;
-    if (ret == CUDART_ERROR_INSUFFICIENT_DRIVER) {
-      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
-      return;
-    }
-    snprintf(buf, buflen, "cudart init failure: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  int version = 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);
-  }
-
-  ret = (*resp->ch.cudaGetDeviceCount)(&resp->num_devices);
-  if (ret != CUDART_SUCCESS) {
-    LOG(resp->ch.verbose, "cudaGetDeviceCount err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->ch.handle);
-    resp->ch.handle = NULL;
-    snprintf(buf, buflen, "unable to get device count: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-}
-
-
-void cudart_bootstrap(cudart_handle_t h, int i, mem_info_t *resp) {
-  resp->err = NULL;
-  cudartMemory_t memInfo = {0,0,0};
-  cudartReturn_t ret;
-  const int buflen = 256;
-  char buf[buflen + 1];
-
-  if (h.handle == NULL) {
-    resp->err = strdup("cudart handle isn't initialized");
-    return;
-  }
-
-  ret = (*h.cudaSetDevice)(i);
-  if (ret != CUDART_SUCCESS) {
-    snprintf(buf, buflen, "cudart device failed to initialize");
-    resp->err = strdup(buf);
-    return;
-  }
-
-  cudaDeviceProp_t props;
-  ret = (*h.cudaGetDeviceProperties)(&props, i);
-  if (ret != CUDART_SUCCESS) {
-    LOG(h.verbose, "[%d] device properties lookup failure: %d\n", i, ret);
-    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
-    resp->major = 0;
-    resp->minor = 0;
-  } else {
-    int allNull = 1;
-    for (int j = 0; j < 16; j++) {
-      if (props.uuid.bytes[j] != 0) {
-        allNull = 0;
-        break;
-      }
-    }
-    if (allNull != 0) {
-      snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
-    } else {
-      // GPU-d110a105-ac29-1d54-7b49-9c90440f215b
-      snprintf(&resp->gpu_id[0], GPU_ID_LEN,
-          "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
-          props.uuid.bytes[0],
-          props.uuid.bytes[1],
-          props.uuid.bytes[2],
-          props.uuid.bytes[3],
-          props.uuid.bytes[4],
-          props.uuid.bytes[5],
-          props.uuid.bytes[6],
-          props.uuid.bytes[7],
-          props.uuid.bytes[8],
-          props.uuid.bytes[9],
-          props.uuid.bytes[10],
-          props.uuid.bytes[11],
-          props.uuid.bytes[12],
-          props.uuid.bytes[13],
-          props.uuid.bytes[14],
-          props.uuid.bytes[15]
-        );
-    }
-    resp->major = props.major;
-    resp->minor = props.minor;
-
-    // TODO add other useful properties from props
-  }
-  ret = (*h.cudaMemGetInfo)(&memInfo.free, &memInfo.total);
-  if (ret != CUDART_SUCCESS) {
-    snprintf(buf, buflen, "cudart device memory info lookup failure %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  resp->total = memInfo.total;
-  resp->free = memInfo.free;
-  resp->used = memInfo.used;
-
-  LOG(h.verbose, "[%s] CUDA totalMem %" PRId64 "\n", resp->gpu_id, resp->total);
-  LOG(h.verbose, "[%s] CUDA freeMem %" PRId64 "\n", resp->gpu_id, resp->free);
-  LOG(h.verbose, "[%s] CUDA usedMem %" PRId64 "\n", resp->gpu_id, resp->used);
-  LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
-}
-
-void cudart_release(cudart_handle_t h) {
-  LOG(h.verbose, "releasing cudart library\n");
-  UNLOAD_LIBRARY(h.handle);
-  h.handle = NULL;
-}
-
-#endif  // __APPLE__
--- a/discover/gpu_info_cudart.h
+++ b/discover/gpu_info_cudart.h
@@ -1,145 +0,0 @@
-#ifndef __APPLE__
-#ifndef __GPU_INFO_CUDART_H__
-#define __GPU_INFO_CUDART_H__
-#include "gpu_info.h"
-
-// Just enough typedef's to dlopen/dlsym for memory information
-typedef enum cudartReturn_enum {
-  CUDART_SUCCESS = 0,
-  CUDART_ERROR_INVALID_VALUE = 1,
-  CUDART_ERROR_MEMORY_ALLOCATION = 2,
-  CUDART_ERROR_INSUFFICIENT_DRIVER = 35,
-  // Other values omitted for now...
-} cudartReturn_t;
-
-typedef enum cudartDeviceAttr_enum {
-  cudartDevAttrComputeCapabilityMajor = 75,
-  cudartDevAttrComputeCapabilityMinor = 76,
-
-  // TODO - not yet wired up but may be useful for Jetson or other
-  // integrated GPU scenarios with shared memory
-  cudaDevAttrIntegrated = 18
-
-} cudartDeviceAttr_t;
-
-typedef void *cudartDevice_t;  // Opaque is sufficient
-typedef struct cudartMemory_st {
-  size_t total;
-  size_t free;
-  size_t used;
-} cudartMemory_t;
-
-typedef struct cudaUUID {
-    unsigned char bytes[16];
-} cudaUUID_t;
-typedef struct cudaDeviceProp {
-    char         name[256];                  /**< ASCII string identifying device */
-    cudaUUID_t   uuid;                       /**< 16-byte unique identifier */
-    char         luid[8];                    /**< 8-byte locally unique identifier. Value is undefined on TCC and non-Windows platforms */
-    unsigned int luidDeviceNodeMask;         /**< LUID device node mask. Value is undefined on TCC and non-Windows platforms */
-    size_t       totalGlobalMem;             /**< Global memory available on device in bytes */
-    size_t       sharedMemPerBlock;          /**< Shared memory available per block in bytes */
-    int          regsPerBlock;               /**< 32-bit registers available per block */
-    int          warpSize;                   /**< Warp size in threads */
-    size_t       memPitch;                   /**< Maximum pitch in bytes allowed by memory copies */
-    int          maxThreadsPerBlock;         /**< Maximum number of threads per block */
-    int          maxThreadsDim[3];           /**< Maximum size of each dimension of a block */
-    int          maxGridSize[3];             /**< Maximum size of each dimension of a grid */
-    int          clockRate;                  /**< Clock frequency in kilohertz */
-    size_t       totalConstMem;              /**< Constant memory available on device in bytes */
-    int          major;                      /**< Major compute capability */
-    int          minor;                      /**< Minor compute capability */
-    size_t       textureAlignment;           /**< Alignment requirement for textures */
-    size_t       texturePitchAlignment;      /**< Pitch alignment requirement for texture references bound to pitched memory */
-    int          deviceOverlap;              /**< Device can concurrently copy memory and execute a kernel. Deprecated. Use instead asyncEngineCount. */
-    int          multiProcessorCount;        /**< Number of multiprocessors on device */
-    int          kernelExecTimeoutEnabled;   /**< Specified whether there is a run time limit on kernels */
-    int          integrated;                 /**< Device is integrated as opposed to discrete */
-    int          canMapHostMemory;           /**< Device can map host memory with cudaHostAlloc/cudaHostGetDevicePointer */
-    int          computeMode;                /**< Compute mode (See ::cudaComputeMode) */
-    int          maxTexture1D;               /**< Maximum 1D texture size */
-    int          maxTexture1DMipmap;         /**< Maximum 1D mipmapped texture size */
-    int          maxTexture1DLinear;         /**< Deprecated, do not use. Use cudaDeviceGetTexture1DLinearMaxWidth() or cuDeviceGetTexture1DLinearMaxWidth() instead. */
-    int          maxTexture2D[2];            /**< Maximum 2D texture dimensions */
-    int          maxTexture2DMipmap[2];      /**< Maximum 2D mipmapped texture dimensions */
-    int          maxTexture2DLinear[3];      /**< Maximum dimensions (width, height, pitch) for 2D textures bound to pitched memory */
-    int          maxTexture2DGather[2];      /**< Maximum 2D texture dimensions if texture gather operations have to be performed */
-    int          maxTexture3D[3];            /**< Maximum 3D texture dimensions */
-    int          maxTexture3DAlt[3];         /**< Maximum alternate 3D texture dimensions */
-    int          maxTextureCubemap;          /**< Maximum Cubemap texture dimensions */
-    int          maxTexture1DLayered[2];     /**< Maximum 1D layered texture dimensions */
-    int          maxTexture2DLayered[3];     /**< Maximum 2D layered texture dimensions */
-    int          maxTextureCubemapLayered[2];/**< Maximum Cubemap layered texture dimensions */
-    int          maxSurface1D;               /**< Maximum 1D surface size */
-    int          maxSurface2D[2];            /**< Maximum 2D surface dimensions */
-    int          maxSurface3D[3];            /**< Maximum 3D surface dimensions */
-    int          maxSurface1DLayered[2];     /**< Maximum 1D layered surface dimensions */
-    int          maxSurface2DLayered[3];     /**< Maximum 2D layered surface dimensions */
-    int          maxSurfaceCubemap;          /**< Maximum Cubemap surface dimensions */
-    int          maxSurfaceCubemapLayered[2];/**< Maximum Cubemap layered surface dimensions */
-    size_t       surfaceAlignment;           /**< Alignment requirements for surfaces */
-    int          concurrentKernels;          /**< Device can possibly execute multiple kernels concurrently */
-    int          ECCEnabled;                 /**< Device has ECC support enabled */
-    int          pciBusID;                   /**< PCI bus ID of the device */
-    int          pciDeviceID;                /**< PCI device ID of the device */
-    int          pciDomainID;                /**< PCI domain ID of the device */
-    int          tccDriver;                  /**< 1 if device is a Tesla device using TCC driver, 0 otherwise */
-    int          asyncEngineCount;           /**< Number of asynchronous engines */
-    int          unifiedAddressing;          /**< Device shares a unified address space with the host */
-    int          memoryClockRate;            /**< Peak memory clock frequency in kilohertz */
-    int          memoryBusWidth;             /**< Global memory bus width in bits */
-    int          l2CacheSize;                /**< Size of L2 cache in bytes */
-    int          persistingL2CacheMaxSize;   /**< Device's maximum l2 persisting lines capacity setting in bytes */
-    int          maxThreadsPerMultiProcessor;/**< Maximum resident threads per multiprocessor */
-    int          streamPrioritiesSupported;  /**< Device supports stream priorities */
-    int          globalL1CacheSupported;     /**< Device supports caching globals in L1 */
-    int          localL1CacheSupported;      /**< Device supports caching locals in L1 */
-    size_t       sharedMemPerMultiprocessor; /**< Shared memory available per multiprocessor in bytes */
-    int          regsPerMultiprocessor;      /**< 32-bit registers available per multiprocessor */
-    int          managedMemory;              /**< Device supports allocating managed memory on this system */
-    int          isMultiGpuBoard;            /**< Device is on a multi-GPU board */
-    int          multiGpuBoardGroupID;       /**< Unique identifier for a group of devices on the same multi-GPU board */
-    int          hostNativeAtomicSupported;  /**< Link between the device and the host supports native atomic operations */
-    int          singleToDoublePrecisionPerfRatio; /**< Ratio of single precision performance (in floating-point operations per second) to double precision performance */
-    int          pageableMemoryAccess;       /**< Device supports coherently accessing pageable memory without calling cudaHostRegister on it */
-    int          concurrentManagedAccess;    /**< Device can coherently access managed memory concurrently with the CPU */
-    int          computePreemptionSupported; /**< Device supports Compute Preemption */
-    int          canUseHostPointerForRegisteredMem; /**< Device can access host registered memory at the same virtual address as the CPU */
-    int          cooperativeLaunch;          /**< Device supports launching cooperative kernels via ::cudaLaunchCooperativeKernel */
-    int          cooperativeMultiDeviceLaunch; /**< Deprecated, cudaLaunchCooperativeKernelMultiDevice is deprecated. */
-    size_t       sharedMemPerBlockOptin;     /**< Per device maximum shared memory per block usable by special opt in */
-    int          pageableMemoryAccessUsesHostPageTables; /**< Device accesses pageable memory via the host's page tables */
-    int          directManagedMemAccessFromHost; /**< Host can directly access managed memory on the device without migration. */
-    int          maxBlocksPerMultiProcessor; /**< Maximum number of resident blocks per multiprocessor */
-    int          accessPolicyMaxWindowSize;  /**< The maximum value of ::cudaAccessPolicyWindow::num_bytes. */
-    size_t       reservedSharedMemPerBlock;  /**< Shared memory reserved by CUDA driver per block in bytes */
-  } cudaDeviceProp_t;
-
-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);
-  cudartReturn_t (*cudaMemGetInfo)(size_t *, size_t *);
-  cudartReturn_t (*cudaGetDeviceCount)(int *);
-  cudartReturn_t (*cudaDeviceGetAttribute)(int* value, cudartDeviceAttr_t attr, int device);
-  cudartReturn_t (*cudaDriverGetVersion) (int *driverVersion);
-  cudartReturn_t (*cudaGetDeviceProperties) (cudaDeviceProp_t* prop, int device);
-} cudart_handle_t;
-
-typedef struct cudart_init_resp {
-  char *err;  // If err is non-null handle is invalid
-  cudart_handle_t ch;
-  int num_devices;
-} cudart_init_resp_t;
-
-void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp);
-void cudart_bootstrap(cudart_handle_t ch, int device_id, mem_info_t *resp);
-// TODO - if we keep this library longer term, add cudart_get_free
-void cudart_release(cudart_handle_t ch);
-
-#endif  // __GPU_INFO_CUDART_H__
-#endif  // __APPLE__
--- a/discover/gpu_info_nvcuda.c
+++ b/discover/gpu_info_nvcuda.c
@@ -1,251 +0,0 @@
-#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
-
-#include <string.h>
-#include <inttypes.h>
-#include "gpu_info_nvcuda.h"
-
-void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
-  LOG(resp->ch.verbose, "initializing %s\n", nvcuda_lib_path);
-  CUresult ret;
-  resp->err = NULL;
-  resp->num_devices = 0;
-  resp->cudaErr = CUDA_SUCCESS;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i;
-
-  struct lookup {
-    char *s;
-    void **p;
-  } l[] = {
-   
-      {"cuInit", (void *)&resp->ch.cuInit},
-      {"cuDriverGetVersion", (void *)&resp->ch.cuDriverGetVersion},
-      {"cuDeviceGetCount", (void *)&resp->ch.cuDeviceGetCount},
-      {"cuDeviceGet", (void *)&resp->ch.cuDeviceGet},
-      {"cuDeviceGetAttribute", (void *)&resp->ch.cuDeviceGetAttribute},
-      {"cuDeviceGetUuid", (void *)&resp->ch.cuDeviceGetUuid},
-      {"cuDeviceGetName", (void *)&resp->ch.cuDeviceGetName},
-      {"cuCtxCreate_v3", (void *)&resp->ch.cuCtxCreate_v3},
-      {"cuMemGetInfo_v2", (void *)&resp->ch.cuMemGetInfo_v2},
-      {"cuCtxDestroy", (void *)&resp->ch.cuCtxDestroy},
-      {NULL, NULL},
-  };
-
-  resp->ch.handle = LOAD_LIBRARY(nvcuda_lib_path, RTLD_LAZY);
-  if (!resp->ch.handle) {
-    char *msg = LOAD_ERR();
-    LOG(resp->ch.verbose, "library %s load err: %s\n", nvcuda_lib_path, msg);
-    snprintf(buf, buflen,
-            "Unable to load %s library to query for Nvidia GPUs: %s",
-            nvcuda_lib_path, msg);
-    free(msg);
-    resp->err = strdup(buf);
-    resp->cudaErr = -1;
-    return;
-  }
-
-  for (i = 0; l[i].s != NULL; i++) {
-    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!*(l[i].p)) {
-      char *msg = LOAD_ERR();
-      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
-      UNLOAD_LIBRARY(resp->ch.handle);
-      resp->ch.handle = NULL;
-      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
-              msg);
-      free(msg);
-      resp->err = strdup(buf);
-      resp->cudaErr = -1;
-      return;
-    }
-    LOG(resp->ch.verbose, "dlsym: %s - %p\n", l[i].s, *l[i].p);
-  }
-
-  LOG(resp->ch.verbose, "calling cuInit\n");
-  ret = (*resp->ch.cuInit)(0);
-  if (ret != CUDA_SUCCESS) {
-    LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->ch.handle);
-    resp->ch.handle = NULL;
-    snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
-    resp->err = strdup(buf);
-    resp->cudaErr = ret;
-    return;
-  }
-
-  int version = 0;
-  resp->ch.driver_major = 0;
-  resp->ch.driver_minor = 0;
-
-  // Report driver version if we're in verbose mode, ignore errors
-  LOG(resp->ch.verbose, "calling cuDriverGetVersion\n");
-  ret = (*resp->ch.cuDriverGetVersion)(&version);
-  if (ret != CUDA_SUCCESS) {
-    LOG(resp->ch.verbose, "cuDriverGetVersion failed: %d\n", ret);
-  } else {
-    LOG(resp->ch.verbose, "raw version 0x%x\n", version);
-    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);
-  }
-
-  LOG(resp->ch.verbose, "calling cuDeviceGetCount\n");
-  ret = (*resp->ch.cuDeviceGetCount)(&resp->num_devices);
-  if (ret != CUDA_SUCCESS) {
-    LOG(resp->ch.verbose, "cuDeviceGetCount err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->ch.handle);
-    resp->ch.handle = NULL;
-    snprintf(buf, buflen, "unable to get device count: %d", ret);
-    resp->err = strdup(buf);
-    resp->cudaErr = ret;
-    return;
-  }
-  LOG(resp->ch.verbose, "device count %d\n", resp->num_devices);
-}
-
-const int buflen = 256;
-void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
-  resp->err = NULL;
-  nvcudaMemory_t memInfo = {0,0};
-  CUresult ret;
-  CUdevice device = -1;
-  CUcontext ctx = NULL;
-  char buf[buflen + 1];
-  CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
-
-  if (h.handle == NULL) {
-    resp->err = strdup("cuda driver library handle isn't initialized");
-    return;
-  }
-
-  ret = (*h.cuDeviceGet)(&device, i);
-  if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "cuda driver library device failed to initialize");
-    resp->err = strdup(buf);
-    return;
-  }
-
-  int major = 0;
-  int minor = 0;
-  ret = (*h.cuDeviceGetAttribute)(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device);
-  if (ret != CUDA_SUCCESS) {
-    LOG(h.verbose, "[%d] device major lookup failure: %d\n", i, ret);
-  } else {
-    ret = (*h.cuDeviceGetAttribute)(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device);
-    if (ret != CUDA_SUCCESS) {
-      LOG(h.verbose, "[%d] device minor lookup failure: %d\n", i, ret);
-    } else {
-      resp->minor = minor;  
-      resp->major = major;  
-    }
-  }
-
-  ret = (*h.cuDeviceGetUuid)(&uuid, device);
-  if (ret != CUDA_SUCCESS) {
-    LOG(h.verbose, "[%d] device uuid lookup failure: %d\n", i, ret);
-    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
-  } else {
-    // GPU-d110a105-ac29-1d54-7b49-9c90440f215b
-    snprintf(&resp->gpu_id[0], GPU_ID_LEN,
-        "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
-        uuid.bytes[0],
-        uuid.bytes[1],
-        uuid.bytes[2],
-        uuid.bytes[3],
-        uuid.bytes[4],
-        uuid.bytes[5],
-        uuid.bytes[6],
-        uuid.bytes[7],
-        uuid.bytes[8],
-        uuid.bytes[9],
-        uuid.bytes[10],
-        uuid.bytes[11],
-        uuid.bytes[12],
-        uuid.bytes[13],
-        uuid.bytes[14],
-        uuid.bytes[15]
-      );
-  }
-
-  ret = (*h.cuDeviceGetName)(&resp->gpu_name[0], GPU_NAME_LEN, device);
-  if (ret != CUDA_SUCCESS) {
-    LOG(h.verbose, "[%d] device name lookup failure: %d\n", i, ret);
-    resp->gpu_name[0] = '\0';
-  }
-
-  // To get memory we have to set (and release) a context
-  ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
-  if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "cuda driver library failed to get device context %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
-  if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "cuda driver library device memory info lookup failure %d", ret);
-    resp->err = strdup(buf);
-    // Best effort on failure...
-    (*h.cuCtxDestroy)(ctx);
-    return;
-  }
-
-  resp->total = memInfo.total;
-  resp->free = memInfo.free;
-
-  LOG(h.verbose, "[%s] CUDA totalMem %" PRId64 "mb\n", resp->gpu_id, resp->total / 1024 / 1024);
-  LOG(h.verbose, "[%s] CUDA freeMem %" PRId64 "mb\n", resp->gpu_id, resp->free / 1024 / 1024);
-  LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
-
-  
-
-  ret = (*h.cuCtxDestroy)(ctx);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library failed to release device context %d", ret);
-  }
-}
-
-void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total) {
-  CUresult ret;
-  CUcontext ctx = NULL;
-  CUdevice device = -1;
-  *free = 0;
-  *total = 0;
-
-  ret = (*h.cuDeviceGet)(&device, i);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library device failed to initialize");
-    return;
-  }
-
-
-  // To get memory we have to set (and release) a context
-  ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library failed to get device context %d", ret);
-    return;
-  }
-
-  ret = (*h.cuMemGetInfo_v2)(free, total);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library device memory info lookup failure %d", ret);
-    // Best effort on failure...
-    (*h.cuCtxDestroy)(ctx);
-    return;
-  }
-
-  ret = (*h.cuCtxDestroy)(ctx);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library failed to release device context %d", ret);
-  }
-}
-
-void nvcuda_release(nvcuda_handle_t h) {
-  LOG(h.verbose, "releasing cuda driver library\n");
-  UNLOAD_LIBRARY(h.handle);
-  // TODO and other context release logic?
-  h.handle = NULL;
-}
-
-#endif  // __APPLE__
--- a/discover/gpu_info_nvcuda.h
+++ b/discover/gpu_info_nvcuda.h
@@ -1,79 +0,0 @@
-#ifndef __APPLE__
-#ifndef __GPU_INFO_NVCUDA_H__
-#define __GPU_INFO_NVCUDA_H__
-#include "gpu_info.h"
-
-// Just enough typedef's to dlopen/dlsym for memory information
-typedef enum cudaError_enum {
-  CUDA_SUCCESS = 0,
-  CUDA_ERROR_INVALID_VALUE = 1,
-  CUDA_ERROR_OUT_OF_MEMORY = 2,
-  CUDA_ERROR_NOT_INITIALIZED = 3,
-  CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
-  CUDA_ERROR_NO_DEVICE = 100,
-  CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
-  CUDA_ERROR_UNKNOWN = 999,
-  // Other values omitted for now...
-} CUresult;
-
-typedef enum CUdevice_attribute_enum {
-  CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75,
-  CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR = 76,
-
-  // TODO - not yet wired up but may be useful for Jetson or other
-  // integrated GPU scenarios with shared memory
-  CU_DEVICE_ATTRIBUTE_INTEGRATED = 18
-
-} CUdevice_attribute;
-
-typedef void *nvcudaDevice_t;  // Opaque is sufficient
-typedef struct nvcudaMemory_st {
-  uint64_t total;
-  uint64_t free;
-} nvcudaMemory_t;
-
-typedef struct nvcudaDriverVersion {
-  int major;
-  int minor;
-} nvcudaDriverVersion_t;
-
-typedef struct CUuuid_st {
-    unsigned char bytes[16];
-} CUuuid;
-
-typedef int CUdevice;
-typedef void* CUcontext;
-
-typedef struct nvcuda_handle {
-  void *handle;
-  uint16_t verbose;
-  int driver_major;
-  int driver_minor;
-  CUresult (*cuInit)(unsigned int Flags);
-  CUresult (*cuDriverGetVersion)(int *driverVersion);
-  CUresult (*cuDeviceGetCount)(int *);
-  CUresult (*cuDeviceGet)(CUdevice* device, int ordinal);
-  CUresult (*cuDeviceGetAttribute)(int* pi, CUdevice_attribute attrib, CUdevice dev);
-  CUresult (*cuDeviceGetUuid)(CUuuid* uuid, CUdevice dev); // signature compatible with cuDeviceGetUuid_v2
-  CUresult (*cuDeviceGetName)(char *name, int len, CUdevice dev);
-
-  // Context specific aspects
-  CUresult (*cuCtxCreate_v3)(CUcontext* pctx, void *params, int len, unsigned int flags, CUdevice dev);
-  CUresult (*cuMemGetInfo_v2)(uint64_t* free, uint64_t* total);
-  CUresult (*cuCtxDestroy)(CUcontext ctx);
-} nvcuda_handle_t;
-
-typedef struct nvcuda_init_resp {
-  char *err;  // If err is non-null handle is invalid
-  nvcuda_handle_t ch;
-  int num_devices;
-  CUresult cudaErr;
-} nvcuda_init_resp_t;
-
-void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);
-void nvcuda_bootstrap(nvcuda_handle_t ch, int device_id, mem_info_t *resp);
-void nvcuda_get_free(nvcuda_handle_t ch,  int device_id, uint64_t *free, uint64_t *total);
-void nvcuda_release(nvcuda_handle_t ch);
-
-#endif  // __GPU_INFO_NVCUDA_H__
-#endif  // __APPLE__
--- a/discover/gpu_info_nvml.c
+++ b/discover/gpu_info_nvml.c
@@ -1,104 +0,0 @@
-#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
-
-#include <string.h>
-
-#include "gpu_info_nvml.h"
-
-void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
-  nvmlReturn_t ret;
-  resp->err = NULL;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i;
-
-  struct lookup {
-    char *s;
-    void **p;
-  } l[] = {
-      {"nvmlInit_v2", (void *)&resp->ch.nvmlInit_v2},
-      {"nvmlShutdown", (void *)&resp->ch.nvmlShutdown},
-      {"nvmlDeviceGetHandleByUUID", (void *)&resp->ch.nvmlDeviceGetHandleByUUID},
-      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.nvmlDeviceGetMemoryInfo},
-      {NULL, NULL},
-  };
-
-  resp->ch.handle = LOAD_LIBRARY(nvml_lib_path, RTLD_LAZY);
-  if (!resp->ch.handle) {
-    char *msg = LOAD_ERR();
-    LOG(resp->ch.verbose, "library %s load err: %s\n", nvml_lib_path, msg);
-    snprintf(buf, buflen,
-             "Unable to load %s library to query for Nvidia GPUs: %s",
-             nvml_lib_path, msg);
-    free(msg);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  // TODO once we've squashed the remaining corner cases remove this log
-  // LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", nvml_lib_path);
-  
-  for (i = 0; l[i].s != NULL; i++) {
-    // TODO once we've squashed the remaining corner cases remove this log
-    // LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
-
-    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!*(l[i].p)) {
-      resp->ch.handle = NULL;
-      char *msg = LOAD_ERR();
-      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
-      UNLOAD_LIBRARY(resp->ch.handle);
-      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
-               msg);
-      free(msg);
-      resp->err = strdup(buf);
-      return;
-    }
-  }
-
-  ret = (*resp->ch.nvmlInit_v2)();
-  if (ret != NVML_SUCCESS) {
-    LOG(resp->ch.verbose, "nvmlInit_v2 err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->ch.handle);
-    resp->ch.handle = NULL;
-    snprintf(buf, buflen, "nvml vram init failure: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-}
-
-
-void nvml_get_free(nvml_handle_t h, char *uuid, uint64_t *free, uint64_t *total, uint64_t *used) {
-    nvmlDevice_t device;
-    nvmlMemory_t memInfo = {0};
-    nvmlReturn_t ret;
-    ret = (*h.nvmlDeviceGetHandleByUUID)((const char *)(uuid), &device);
-    if (ret != NVML_SUCCESS) {
-        LOG(1, "unable to get device handle %s: %d", uuid, ret);
-        *free = 0;
-        return;
-    }
-
-    ret = (*h.nvmlDeviceGetMemoryInfo)(device, &memInfo);
-    if (ret != NVML_SUCCESS) {
-        LOG(1, "device memory info lookup failure %s: %d", uuid, ret);
-        *free = 0;
-        return;
-    }
-    *free = memInfo.free;
-    *total = memInfo.total;
-    *used = memInfo.used;
-}
-
-
-void nvml_release(nvml_handle_t h) {
-  LOG(h.verbose, "releasing nvml library\n");
-  nvmlReturn_t ret;
-  ret = (*h.nvmlShutdown)();
-  if (ret != NVML_SUCCESS) {
-    LOG(1, "error during nvmlShutdown %d", ret);
-  }
-  UNLOAD_LIBRARY(h.handle);
-  h.handle = NULL;
-}
-
-#endif  // __APPLE__
--- a/discover/gpu_info_nvml.h
+++ b/discover/gpu_info_nvml.h
@@ -1,48 +0,0 @@
-#ifndef __APPLE__
-#ifndef __GPU_INFO_NVML_H__
-#define __GPU_INFO_NVML_H__
-#include "gpu_info.h"
-
-// Just enough typedef's to dlopen/dlsym for memory information
-typedef enum nvmlReturn_enum {
-  NVML_SUCCESS = 0,
-  // Other values omitted for now...
-} nvmlReturn_t;
-typedef void *nvmlDevice_t;  // Opaque is sufficient
-typedef struct nvmlMemory_st {
-  unsigned long long total;
-  unsigned long long free;
-  unsigned long long used;
-} nvmlMemory_t;
-
-typedef enum nvmlBrandType_enum
-{
-    NVML_BRAND_UNKNOWN          = 0,
-} nvmlBrandType_t;
-
-typedef struct nvml_handle {
-  void *handle;
-  uint16_t verbose;
-  nvmlReturn_t (*nvmlInit_v2)(void);
-  nvmlReturn_t (*nvmlShutdown)(void);
-  nvmlReturn_t (*nvmlDeviceGetHandleByUUID)(const char *, nvmlDevice_t *);
-  nvmlReturn_t (*nvmlDeviceGetMemoryInfo)(nvmlDevice_t, nvmlMemory_t *);
-} nvml_handle_t;
-
-typedef struct nvml_init_resp {
-  char *err;  // If err is non-null handle is invalid
-  nvml_handle_t ch;
-} nvml_init_resp_t;
-
-typedef struct nvml_compute_capability {
-  char *err;
-  int major;
-  int minor;
-} nvml_compute_capability_t;
-
-void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp);
-void nvml_get_free(nvml_handle_t ch, char *uuid, uint64_t *free, uint64_t *total, uint64_t *used);
-void nvml_release(nvml_handle_t ch);
-
-#endif  // __GPU_INFO_NVML_H__
-#endif  // __APPLE__
--- a/discover/gpu_info_oneapi.c
+++ b/discover/gpu_info_oneapi.c
@@ -1,259 +0,0 @@
-#ifndef __APPLE__
-
-#include "gpu_info_oneapi.h"
-
-#include <string.h>
-
-void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
-  ze_result_t ret;
-  resp->err = NULL;
-  resp->oh.devices = NULL;
-  resp->oh.num_devices = NULL;
-  resp->oh.drivers = NULL;
-  resp->oh.num_drivers = 0;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i, d;
-  struct lookup {
-    char *s;
-    void **p;
-  } l[] = {
-      {"zesInit", (void *)&resp->oh.zesInit},
-      {"zesDriverGet", (void *)&resp->oh.zesDriverGet},
-      {"zesDeviceGet", (void *)&resp->oh.zesDeviceGet},
-      {"zesDeviceGetProperties", (void *)&resp->oh.zesDeviceGetProperties},
-      {"zesDeviceEnumMemoryModules",
-       (void *)&resp->oh.zesDeviceEnumMemoryModules},
-      {"zesMemoryGetProperties", (void *)&resp->oh.zesMemoryGetProperties},
-      {"zesMemoryGetState", (void *)&resp->oh.zesMemoryGetState},
-      {NULL, NULL},
-  };
-
-  resp->oh.handle = LOAD_LIBRARY(oneapi_lib_path, RTLD_LAZY);
-  if (!resp->oh.handle) {
-    char *msg = LOAD_ERR();
-    snprintf(buf, buflen,
-             "Unable to load %s library to query for Intel GPUs: %s\n",
-             oneapi_lib_path, msg);
-    free(msg);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  // TODO once we've squashed the remaining corner cases remove this log
-  LOG(resp->oh.verbose,
-      "wiring Level-Zero management library functions in %s\n",
-      oneapi_lib_path);
-
-  for (i = 0; l[i].s != NULL; i++) {
-    // TODO once we've squashed the remaining corner cases remove this log
-    LOG(resp->oh.verbose, "dlsym: %s\n", l[i].s);
-
-    *l[i].p = LOAD_SYMBOL(resp->oh.handle, l[i].s);
-    if (!*(l[i].p)) {
-      resp->oh.handle = NULL;
-      char *msg = LOAD_ERR();
-      LOG(resp->oh.verbose, "dlerr: %s\n", msg);
-      UNLOAD_LIBRARY(resp->oh.handle);
-      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s, msg);
-      free(msg);
-      resp->err = strdup(buf);
-      return;
-    }
-  }
-
-  LOG(resp->oh.verbose, "calling zesInit\n");
-
-  ret = (*resp->oh.zesInit)(0);
-  if (ret != ZE_RESULT_SUCCESS) {
-    LOG(resp->oh.verbose, "zesInit err: %x\n", ret);
-    snprintf(buf, buflen, "oneapi vram init failure: %x", ret);
-    resp->err = strdup(buf);
-    oneapi_release(resp->oh);
-    return;
-  }
-
-  LOG(resp->oh.verbose, "calling zesDriverGet\n");
-  ret = (*resp->oh.zesDriverGet)(&resp->oh.num_drivers, NULL);
-  if (ret != ZE_RESULT_SUCCESS) {
-    LOG(resp->oh.verbose, "zesDriverGet err: %x\n", ret);
-    snprintf(buf, buflen, "unable to get driver count: %x", ret);
-    resp->err = strdup(buf);
-    oneapi_release(resp->oh);
-    return;
-  }
-  LOG(resp->oh.verbose, "oneapi driver count: %d\n", resp->oh.num_drivers);
-  resp->oh.drivers = malloc(resp->oh.num_drivers * sizeof(zes_driver_handle_t));
-  resp->oh.num_devices = malloc(resp->oh.num_drivers * sizeof(uint32_t));
-  memset(&resp->oh.num_devices[0], 0, resp->oh.num_drivers * sizeof(uint32_t));
-  resp->oh.devices =
-      malloc(resp->oh.num_drivers * sizeof(zes_device_handle_t *));
-  ret = (*resp->oh.zesDriverGet)(&resp->oh.num_drivers, &resp->oh.drivers[0]);
-  if (ret != ZE_RESULT_SUCCESS) {
-    LOG(resp->oh.verbose, "zesDriverGet err: %x\n", ret);
-    snprintf(buf, buflen, "unable to get driver count: %x", ret);
-    resp->err = strdup(buf);
-    oneapi_release(resp->oh);
-    return;
-  }
-
-  for (d = 0; d < resp->oh.num_drivers; d++) {
-    LOG(resp->oh.verbose, "calling zesDeviceGet count %d: %p\n", d, resp->oh.drivers[d]);
-    ret = (*resp->oh.zesDeviceGet)(resp->oh.drivers[d],
-                                   &resp->oh.num_devices[d], NULL);
-    if (ret != ZE_RESULT_SUCCESS) {
-      LOG(resp->oh.verbose, "zesDeviceGet err: %x\n", ret);
-      snprintf(buf, buflen, "unable to get device count: %x", ret);
-      resp->err = strdup(buf);
-      oneapi_release(resp->oh);
-      return;
-    }
-    resp->oh.devices[d] =
-        malloc(resp->oh.num_devices[d] * sizeof(zes_device_handle_t));
-    ret = (*resp->oh.zesDeviceGet)(
-        resp->oh.drivers[d], &resp->oh.num_devices[d], resp->oh.devices[d]);
-    if (ret != ZE_RESULT_SUCCESS) {
-      LOG(resp->oh.verbose, "zesDeviceGet err: %x\n", ret);
-      snprintf(buf, buflen, "unable to get device count: %x", ret);
-      resp->err = strdup(buf);
-      oneapi_release(resp->oh);
-      return;
-    }
-  }
-
-  return;
-}
-
-void oneapi_check_vram(oneapi_handle_t h, int driver, int device,
-                       mem_info_t *resp) {
-  ze_result_t ret;
-  resp->err = NULL;
-  uint64_t totalMem = 0;
-  uint64_t usedMem = 0;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i, d, m;
-
-  if (h.handle == NULL) {
-    resp->err = strdup("Level-Zero handle not initialized");
-    return;
-  }
-
-  if (driver > h.num_drivers || device > h.num_devices[driver]) {
-    resp->err = strdup("driver of device index out of bounds");
-    return;
-  }
-
-  resp->total = 0;
-  resp->free = 0;
-
-  zes_device_ext_properties_t ext_props;
-  ext_props.stype = ZES_STRUCTURE_TYPE_DEVICE_EXT_PROPERTIES;
-  ext_props.pNext = NULL;
-
-  zes_device_properties_t props;
-  props.stype = ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES;
-  props.pNext = &ext_props;
-
-  ret = (*h.zesDeviceGetProperties)(h.devices[driver][device], &props);
-  if (ret != ZE_RESULT_SUCCESS) {
-    snprintf(buf, buflen, "unable to get device properties: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  snprintf(&resp->gpu_name[0], GPU_NAME_LEN, "%s", props.modelName);
-
-  // TODO this needs to map to ONEAPI_DEVICE_SELECTOR syntax
-  // (this is probably wrong...)
-  // TODO - the driver isn't included - what if there are multiple drivers?
-  snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", device);
-
-  if (h.verbose) {
-    // When in verbose mode, report more information about
-    // the card we discover.
-    LOG(h.verbose, "[%d:%d] oneAPI device name: %s\n", driver, device,
-        props.modelName);
-    LOG(h.verbose, "[%d:%d] oneAPI brand: %s\n", driver, device,
-        props.brandName);
-    LOG(h.verbose, "[%d:%d] oneAPI vendor: %s\n", driver, device,
-        props.vendorName);
-    LOG(h.verbose, "[%d:%d] oneAPI S/N: %s\n", driver, device,
-        props.serialNumber);
-    LOG(h.verbose, "[%d:%d] oneAPI board number: %s\n", driver, device,
-        props.boardNumber);
-  }
-
-  // TODO
-  // Compute Capability equivalent in resp->major, resp->minor, resp->patch
-
-  uint32_t memCount = 0;
-  ret = (*h.zesDeviceEnumMemoryModules)(h.devices[driver][device], &memCount,
-                                        NULL);
-  if (ret != ZE_RESULT_SUCCESS) {
-    snprintf(buf, buflen, "unable to enumerate Level-Zero memory modules: %x",
-             ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  LOG(h.verbose, "discovered %d Level-Zero memory modules\n", memCount);
-
-  zes_mem_handle_t *mems = malloc(memCount * sizeof(zes_mem_handle_t));
-  (*h.zesDeviceEnumMemoryModules)(h.devices[driver][device], &memCount, mems);
-
-  for (m = 0; m < memCount; m++) {
-    zes_mem_state_t state;
-    state.stype = ZES_STRUCTURE_TYPE_MEM_STATE;
-    state.pNext = NULL;
-    ret = (*h.zesMemoryGetState)(mems[m], &state);
-    if (ret != ZE_RESULT_SUCCESS) {
-      snprintf(buf, buflen, "unable to get memory state: %x", ret);
-      resp->err = strdup(buf);
-      free(mems);
-      return;
-    }
-
-    resp->total += state.size;
-    resp->free += state.free;
-  }
-
-  free(mems);
-}
-
-void oneapi_release(oneapi_handle_t h) {
-  int d;
-  LOG(h.verbose, "releasing oneapi library\n");
-  for (d = 0; d < h.num_drivers; d++) {
-    if (h.devices != NULL && h.devices[d] != NULL) {
-      free(h.devices[d]);
-    }
-  }
-  if (h.devices != NULL) {
-    free(h.devices);
-    h.devices = NULL;
-  }
-  if (h.num_devices != NULL) {
-    free(h.num_devices);
-    h.num_devices = NULL;
-  }
-  if (h.drivers != NULL) {
-    free(h.drivers);
-    h.drivers = NULL;
-  }
-  h.num_drivers = 0;
-  UNLOAD_LIBRARY(h.handle);
-  h.handle = NULL;
-}
-
-int oneapi_get_device_count(oneapi_handle_t h, int driver) {
-  if (h.handle == NULL || h.num_devices == NULL) {
-    return 0;
-  }
-  if (driver > h.num_drivers) {
-    return 0;
-  }
-  return (int)h.num_devices[driver];
-}
-
-#endif // __APPLE__
--- a/discover/gpu_info_oneapi.h
+++ b/discover/gpu_info_oneapi.h
@@ -1,203 +0,0 @@
-#ifndef __APPLE__
-#ifndef __GPU_INFO_ONEAPI_H__
-#define __GPU_INFO_ONEAPI_H__
-#include "gpu_info.h"
-
-#define ZE_MAX_DEVICE_NAME 256
-#define ZE_MAX_DEVICE_UUID_SIZE 16
-#define ZES_STRING_PROPERTY_SIZE 64
-#define ZE_BIT(_i) (1 << _i)
-
-// Just enough typedef's to dlopen/dlsym for memory information
-typedef enum ze_result_t {
-  ZE_RESULT_SUCCESS = 0,
-  // Other values omitted for now...
-} ze_result_t;
-
-typedef uint8_t ze_bool_t;
-typedef struct _zes_driver_handle_t *zes_driver_handle_t;
-typedef struct _zes_device_handle_t *zes_device_handle_t;
-typedef struct _zes_mem_handle_t *zes_mem_handle_t;
-
-typedef enum _ze_structure_type_t {
-  ZE_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
-} ze_structure_type_t;
-
-typedef enum _zes_structure_type_t {
-  ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES = 0x1,
-  ZES_STRUCTURE_TYPE_MEM_PROPERTIES = 0xb,
-  ZES_STRUCTURE_TYPE_MEM_STATE = 0x1e,
-  ZES_STRUCTURE_TYPE_DEVICE_EXT_PROPERTIES = 0x2d,
-  ZES_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
-} zes_structure_type_t;
-
-typedef enum _zes_mem_type_t {
-  ZES_MEM_TYPE_FORCE_UINT32 = 0x7fffffff
-} zes_mem_type_t;
-
-typedef enum _zes_mem_loc_t {
-  ZES_MEM_LOC_SYSTEM = 0,
-  ZES_MEM_LOC_DEVICE = 1,
-  ZES_MEM_LOC_FORCE_UINT32 = 0x7fffffff
-} zes_mem_loc_t;
-
-typedef enum _zes_mem_health_t {
-  ZES_MEM_HEALTH_FORCE_UINT32 = 0x7fffffff
-} zes_mem_health_t;
-
-typedef struct _ze_device_uuid_t {
-  uint8_t id[ZE_MAX_DEVICE_UUID_SIZE];
-} ze_device_uuid_t;
-
-typedef struct _zes_uuid_t {
-  uint8_t id[ZE_MAX_DEVICE_UUID_SIZE];
-} zes_uuid_t;
-
-typedef enum _ze_device_type_t {
-  ZE_DEVICE_TYPE_GPU = 1,
-  ZE_DEVICE_TYPE_CPU = 2,
-  ZE_DEVICE_TYPE_FPGA = 3,
-  ZE_DEVICE_TYPE_MCA = 4,
-  ZE_DEVICE_TYPE_VPU = 5,
-  ZE_DEVICE_TYPE_FORCE_UINT32 = 0x7fffffff
-} ze_device_type_t;
-
-typedef enum _zes_device_type_t {
-  ZES_DEVICE_TYPE_GPU = 1,
-  ZES_DEVICE_TYPE_CPU = 2,
-  ZES_DEVICE_TYPE_FPGA = 3,
-  ZES_DEVICE_TYPE_MCA = 4,
-  ZES_DEVICE_TYPE_VPU = 5,
-  ZES_DEVICE_TYPE_FORCE_UINT32 = 0x7fffffff
-} zes_device_type_t;
-
-typedef uint32_t ze_device_property_flags_t;
-typedef enum _ze_device_property_flag_t {
-  ZE_DEVICE_PROPERTY_FLAG_INTEGRATED = ZE_BIT(0),
-  ZE_DEVICE_PROPERTY_FLAG_SUBDEVICE = ZE_BIT(1),
-  ZE_DEVICE_PROPERTY_FLAG_ECC = ZE_BIT(2),
-  ZE_DEVICE_PROPERTY_FLAG_ONDEMANDPAGING = ZE_BIT(3),
-  ZE_DEVICE_PROPERTY_FLAG_FORCE_UINT32 = 0x7fffffff
-} ze_device_property_flag_t;
-
-typedef uint32_t zes_device_property_flags_t;
-typedef enum _zes_device_property_flag_t {
-  ZES_DEVICE_PROPERTY_FLAG_INTEGRATED = ZE_BIT(0),
-  ZES_DEVICE_PROPERTY_FLAG_SUBDEVICE = ZE_BIT(1),
-  ZES_DEVICE_PROPERTY_FLAG_ECC = ZE_BIT(2),
-  ZES_DEVICE_PROPERTY_FLAG_ONDEMANDPAGING = ZE_BIT(3),
-  ZES_DEVICE_PROPERTY_FLAG_FORCE_UINT32 = 0x7fffffff
-} zes_device_property_flag_t;
-
-typedef struct _ze_device_properties_t {
-  ze_structure_type_t stype;
-  void *pNext;
-  ze_device_type_t type;
-  uint32_t vendorId;
-  uint32_t deviceId;
-  ze_device_property_flags_t flags;
-  uint32_t subdeviceId;
-  uint32_t coreClockRate;
-  uint64_t maxMemAllocSize;
-  uint32_t maxHardwareContexts;
-  uint32_t maxCommandQueuePriority;
-  uint32_t numThreadsPerEU;
-  uint32_t physicalEUSimdWidth;
-  uint32_t numEUsPerSubslice;
-  uint32_t numSubslicesPerSlice;
-  uint32_t numSlices;
-  uint64_t timerResolution;
-  uint32_t timestampValidBits;
-  uint32_t kernelTimestampValidBits;
-  ze_device_uuid_t uuid;
-  char name[ZE_MAX_DEVICE_NAME];
-} ze_device_properties_t;
-
-typedef struct _zes_device_properties_t {
-  zes_structure_type_t stype;
-  void *pNext;
-  ze_device_properties_t core;
-  uint32_t numSubdevices;
-  char serialNumber[ZES_STRING_PROPERTY_SIZE];
-  char boardNumber[ZES_STRING_PROPERTY_SIZE];
-  char brandName[ZES_STRING_PROPERTY_SIZE];
-  char modelName[ZES_STRING_PROPERTY_SIZE];
-  char vendorName[ZES_STRING_PROPERTY_SIZE];
-  char driverVersion[ZES_STRING_PROPERTY_SIZE];
-} zes_device_properties_t;
-
-typedef struct _zes_device_ext_properties_t {
-  zes_structure_type_t stype;
-  void *pNext;
-  zes_uuid_t uuid;
-  zes_device_type_t type;
-  zes_device_property_flags_t flags;
-} zes_device_ext_properties_t;
-
-typedef struct _zes_mem_properties_t {
-  zes_structure_type_t stype;
-  void *pNext;
-  zes_mem_type_t type;
-  ze_bool_t onSubdevice;
-  uint32_t subdeviceId;
-  zes_mem_loc_t location;
-  uint64_t physicalSize;
-  int32_t busWidth;
-  int32_t numChannels;
-} zes_mem_properties_t;
-
-typedef struct _zes_mem_state_t {
-  zes_structure_type_t stype;
-  const void *pNext;
-  zes_mem_health_t health;
-  uint64_t free;
-  uint64_t size;
-} zes_mem_state_t;
-
-typedef struct oneapi_handle {
-  void *handle;
-  uint16_t verbose;
-
-  uint32_t num_drivers;
-  zes_driver_handle_t *drivers;
-  uint32_t *num_devices;
-  zes_device_handle_t **devices;
-
-  // TODO Driver major, minor information
-  // int driver_major;
-  // int driver_minor;
-
-  ze_result_t (*zesInit)(int);
-  ze_result_t (*zesDriverGet)(uint32_t *pCount, zes_driver_handle_t *phDrivers);
-  ze_result_t (*zesDeviceGet)(zes_driver_handle_t hDriver, uint32_t *pCount,
-                              zes_device_handle_t *phDevices);
-  ze_result_t (*zesDeviceGetProperties)(zes_device_handle_t hDevice,
-                                        zes_device_properties_t *pProperties);
-  ze_result_t (*zesDeviceEnumMemoryModules)(zes_device_handle_t hDevice,
-                                            uint32_t *pCount,
-                                            zes_mem_handle_t *phMemory);
-  ze_result_t (*zesMemoryGetProperties)(zes_mem_handle_t hMemory,
-                                        zes_mem_properties_t *pProperties);
-  ze_result_t (*zesMemoryGetState)(zes_mem_handle_t hMemory,
-                                   zes_mem_state_t *pState);
-
-} oneapi_handle_t;
-
-typedef struct oneapi_init_resp {
-  char *err; // If err is non-null handle is invalid
-  oneapi_handle_t oh;
-} oneapi_init_resp_t;
-
-typedef struct oneapi_version_resp {
-  ze_result_t status;
-  char *str; // Contains version or error string if status != 0
-} oneapi_version_resp_t;
-
-void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp);
-void oneapi_check_vram(oneapi_handle_t h, int driver, int device,
-                       mem_info_t *resp);
-void oneapi_release(oneapi_handle_t h);
-int oneapi_get_device_count(oneapi_handle_t h, int driver);
-
-#endif // __GPU_INFO_INTEL_H__
-#endif // __APPLE__
--- a/discover/gpu_oneapi.go
+++ b/discover/gpu_oneapi.go
@@ -1,21 +0,0 @@
-//go:build linux || windows
-
-package discover
-
-import (
-	"log/slog"
-	"strings"
-)
-
-func oneapiGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
-	ids := []string{}
-	for _, info := range gpuInfo {
-		if info.Library != "oneapi" {
-			// TODO shouldn't happen if things are wired correctly...
-			slog.Debug("oneapiGetVisibleDevicesEnv skipping over non-sycl device", "library", info.Library)
-			continue
-		}
-		ids = append(ids, info.ID)
-	}
-	return "ONEAPI_DEVICE_SELECTOR", "level_zero:" + strings.Join(ids, ",")
-}
--- a/discover/gpu_test.go
+++ b/discover/gpu_test.go
@@ -1,60 +0,0 @@
-package discover
-
-import (
-	"runtime"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestBasicGetGPUInfo(t *testing.T) {
-	info := GetGPUInfo()
-	assert.NotEmpty(t, len(info))
-	assert.Contains(t, "cuda rocm cpu metal", info[0].Library)
-	if info[0].Library != "cpu" {
-		assert.Greater(t, info[0].TotalMemory, uint64(0))
-		assert.Greater(t, info[0].FreeMemory, uint64(0))
-	}
-}
-
-func TestCPUMemInfo(t *testing.T) {
-	info, err := GetCPUMem()
-	require.NoError(t, err)
-	switch runtime.GOOS {
-	case "darwin":
-		t.Skip("CPU memory not populated on darwin")
-	case "linux", "windows":
-		assert.Greater(t, info.TotalMemory, uint64(0))
-		assert.Greater(t, info.FreeMemory, uint64(0))
-	default:
-		return
-	}
-}
-
-func TestByLibrary(t *testing.T) {
-	type testCase struct {
-		input  []GpuInfo
-		expect int
-	}
-
-	testCases := map[string]*testCase{
-		"empty":                    {input: []GpuInfo{}, expect: 0},
-		"cpu":                      {input: []GpuInfo{{Library: "cpu"}}, expect: 1},
-		"cpu + GPU":                {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda"}}, expect: 2},
-		"cpu + 2 GPU no variant":   {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda"}, {Library: "cuda"}}, expect: 2},
-		"cpu + 2 GPU same variant": {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda", Variant: "v11"}, {Library: "cuda", Variant: "v11"}}, expect: 2},
-		"cpu + 2 GPU diff variant": {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda", Variant: "v11"}, {Library: "cuda", Variant: "v12"}}, expect: 3},
-	}
-
-	for k, v := range testCases {
-		t.Run(k, func(t *testing.T) {
-			resp := (GpuInfoList)(v.input).ByLibrary()
-			if len(resp) != v.expect {
-				t.Fatalf("expected length %d, got %d => %+v", v.expect, len(resp), resp)
-			}
-		})
-	}
-}
-
-// TODO - add some logic to figure out card type through other means and actually verify we got back what we expected
--- a/discover/runner.go
+++ b/discover/runner.go
@@ -0,0 +1,590 @@
+package discover
+
+// Runner based GPU discovery
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"math/rand"
+	"net"
+	"net/http"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/logutil"
+	"github.com/ollama/ollama/ml"
+)
+
+var (
+	deviceMu     sync.Mutex
+	devices      []ml.DeviceInfo
+	libDirs      map[string]struct{}
+	rocmDir      string
+	exe          string
+	bootstrapped bool
+)
+
+func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.DeviceInfo {
+	deviceMu.Lock()
+	defer deviceMu.Unlock()
+	startDiscovery := time.Now()
+	msg := "overall device VRAM discovery took"
+	defer func() {
+		slog.Debug(msg, "duration", time.Since(startDiscovery))
+	}()
+
+	if !bootstrapped {
+		msg = "GPU bootstrap discovery took"
+		libDirs = make(map[string]struct{})
+		var err error
+		exe, err = os.Executable()
+		if err != nil {
+			slog.Error("unable to lookup executable path", "error", err)
+			return nil
+		}
+		if eval, err := filepath.EvalSymlinks(exe); err == nil {
+			exe = eval
+		}
+		files, err := filepath.Glob(filepath.Join(LibOllamaPath, "*", "*ggml-*"))
+		if err != nil {
+			slog.Debug("unable to lookup runner library directories", "error", err)
+		}
+		for _, file := range files {
+			libDirs[filepath.Dir(file)] = struct{}{}
+		}
+
+		// Our current packaging model places ggml-hip in the main directory
+		// but keeps rocm in an isolated directory.  We have to add it to
+		// the [LD_LIBRARY_]PATH so ggml-hip will load properly
+		rocmDir = filepath.Join(LibOllamaPath, "rocm")
+		if _, err := os.Stat(rocmDir); err != nil {
+			rocmDir = ""
+		}
+
+		if len(libDirs) == 0 {
+			libDirs[""] = struct{}{}
+		}
+
+		slog.Info("discovering available GPUs...")
+		requested := envconfig.LLMLibrary()
+		jetpack := cudaJetpack()
+
+		// For our initial discovery pass, we gather all the known GPUs through
+		// all the libraries that were detected. This pass may include GPUs that
+		// are enumerated, but not actually supported.
+		// We run this in serial to avoid potentially initializing a GPU multiple
+		// times concurrently leading to memory contention
+		for dir := range libDirs {
+			var dirs []string
+			if dir != "" {
+				if requested != "" && filepath.Base(dir) != requested {
+					slog.Debug("skipping available library at users request", "requested", requested, "libDir", dir)
+					continue
+				} else if jetpack != "" && filepath.Base(dir) != "cuda_"+jetpack {
+					continue
+				}
+			}
+			if dir == "" {
+				dirs = []string{LibOllamaPath}
+			} else {
+				dirs = []string{LibOllamaPath, dir}
+			}
+			// Typically bootstrapping takes < 1s, but on some systems, with devices
+			// in low power/idle mode, initialization can take multiple seconds.  We
+			// set a long timeout just for bootstrap discovery to reduce the chance
+			// of giving up too quickly
+			ctx1stPass, cancel := context.WithTimeout(ctx, 30*time.Second)
+			defer cancel()
+
+			// For this pass, we retain duplicates in case any are incompatible with some libraries
+			devices = append(devices, bootstrapDevices(ctx1stPass, dirs, nil)...)
+		}
+
+		// In the second pass, we more deeply initialize the GPUs to weed out devices that
+		// aren't supported by a given library.  We run this phase in parallel to speed up discovery.
+		slog.Debug("filtering out unsupported or overlapping GPU library combinations", "count", len(devices))
+		ctx2ndPass, cancel := context.WithTimeout(ctx, 30*time.Second)
+		defer cancel()
+		var wg sync.WaitGroup
+		needsDelete := make([]bool, len(devices))
+		supportedMu := sync.Mutex{}
+		supported := make(map[string]map[string]map[string]int) // [Library][libDir][ID] = pre-deletion devices index
+		for i := range devices {
+			libDir := devices[i].LibraryPath[len(devices[i].LibraryPath)-1]
+			if devices[i].Library == "Metal" {
+				continue
+			}
+			slog.Debug("verifying GPU is supported", "library", libDir, "description", devices[i].Description, "compute", devices[i].Compute(), "pci_id", devices[i].PCIID)
+			wg.Add(1)
+			go func(i int) {
+				defer wg.Done()
+				var envVar string
+				if devices[i].Library == "ROCm" {
+					if runtime.GOOS != "linux" {
+						envVar = "HIP_VISIBLE_DEVICES"
+					} else {
+						envVar = "ROCR_VISIBLE_DEVICES"
+					}
+				} else {
+					envVar = "CUDA_VISIBLE_DEVICES"
+				}
+
+				extraEnvs := []string{
+					"GGML_CUDA_INIT=1",           // force deep initialization to trigger crash on unsupported GPUs
+					envVar + "=" + devices[i].ID, // Filter to just this one GPU
+				}
+				if len(bootstrapDevices(ctx2ndPass, devices[i].LibraryPath, extraEnvs)) == 0 {
+					needsDelete[i] = true
+				} else {
+					supportedMu.Lock()
+					if _, ok := supported[devices[i].Library]; !ok {
+						supported[devices[i].Library] = make(map[string]map[string]int)
+					}
+					if _, ok := supported[devices[i].Library][libDir]; !ok {
+						supported[devices[i].Library][libDir] = make(map[string]int)
+					}
+					supported[devices[i].Library][libDir][devices[i].ID] = i
+					supportedMu.Unlock()
+				}
+			}(i)
+		}
+		wg.Wait()
+		logutil.Trace("supported GPU library combinations", "supported", supported)
+
+		// Mark for deletion any overlaps - favoring the library version that can cover all GPUs if possible
+		filterOverlapByLibrary(supported, needsDelete)
+
+		// TODO if we ever support multiple ROCm library versions this algorithm will need to be adjusted to keep the rocmID numeric value correct
+		rocmID := 0
+		for i := 0; i < len(needsDelete); i++ {
+			if needsDelete[i] {
+				logutil.Trace("removing unsupported or overlapping GPU combination", "libDir", devices[i].LibraryPath[len(devices[i].LibraryPath)-1], "description", devices[i].Description, "compute", devices[i].Compute(), "pci_id", devices[i].PCIID)
+				devices = append(devices[:i], devices[i+1:]...)
+				needsDelete = append(needsDelete[:i], needsDelete[i+1:]...)
+				i--
+			} else if devices[i].Library == "ROCm" {
+				if _, err := strconv.Atoi(devices[i].ID); err == nil {
+					// Replace the numeric ID with the post-filtered IDs
+					devices[i].FilteredID = devices[i].ID
+					devices[i].ID = strconv.Itoa(rocmID)
+				}
+				rocmID++
+			}
+		}
+
+		// Now filter out any overlap with different libraries (favor CUDA/ROCm over others)
+		for i := 0; i < len(devices); i++ {
+			for j := i + 1; j < len(devices); j++ {
+				// For this pass, we only drop exact duplicates
+				switch devices[i].Compare(devices[j]) {
+				case ml.SameBackendDevice:
+					// Same library and device, skip it
+					devices = append(devices[:j], devices[j+1:]...)
+					j--
+					continue
+				case ml.DuplicateDevice:
+					// Different library, choose based on priority
+					var droppedDevice ml.DeviceInfo
+					if devices[i].Library == "CUDA" || devices[i].Library == "ROCm" {
+						droppedDevice = devices[j]
+					} else {
+						droppedDevice = devices[i]
+						devices[i] = devices[j]
+					}
+					devices = append(devices[:j], devices[j+1:]...)
+					j--
+
+					typeStr := "discrete"
+					if droppedDevice.Integrated {
+						typeStr = "iGPU"
+					}
+					slog.Debug("dropping duplicate device",
+						"id", droppedDevice.ID,
+						"library", droppedDevice.Library,
+						"compute", droppedDevice.Compute(),
+						"name", droppedDevice.Name,
+						"description", droppedDevice.Description,
+						"libdirs", strings.Join(droppedDevice.LibraryPath, ","),
+						"driver", droppedDevice.Driver(),
+						"pci_id", droppedDevice.PCIID,
+						"type", typeStr,
+						"total", format.HumanBytes2(droppedDevice.TotalMemory),
+						"available", format.HumanBytes2(droppedDevice.FreeMemory),
+					)
+					continue
+				}
+			}
+		}
+
+		// Reset the libDirs to what we actually wind up using for future refreshes
+		libDirs = make(map[string]struct{})
+		for _, dev := range devices {
+			dir := dev.LibraryPath[len(dev.LibraryPath)-1]
+			if dir != LibOllamaPath {
+				libDirs[dir] = struct{}{}
+			}
+		}
+		if len(libDirs) == 0 {
+			libDirs[""] = struct{}{}
+		}
+
+		bootstrapped = true
+	} else {
+		if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
+			// metal never updates free VRAM
+			return devices
+		}
+
+		slog.Debug("refreshing free memory")
+		updated := make([]bool, len(devices))
+		allDone := func() bool {
+			allDone := true
+			for _, done := range updated {
+				if !done {
+					allDone = false
+					break
+				}
+			}
+			return allDone
+		}
+
+		// First try to use existing runners to refresh VRAM since they're already
+		// active on GPU(s)
+		for _, runner := range runners {
+			if runner == nil {
+				continue
+			}
+			deviceIDs := runner.GetActiveDeviceIDs()
+			if len(deviceIDs) == 0 {
+				// Skip this runner since it doesn't have active GPU devices
+				continue
+			}
+
+			// Check to see if this runner is active on any devices that need a refresh
+			skip := true
+		devCheck:
+			for _, dev := range deviceIDs {
+				for i := range devices {
+					if dev == devices[i].DeviceID {
+						if !updated[i] {
+							skip = false
+							break devCheck
+						}
+					}
+				}
+			}
+			if skip {
+				continue
+			}
+
+			// Typical refresh on existing runner is ~500ms but allow longer if the system
+			// is under stress before giving up and using stale data.
+			ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
+			defer cancel()
+			start := time.Now()
+			updatedDevices := runner.GetDeviceInfos(ctx)
+			slog.Debug("existing runner discovery took", "duration", time.Since(start))
+			for _, u := range updatedDevices {
+				for i := range devices {
+					if u.DeviceID == devices[i].DeviceID {
+						updated[i] = true
+						devices[i].FreeMemory = u.FreeMemory
+						break
+					}
+				}
+			}
+			// Short circuit if we've updated all the devices
+			if allDone() {
+				break
+			}
+		}
+		if !allDone() {
+			slog.Debug("unable to refresh all GPUs with existing runners, performing bootstrap discovery")
+
+			// Bootstrapping may take longer in some cases (AMD windows), but we
+			// would rather use stale free data to get the model running sooner
+			ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
+			defer cancel()
+
+			for dir := range libDirs {
+				updatedDevices := bootstrapDevices(ctx, []string{LibOllamaPath, dir}, nil)
+				for _, u := range updatedDevices {
+					for i := range devices {
+						if u.DeviceID == devices[i].DeviceID {
+							updated[i] = true
+							devices[i].FreeMemory = u.FreeMemory
+							break
+						}
+					}
+					// TODO - consider evaluating if new devices have appeared (e.g. hotplug)
+				}
+				if allDone() {
+					break
+				}
+			}
+			if !allDone() {
+				slog.Warn("unable to refresh free memory, using old values")
+			}
+		}
+	}
+
+	// Apply any iGPU workarounds
+	iGPUWorkarounds(devices)
+
+	return devices
+}
+
+func filterOverlapByLibrary(supported map[string]map[string]map[string]int, needsDelete []bool) {
+	// For multi-GPU systems, use the newest version that supports all the GPUs
+	for _, byLibDirs := range supported {
+		libDirs := make([]string, 0, len(byLibDirs))
+		for libDir := range byLibDirs {
+			libDirs = append(libDirs, libDir)
+		}
+		sort.Sort(sort.Reverse(sort.StringSlice(libDirs)))
+		anyMissing := false
+		var newest string
+		for _, newest = range libDirs {
+			for _, libDir := range libDirs {
+				if libDir == newest {
+					continue
+				}
+				if len(byLibDirs[newest]) != len(byLibDirs[libDir]) {
+					anyMissing = true
+					break
+				}
+				for dev := range byLibDirs[newest] {
+					if _, found := byLibDirs[libDir][dev]; !found {
+						anyMissing = true
+						break
+					}
+				}
+			}
+			if !anyMissing {
+				break
+			}
+		}
+		// Now we can mark overlaps for deletion
+		for _, libDir := range libDirs {
+			if libDir == newest {
+				continue
+			}
+			for dev, i := range byLibDirs[libDir] {
+				if _, found := byLibDirs[newest][dev]; found {
+					needsDelete[i] = true
+				}
+			}
+		}
+	}
+}
+
+type bootstrapRunner struct {
+	port int
+	cmd  *exec.Cmd
+}
+
+func (r *bootstrapRunner) GetPort() int {
+	return r.port
+}
+
+func (r *bootstrapRunner) HasExited() bool {
+	if r.cmd != nil && r.cmd.ProcessState != nil {
+		return true
+	}
+	return false
+}
+
+func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []string) []ml.DeviceInfo {
+	// TODO DRY out with llm/server.go
+	slog.Debug("spawing runner with", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs)
+	start := time.Now()
+	defer func() {
+		slog.Debug("bootstrap discovery took", "duration", time.Since(start), "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs)
+	}()
+	port := 0
+	if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
+		var l *net.TCPListener
+		if l, err = net.ListenTCP("tcp", a); err == nil {
+			port = l.Addr().(*net.TCPAddr).Port
+			l.Close()
+		}
+	}
+	if port == 0 {
+		slog.Debug("ResolveTCPAddr failed, using random port")
+		port = rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
+	}
+	params := []string{"runner", "--ollama-engine", "--port", strconv.Itoa(port)}
+	var pathEnv string
+	switch runtime.GOOS {
+	case "windows":
+		pathEnv = "PATH"
+	case "darwin":
+		pathEnv = "DYLD_LIBRARY_PATH"
+	default:
+		pathEnv = "LD_LIBRARY_PATH"
+	}
+	libraryPaths := append([]string{LibOllamaPath}, ollamaLibDirs...)
+	if rocmDir != "" {
+		libraryPaths = append(libraryPaths, rocmDir)
+	}
+	// Note: we always put our dependency paths first
+	// since these are the exact version we compiled/linked against
+	if libraryPath, ok := os.LookupEnv(pathEnv); ok {
+		libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
+	}
+
+	cmd := exec.Command(exe, params...)
+	cmd.Env = os.Environ()
+	if envconfig.LogLevel() == logutil.LevelTrace {
+		cmd.Stdout = os.Stdout
+		cmd.Stderr = os.Stderr
+	}
+	// cmd.SysProcAttr = llm.LlamaServerSysProcAttr // circular dependency - bring back once refactored
+	pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))
+	pathNeeded := true
+	ollamaPathNeeded := true
+	extraDone := make([]bool, len(extraEnvs))
+	for i := range cmd.Env {
+		cmp := strings.SplitN(cmd.Env[i], "=", 2)
+		if strings.EqualFold(cmp[0], pathEnv) {
+			cmd.Env[i] = pathEnv + "=" + pathEnvVal
+			pathNeeded = false
+		} else if strings.EqualFold(cmp[0], "OLLAMA_LIBRARY_PATH") {
+			cmd.Env[i] = "OLLAMA_LIBRARY_PATH=" + strings.Join(ollamaLibDirs, string(filepath.ListSeparator))
+			ollamaPathNeeded = false
+		} else {
+			for j := range extraEnvs {
+				if extraDone[j] {
+					continue
+				}
+				extra := strings.SplitN(extraEnvs[j], "=", 2)
+				if cmp[0] == extra[0] {
+					cmd.Env[i] = extraEnvs[j]
+					extraDone[j] = true
+				}
+			}
+		}
+	}
+	if pathNeeded {
+		cmd.Env = append(cmd.Env, pathEnv+"="+pathEnvVal)
+	}
+	if ollamaPathNeeded {
+		cmd.Env = append(cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ollamaLibDirs, string(filepath.ListSeparator)))
+	}
+	for i := range extraDone {
+		if !extraDone[i] {
+			cmd.Env = append(cmd.Env, extraEnvs[i])
+		}
+	}
+	logutil.Trace("starting runner for device discovery", "env", cmd.Env, "cmd", cmd)
+	if err := cmd.Start(); err != nil {
+		slog.Warn("unable to start discovery subprocess", "cmd", cmd, "error", err)
+		return nil
+	}
+	go func() {
+		cmd.Wait() // exit status ignored
+	}()
+
+	defer cmd.Process.Kill()
+	devices, err := GetDevicesFromRunner(ctx, &bootstrapRunner{port: port, cmd: cmd})
+	if err != nil {
+		if cmd.ProcessState != nil && cmd.ProcessState.ExitCode() >= 0 {
+			// Expected during bootstrapping while we filter out unsupported AMD GPUs
+			logutil.Trace("runner exited", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs, "code", cmd.ProcessState.ExitCode())
+		} else {
+			slog.Info("failure during GPU discovery", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs, "error", err)
+		}
+	}
+	logutil.Trace("runner enumerated devices", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "devices", devices)
+	return devices
+}
+
+func GetDevicesFromRunner(ctx context.Context, runner BaseRunner) ([]ml.DeviceInfo, error) {
+	var moreDevices []ml.DeviceInfo
+	port := runner.GetPort()
+	tick := time.Tick(10 * time.Millisecond)
+	for {
+		select {
+		case <-ctx.Done():
+			return nil, fmt.Errorf("failed to finish discovery before timeout")
+		case <-tick:
+			r, err := http.NewRequestWithContext(ctx, http.MethodGet, fmt.Sprintf("http://127.0.0.1:%d/info", port), nil)
+			if err != nil {
+				return nil, fmt.Errorf("failed to create request: %w", err)
+			}
+			r.Header.Set("Content-Type", "application/json")
+
+			resp, err := http.DefaultClient.Do(r)
+			if err != nil {
+				// slog.Warn("failed to send request", "error", err)
+				if runner.HasExited() {
+					return nil, fmt.Errorf("runner crashed")
+				}
+				continue
+			}
+			defer resp.Body.Close()
+
+			if resp.StatusCode == http.StatusNotFound {
+				// old runner, fall back to bootstrapping model
+				return nil, fmt.Errorf("llamarunner free vram reporting not supported")
+			}
+
+			body, err := io.ReadAll(resp.Body)
+			if err != nil {
+				slog.Warn("failed to read response", "error", err)
+				continue
+			}
+			if resp.StatusCode != 200 {
+				logutil.Trace("runner failed to discover free VRAM", "status", resp.StatusCode, "response", body)
+				return nil, fmt.Errorf("runner error: %s", string(body))
+			}
+
+			if err := json.Unmarshal(body, &moreDevices); err != nil {
+				slog.Warn("unmarshal encode response", "error", err)
+				continue
+			}
+			return moreDevices, nil
+		}
+	}
+}
+
+func iGPUWorkarounds(devices []ml.DeviceInfo) {
+	// short circuit if we have no iGPUs
+	anyiGPU := false
+	for i := range devices {
+		if devices[i].Integrated {
+			anyiGPU = true
+			break
+		}
+	}
+	if !anyiGPU {
+		return
+	}
+
+	memInfo, err := GetCPUMem()
+	if err != nil {
+		slog.Debug("failed to fetch system memory information for iGPU", "error", err)
+		return
+	}
+	for i := range devices {
+		if !devices[i].Integrated {
+			continue
+		}
+		// NVIDIA iGPUs return useless free VRAM data which ignores system buff/cache
+		if devices[i].Library == "CUDA" {
+			devices[i].FreeMemory = memInfo.FreeMemory
+		}
+	}
+}
--- a/discover/runner_test.go
+++ b/discover/runner_test.go
@@ -0,0 +1,108 @@
+package discover
+
+import (
+	"testing"
+
+	"github.com/ollama/ollama/app/lifecycle"
+)
+
+func init() {
+	lifecycle.InitLogging()
+}
+
+func TestFilterOverlapByLibrary(t *testing.T) {
+	type testcase struct {
+		name string
+		inp  map[string]map[string]map[string]int
+		exp  []bool
+	}
+	for _, tc := range []testcase{
+		{
+			name: "empty",
+			inp:  map[string]map[string]map[string]int{},
+			exp:  []bool{}, // needs deletion
+		},
+		{
+			name: "single no overlap",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v12": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+					},
+				},
+			},
+			exp: []bool{false},
+		},
+		{
+			name: "100% overlap pick 2nd",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v12": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 1,
+					},
+					"cuda_v13": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 2,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 3,
+					},
+				},
+			},
+			exp: []bool{true, true, false, false},
+		},
+		{
+			name: "100% overlap pick 1st",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v13": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 1,
+					},
+					"cuda_v12": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 2,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 3,
+					},
+				},
+			},
+			exp: []bool{false, false, true, true},
+		},
+		{
+			name: "partial overlap pick older",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v13": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+					},
+					"cuda_v12": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 1,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 2,
+					},
+				},
+			},
+			exp: []bool{true, false, false},
+		},
+		{
+			name: "no overlap",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v13": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+					},
+					"cuda_v12": {
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 1,
+					},
+				},
+			},
+			exp: []bool{false, false},
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			needsDelete := make([]bool, len(tc.exp))
+			filterOverlapByLibrary(tc.inp, needsDelete)
+			for i, exp := range tc.exp {
+				if needsDelete[i] != exp {
+					t.Fatalf("expected: %v\ngot: %v", tc.exp, needsDelete)
+				}
+			}
+		})
+	}
+}
--- a/discover/types.go
+++ b/discover/types.go
@@ -1,10 +1,14 @@
 package discover

 import (
-	"fmt"
+	"context"
 	"log/slog"
+	"path/filepath"
+	"runtime"
+	"strings"

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

 type memInfo struct {
@@ -15,8 +19,8 @@ type memInfo struct {

 // Beginning of an `ollama info` command
 type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
+	ml.DeviceID
 	memInfo
-	Library string `json:"library,omitempty"`

 	// Optional variant to select (e.g. versions, cpu feature flags)
 	Variant string `json:"variant"`
@@ -27,18 +31,16 @@ type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
 	// Any extra PATH/LD_LIBRARY_PATH dependencies required for the Library to operate properly
 	DependencyPath []string `json:"lib_path,omitempty"`

-	// Extra environment variables specific to the GPU as list of [key,value]
-	EnvWorkarounds [][2]string `json:"envs,omitempty"`
-
 	// Set to true if we can NOT reliably discover FreeMemory.  A value of true indicates
 	// the FreeMemory is best effort, and may over or under report actual memory usage
 	// False indicates FreeMemory can generally be trusted on this GPU
 	UnreliableFreeMemory bool

 	// GPU information
-	ID      string `json:"gpu_id"`  // string to use for selection of this specific GPU
-	Name    string `json:"name"`    // user friendly name if available
-	Compute string `json:"compute"` // Compute Capability or gfx
+	filterID     string // AMD Workaround: The numeric ID of the device used to filter out other devices
+	Name         string `json:"name"`          // user friendly name if available
+	ComputeMajor int    `json:"compute_major"` // Compute Capability or gfx
+	ComputeMinor int    `json:"compute_minor"`

 	// Driver Information - TODO no need to put this on each GPU
 	DriverMajor int `json:"driver_major,omitempty"`
@@ -69,37 +71,8 @@ type CPU struct {
 	ThreadCount         int
 }

-type CudaGPUInfo struct {
-	GpuInfo
-	OSOverhead   uint64 // Memory overhead between the driver library and management library
-	index        int    //nolint:unused,nolintlint
-	computeMajor int    //nolint:unused,nolintlint
-	computeMinor int    //nolint:unused,nolintlint
-}
-type CudaGPUInfoList []CudaGPUInfo
-
-type RocmGPUInfo struct {
-	GpuInfo
-	usedFilepath string //nolint:unused,nolintlint
-	index        int    //nolint:unused,nolintlint
-}
-type RocmGPUInfoList []RocmGPUInfo
-
-type OneapiGPUInfo struct {
-	GpuInfo
-	driverIndex int //nolint:unused,nolintlint
-	gpuIndex    int //nolint:unused,nolintlint
-}
-type OneapiGPUInfoList []OneapiGPUInfo
-
 type GpuInfoList []GpuInfo

-type UnsupportedGPUInfo struct {
-	GpuInfo
-	Reason string `json:"reason"`
-}
-
-// Split up the set of gpu info's by Library and variant
 func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	resp := []GpuInfoList{}
 	libs := []string{}
@@ -124,18 +97,47 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	return resp
 }

-// Report the GPU information into the log an Info level
-func (l GpuInfoList) LogDetails() {
-	for _, g := range l {
+func LogDetails(devices []ml.DeviceInfo) {
+	for _, dev := range devices {
+		var libs []string
+		for _, dir := range dev.LibraryPath {
+			if strings.Contains(dir, filepath.Join("lib", "ollama")) {
+				libs = append(libs, filepath.Base(dir))
+			}
+		}
+		typeStr := "discrete"
+		if dev.Integrated {
+			typeStr = "iGPU"
+		}
 		slog.Info("inference compute",
-			"id", g.ID,
-			"library", g.Library,
-			"variant", g.Variant,
-			"compute", g.Compute,
-			"driver", fmt.Sprintf("%d.%d", g.DriverMajor, g.DriverMinor),
-			"name", g.Name,
-			"total", format.HumanBytes2(g.TotalMemory),
-			"available", format.HumanBytes2(g.FreeMemory),
+			"id", dev.ID,
+			"library", dev.Library,
+			"compute", dev.Compute(),
+			"name", dev.Name,
+			"description", dev.Description,
+			"libdirs", strings.Join(libs, ","),
+			"driver", dev.Driver(),
+			"pci_id", dev.PCIID,
+			"type", typeStr,
+			"total", format.HumanBytes2(dev.TotalMemory),
+			"available", format.HumanBytes2(dev.FreeMemory),
+		)
+	}
+	// CPU inference
+	if len(devices) == 0 {
+		dev, _ := GetCPUMem()
+		slog.Info("inference compute",
+			"id", "cpu",
+			"library", "cpu",
+			"compute", "",
+			"name", "cpu",
+			"description", "cpu",
+			"libdirs", "ollama",
+			"driver", "",
+			"pci_id", "",
+			"type", "",
+			"total", format.HumanBytes2(dev.TotalMemory),
+			"available", format.HumanBytes2(dev.FreeMemory),
 		)
 	}
 }
@@ -148,16 +150,15 @@ func (a ByFreeMemory) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
 func (a ByFreeMemory) Less(i, j int) bool { return a[i].FreeMemory < a[j].FreeMemory }

 type SystemInfo struct {
-	System          CPUInfo              `json:"system"`
-	GPUs            []GpuInfo            `json:"gpus"`
-	UnsupportedGPUs []UnsupportedGPUInfo `json:"unsupported_gpus"`
-	DiscoveryErrors []string             `json:"discovery_errors"`
+	System CPUInfo   `json:"system"`
+	GPUs   []GpuInfo `json:"gpus"`
 }

 // Return the optimal number of threads to use for inference
 func (si SystemInfo) GetOptimalThreadCount() int {
 	if len(si.System.CPUs) == 0 {
-		return 0
+		// Fall back to Go's num CPU
+		return runtime.NumCPU()
 	}

 	coreCount := 0
@@ -172,9 +173,9 @@ func (si SystemInfo) GetOptimalThreadCount() int {
 func (l GpuInfoList) FlashAttentionSupported() bool {
 	for _, gpu := range l {
 		supportsFA := gpu.Library == "cpu" ||
-			gpu.Library == "metal" ||
-			(gpu.Library == "cuda" && gpu.DriverMajor >= 7) ||
-			gpu.Library == "rocm"
+			gpu.Name == "Metal" || gpu.Library == "Metal" ||
+			(gpu.Library == "CUDA" && gpu.DriverMajor >= 7 && !(gpu.ComputeMajor == 7 && gpu.ComputeMinor == 2)) || // We don't have kernels for Jetson Xavier
+			gpu.Library == "ROCm"

 		if !supportsFA {
 			return false
@@ -182,3 +183,31 @@ func (l GpuInfoList) FlashAttentionSupported() bool {
 	}
 	return true
 }
+
+type BaseRunner interface {
+	// GetPort returns the localhost port number the runner is running on
+	GetPort() int
+
+	// HasExited indicates if the runner is no longer running.  This can be used during
+	// bootstrap to detect if a given filtered device is incompatible and triggered an assert
+	HasExited() bool
+}
+
+type RunnerDiscovery interface {
+	BaseRunner
+
+	// GetDeviceInfos will perform a query of the underlying device libraries
+	// for device identification and free VRAM information
+	// During bootstrap scenarios, this routine may take seconds to complete
+	GetDeviceInfos(ctx context.Context) []ml.DeviceInfo
+}
+
+type FilteredRunnerDiscovery interface {
+	RunnerDiscovery
+
+	// GetActiveDeviceIDs returns the filtered set of devices actively in
+	// use by this runner for running models.  If the runner is a bootstrap runner, no devices
+	// will be active yet so no device IDs are returned.
+	// This routine will not query the underlying device and will return immediately
+	GetActiveDeviceIDs() []ml.DeviceID
+}
--- a/docs/api.md
+++ b/docs/api.md
@@ -1708,6 +1708,7 @@ Advanced parameters:
 - `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
+- `dimensions`: number of dimensions for the embedding

 ### Examples

--- a/docs/cloud.md
+++ b/docs/cloud.md
@@ -0,0 +1,40 @@
+# Cloud
+
+| Ollama's cloud is currently in preview. For full documentation, see [Ollama's documentation](https://docs.ollama.com/cloud).
+
+## Cloud Models
+
+[Cloud models](https://ollama.com/cloud) are a new kind of model in Ollama that can run without a powerful GPU. Instead, cloud models are automatically offloaded to Ollama's cloud while offering the same capabilities as local models, making it possible to keep using your local tools while running larger models that wouldn’t fit on a personal computer.
+
+Ollama currently supports the following cloud models, with more coming soon:
+
+- `gpt-oss:20b-cloud`
+- `gpt-oss:120b-cloud`
+- `deepseek-v3.1:671b-cloud`
+- `qwen3-coder:480b-cloud`
+
+### Get started
+
+To run a cloud model, open the terminal and run:
+
+```
+ollama run gpt-oss:120b-cloud
+```
+
+To run cloud models with integrations that work with Ollama, first download the cloud model:
+
+```
+ollama pull qwen3-coder:480b-cloud
+```
+
+Then sign in to Ollama:
+
+```
+ollama signin
+```
+
+Finally, access the model using the model name `qwen3-coder:480b-cloud` via Ollama's local API or tooling.
+
+## Cloud API access
+
+Cloud models can also be accessed directly on ollama.com's API. For more information, see the [docs](https://docs.ollama.com/cloud).
--- a/docs/development.md
+++ b/docs/development.md
@@ -11,6 +11,10 @@ Then build and run Ollama from the root directory of the repository:
 go run . serve
 ```

+> [!NOTE]
+> Ollama includes native code compiled with CGO.  From time to time these data structures can change and CGO can get out of sync resulting in unexpected crashes.  You can force a full build of the native code by running `go clean -cache` first. 
+
+
 ## macOS (Apple Silicon)

 macOS Apple Silicon supports Metal which is built-in to the Ollama binary. No additional steps are required.
--- 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/gpu.md
+++ b/docs/gpu.md
@@ -51,20 +51,23 @@ sudo modprobe nvidia_uvm`
 Ollama supports the following AMD GPUs:

 ### Linux Support
-| Family         | Cards and accelerators                                                                                                               |
-| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
-| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56`    |
-| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` |
-| AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50`                                                               |
+| Family         | Cards and accelerators                                                                                                                   |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
+| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64`            |
+| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `SSG` |
+| AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60`                                                                |

 ### Windows Support
-With ROCm v6.1, the following GPUs are supported on Windows.
+With ROCm v6.2, the following GPUs are supported on Windows.

 | Family         | Cards and accelerators                                                                                                               |
 | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
 | AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800`    |
 | AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` |

+### Known Workarounds
+
+- The RX Vega 56 requires `HSA_ENABLE_SDMA=0` to disable SDMA

 ### Overrides on Linux
 Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
@@ -85,8 +88,6 @@ At this time, the known supported GPU types on linux are the following LLVM Targ
 This table shows some example GPUs that map to these LLVM targets:
 | **LLVM Target** | **An Example GPU** |
 |-----------------|---------------------|
-| gfx900 | Radeon RX Vega 56 |
-| gfx906 | Radeon Instinct MI50 |
 | gfx908 | Radeon Instinct MI100 |
 | gfx90a | Radeon Instinct MI210 |
 | gfx940 | Radeon Instinct MI300 |
--- a/docs/linux.md
+++ b/docs/linux.md
@@ -11,12 +11,13 @@ curl -fsSL https://ollama.com/install.sh | sh
 ## Manual install

 > [!NOTE]
-> If you are upgrading from a prior version, you should remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.
+> If you are upgrading from a prior version, you **MUST** remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.

 Download and extract the package:

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

--- a/docs/macos.md
+++ b/docs/macos.md
@@ -2,7 +2,7 @@

 ## System Requirements

-* MacOS Monterey (v12) or newer
+* MacOS Sonoma (v14) or newer
 * Apple M series (CPU and GPU support) or x86 (CPU only)


--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@@ -38,26 +38,14 @@ 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.
-
-In the server log, you will see a message that looks something like this (varies from release to release):
-
-```
-Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v12 rocm_v5]
-```
+Ollama includes multiple LLM libraries compiled for different GPU libraries and versions. 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.

 **Experimental LLM Library Override**

-You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to bypass autodetection, so for example, if you have a CUDA card, but want to force the CPU LLM library with AVX2 vector support, use:
+You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to limit autodetection, so for example, if you have both CUDA and AMD GPUs, but want to force the CUDA v13 only, use:

 ```shell
-OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
-```
-
-You can see what features your CPU has with the following.
-
-```shell
-cat /proc/cpuinfo| grep flags | head -1
+OLLAMA_LLM_LIBRARY="cuda_v13" ollama serve
 ```

 ## Installing older or pre-release versions on Linux
@@ -92,6 +80,9 @@ If none of those resolve the problem, gather additional information and file an
 - Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`

+You may get more details for initialization failures by enabling debug prints in the uvm driver.  You should only use this temporarily while troubleshooting
+- `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm uvm_debug_prints=1`
+

 ## AMD GPU Discovery

--- 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/envconfig/config.go
+++ b/envconfig/config.go
@@ -134,8 +134,19 @@ func LoadTimeout() (loadTimeout time.Duration) {
 	return loadTimeout
 }

-func Bool(k string) func() bool {
-	return func() bool {
+func Remotes() []string {
+	var r []string
+	raw := strings.TrimSpace(Var("OLLAMA_REMOTES"))
+	if raw == "" {
+		r = []string{"ollama.com"}
+	} else {
+		r = strings.Split(raw, ",")
+	}
+	return r
+}
+
+func BoolWithDefault(k string) func(defaultValue bool) bool {
+	return func(defaultValue bool) bool {
 		if s := Var(k); s != "" {
 			b, err := strconv.ParseBool(s)
 			if err != nil {
@@ -145,7 +156,14 @@ func Bool(k string) func() bool {
 			return b
 		}

-		return false
+		return defaultValue
+	}
+}
+
+func Bool(k string) func() bool {
+	withDefault := BoolWithDefault(k)
+	return func() bool {
+		return withDefault(false)
 	}
 }

@@ -166,7 +184,7 @@ func LogLevel() slog.Level {

 var (
 	// FlashAttention enables the experimental flash attention feature.
-	FlashAttention = Bool("OLLAMA_FLASH_ATTENTION")
+	FlashAttention = BoolWithDefault("OLLAMA_FLASH_ATTENTION")
 	// KvCacheType is the quantization type for the K/V cache.
 	KvCacheType = String("OLLAMA_KV_CACHE_TYPE")
 	// NoHistory disables readline history.
@@ -185,8 +203,6 @@ var (
 	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 {
@@ -254,7 +270,7 @@ 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_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
+		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(false), "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)"},
 		"OLLAMA_HOST":              {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
@@ -272,7 +288,7 @@ 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"},
+		"OLLAMA_REMOTES":           {"OLLAMA_REMOTES", Remotes(), "Allowed hosts for remote models (default \"ollama.com\")"},

 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -7,9 +7,11 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
+	"math"
 	"slices"
 	"strings"

+	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/util/bufioutil"
 )

@@ -55,10 +57,28 @@ func (kv KV) EmbeddingLength() uint64 {
 	return uint64(kv.Uint("embedding_length"))
 }

+func (kv KV) HeadCount() []uint64 {
+	headCountDefault := uint32(1)
+	headCount := kv.UintOrArrayValueAsArray("attention.head_count", headCountDefault)
+	if len(headCount) == 1 {
+		headCountDefault = headCount[0]
+	}
+	nLayers := int(kv.BlockCount())
+	if len(headCount) > nLayers {
+		slog.Warn("got more elements of attention.head_count than layers", "len(headCount)", len(headCount), "layers", nLayers)
+	}
+	out := make([]uint64, nLayers)
+	for i := range nLayers {
+		if i >= len(headCount) {
+			out[i] = uint64(headCountDefault)
+		} else {
+			out[i] = uint64(headCount[i])
+		}
+	}
+	return out
+}
+
 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))
 }

@@ -66,6 +86,27 @@ func (kv KV) HeadCountMin() uint64 {
 	return uint64(kv.UintOrMinArrayValue("attention.head_count", 1))
 }

+func (kv KV) HeadCountKV() []uint64 {
+	headCountKVDefault := uint32(1)
+	headCountKV := kv.UintOrArrayValueAsArray("attention.head_count_kv", headCountKVDefault)
+	if len(headCountKV) == 1 {
+		headCountKVDefault = headCountKV[0]
+	}
+	nLayers := int(kv.BlockCount())
+	if len(headCountKV) > nLayers {
+		slog.Warn("got more elements of attention.head_count than layers", "len(headCountKV)", len(headCountKV), "layers", nLayers)
+	}
+	out := make([]uint64, nLayers)
+	for i := range nLayers {
+		if i >= len(headCountKV) {
+			out[i] = uint64(headCountKVDefault)
+		} else {
+			out[i] = uint64(headCountKV[i])
+		}
+	}
+	return out
+}
+
 func (kv KV) HeadCountKVMax() uint64 {
 	return uint64(kv.UintOrMaxArrayValue("attention.head_count_kv", 1))
 }
@@ -98,6 +139,26 @@ func (kv KV) ChatTemplate() string {
 	return kv.String("tokenizer.chat_template")
 }

+// ssm architecture parameters
+
+func (kv KV) SSMConvKernel() uint64 {
+	return uint64(kv.Uint("ssm.conv_kernel"))
+}
+
+func (kv KV) SSMInnerSize() uint64 {
+	return uint64(kv.Uint("ssm.inner_size"))
+}
+
+func (kv KV) SSMStateSize() uint64 {
+	return uint64(kv.Uint("ssm.state_size"))
+}
+
+func (kv KV) SSMGroupCount() uint64 {
+	return uint64(kv.Uint("ssm.group_count"))
+}
+
+// general types
+
 func (kv KV) String(key string, defaultValue ...string) string {
 	val, _ := keyValue(kv, key, append(defaultValue, "")...)
 	return val
@@ -129,22 +190,27 @@ func (kv KV) UintOrMinArrayValue(key string, defaultValue uint32) uint32 {
 }

 func (kv KV) UintOrArrayValue(key string, defaultValue uint32) (uint32, uint32) {
+	arrVal := kv.UintOrArrayValueAsArray(key, defaultValue)
+	return slices.Min(arrVal), slices.Max(arrVal)
+}
+
+func (kv KV) UintOrArrayValueAsArray(key string, defaultValue uint32) []uint32 {
 	if u32, ok := keyValue(kv, key, uint32(0)); ok {
-		return u32, u32
+		return []uint32{u32}
 	} else if u32s, ok := keyValue(kv, key, &array[uint32]{}); ok {
-		min := slices.Min(u32s.values)
-		max := slices.Max(u32s.values)
-		return min, max
+		return u32s.values
 	} 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)
+		dst := make([]uint32, len(i32s.values))
+		for i, v := range i32s.values {
+			if v < 0 {
+				slog.Warn("array values are unexpectedly negative", "key", key, "i", i, "v", v)
+			}
+			dst[i] = uint32(v)
 		}
-		return uint32(min), uint32(max)
+		return dst
 	}

-	return defaultValue, defaultValue
+	return []uint32{defaultValue}
 }

 func (kv KV) Strings(key string, defaultValue ...[]string) []string {
@@ -177,6 +243,8 @@ func (kv KV) OllamaEngineRequired() bool {
 		"gemma3",
 		"gemma3n",
 		"mistral3",
+		"qwen3",
+		"qwen3moe",
 		"llama4",
 		"mllama",
 		"qwen25vl",
@@ -275,7 +343,7 @@ type Tensor struct {

 func (t Tensor) block() (n int) {
 	if _, err := fmt.Sscanf(t.Name, "blk.%d.", &n); err != nil {
-		return -1
+		return math.MaxInt
 	}

 	return
@@ -288,24 +356,24 @@ func (t Tensor) blockSize() uint64 {
 func (t TensorType) BlockSize() uint64 {
 	switch t {
 	case
-		0,  // F32
-		1,  // F16
-		24, // I8
-		25, // I16
-		26, // I32
-		27, // I64
-		28, // F64
-		30: // BF16
+		TensorTypeF32,
+		TensorTypeF16,
+		TensorTypeI8,
+		TensorTypeI16,
+		TensorTypeI32,
+		TensorTypeI64,
+		TensorTypeF64,
+		TensorTypeBF16:
 		return 1
 	case
-		2,  // Q4_0
-		3,  // Q4_1
-		4,  // MXFP4
-		6,  // Q5_0
-		7,  // Q5_1
-		8,  // Q8_0
-		9,  // Q8_1
-		20: // IQ4_NL
+		TensorTypeQ4_0,
+		TensorTypeQ4_1,
+		TensorTypeQ5_0,
+		TensorTypeQ5_1,
+		TensorTypeQ8_0,
+		TensorTypeQ8_1,
+		tensorTypeIQ4_NL,
+		4, TensorTypeMXFP4:
 		return 32
 	default:
 		return 256
@@ -328,8 +396,6 @@ 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:
@@ -380,6 +446,8 @@ func (t TensorType) TypeSize() uint64 {
 		return blockSize/8 + blockSize/16 + blockSize/32
 	case TensorTypeBF16:
 		return 2
+	case 4, TensorTypeMXFP4:
+		return 1 + blockSize/2
 	default:
 		return 0
 	}
@@ -479,12 +547,14 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	}, nil
 }

-func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention bool) (kv []uint64, partialOffload, fullOffload uint64) {
 	context *= uint64(numParallel)

 	embedding := f.KV().EmbeddingLength()
 	heads := f.KV().HeadCountMax()
+	headsArr := f.KV().HeadCount()
 	headsKV := f.KV().HeadCountKVMax()
+	headsKVArr := f.KV().HeadCountKV()
 	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array[string]).size)

 	embeddingHeads := f.KV().EmbeddingHeadCountMax()
@@ -494,12 +564,51 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 	layers := f.Tensors().GroupLayers()

 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
+
+	// Default for models unless special-cased below. These defaults mirror the
+	// cache usage in llama.cpp under the assumption that models without special
+	// cases below will use the llamarunner and caching will be handled by the
+	// llama.cpp layer.
+	//
+	// This also assumes that a layer without heads or headsKV set is recurrent
+	// which is usually the case. Some models (eg nemotronh) use "blocks" in
+	// place of layers where some are MLP blocks that don't have any cache.
+	// Models like this will need a special case below to be accurately
+	// estimated.
 	var kvTotal uint64
 	kv = make([]uint64, f.KV().BlockCount())
+	kvSizeAttn := uint64(0)
+	kvSizeRecurrent := uint64(0)
 	for i := range kv {
-		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+		headsL := headsArr[i]
+		headsKVL := headsKVArr[i]
+		if headsL > 0 && headsKVL > 0 {
+			// full attention layer
+			// NOTE: Assumes uniform values for all attn layers
+			kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKVL) * bytesPerElement)
+			kvSizeAttn += kv[i]
+		} else {
+			// recurrent layer
+			ssmDConv := f.KV().SSMConvKernel()
+			ssmDState := f.KV().SSMStateSize()
+			ssmDInner := f.KV().SSMInnerSize()
+			ssmNGroups := f.KV().SSMGroupCount()
+			nEmbdR := uint64(0)
+			if ssmDConv > 0 {
+				nEmbdR = (ssmDConv - 1) * (ssmDInner + 2*ssmNGroups*ssmDState)
+			}
+			nEmbdS := ssmDState * ssmDInner
+
+			// recurrent always uses F32 in llama.cpp backend
+			// https://github.com/ggml-org/llama.cpp/blob/master/src/llama-model.cpp#L18644
+			bytesPerElementRecurrent := kvCacheBytesPerElement("f32")
+
+			kv[i] = (nEmbdR + nEmbdS) * uint64(bytesPerElementRecurrent)
+			kvSizeRecurrent += kv[i]
+		}
 		kvTotal += kv[i]
 	}
+	slog.Debug("default cache size estimate", "attention MiB", float32(kvSizeAttn)/(1024.*1024.), "attention bytes", kvSizeAttn, "recurrent MiB", float32(kvSizeRecurrent)/(1024.*1024.), "recurrent bytes", kvSizeRecurrent)

 	switch f.KV().Architecture() {
 	case "llama", "llama4":
@@ -677,7 +786,12 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 				kv[i] *= context
 			}
 		}
+
 		partialOffload = 2 * f.KV().HeadCountMax() / cmp.Or(f.KV().HeadCountKVMin(), 1) * kvTotal / 6
+		if useFlashAttention {
+			// rough estimate of graph size with flash attention on
+			partialOffload = (4*uint64(numParallel) + context>>10 + 110) * format.MebiByte
+		}
 	}

 	return
@@ -752,12 +866,11 @@ 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"
+	if cacheType == "" || cacheType == "f16" {
+		return true
 	}
-	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
+
+	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }

 // SupportsFlashAttention checks if the model supports flash attention
@@ -767,12 +880,25 @@ func (f GGML) SupportsFlashAttention() bool {
 		return false
 	}

+	if arch := f.KV().Architecture(); slices.Contains([]string{"gemma2"}, arch) {
+		return false
+	}
+
 	// Check head counts match and are non-zero
 	headCountK := f.KV().EmbeddingHeadCountK()
 	headCountV := f.KV().EmbeddingHeadCountV()
 	return headCountK != 0 && headCountV != 0 && headCountK == headCountV
 }

+// FlashAttention checks if the model should enable flash attention
+func (f GGML) FlashAttention() bool {
+	return slices.Contains([]string{
+		"gptoss", "gpt-oss",
+		"qwen3",
+		"qwen3moe",
+	}, f.KV().String("general.architecture"))
+}
+
 // kvCacheBytesPerElement returns the number of bytes per element for a given KV cache type
 func kvCacheBytesPerElement(cacheType string) float64 {
 	switch cacheType {
@@ -780,6 +906,8 @@ func kvCacheBytesPerElement(cacheType string) float64 {
 		return 1 // 1/2 of fp16
 	case "q4_0":
 		return 0.5 // 1/4 of fp16
+	case "f32":
+		return 4 // f32 (default for recurrent)
 	default:
 		return 2 // f16 (default)
 	}
--- a/fs/ggml/gguf.go
+++ b/fs/ggml/gguf.go
@@ -533,12 +533,15 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		}
 	}

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

 	var s uint64
 	for i := range ts {
--- a/fs/ggml/gguf_test.go
+++ b/fs/ggml/gguf_test.go
@@ -11,24 +11,24 @@ import (
 )

 func TestWriteGGUF(t *testing.T) {
-	r := rand.New(rand.NewPCG(0, 0))
+	b := bytes.NewBuffer(make([]byte, 2*3))
 	for range 8 {
 		t.Run("shuffle", func(t *testing.T) {
 			t.Parallel()

 			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))},
+				{Name: "token_embd.weight", Shape: []uint64{2, 3}, WriterTo: b},
+				{Name: "blk.0.ffn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
+				{Name: "blk.0.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
+				{Name: "blk.1.ffn_up.weight", Shape: []uint64{2, 3}, WriterTo: b},
+				{Name: "blk.2.ffn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
+				{Name: "blk.1.ffn_down.weight", Shape: []uint64{2, 3}, WriterTo: b},
+				{Name: "blk.0.attn_k.weight", Shape: []uint64{2, 3}, WriterTo: b},
+				{Name: "output_norm.weight", Shape: []uint64{3, 2}, WriterTo: b},
+				{Name: "output.weight", Shape: []uint64{3, 2}, WriterTo: b},
 			}

-			r.Shuffle(len(ts), func(i, j int) {
+			rand.Shuffle(len(ts), func(i, j int) {
 				ts[i], ts[j] = ts[j], ts[i]
 			})

@@ -63,14 +63,14 @@ func TestWriteGGUF(t *testing.T) {
 			}

 			if diff := cmp.Diff(Tensors{
-				Offset: 608,
+				Offset: 592,
 				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: "blk.0.attn_k.weight", Offset: 0, Shape: []uint64{2, 3}},
+					{Name: "blk.0.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},
+					{Name: "blk.0.ffn_norm.weight", Offset: 64, Shape: []uint64{2, 3}},
+					{Name: "blk.1.ffn_down.weight", Offset: 96, Shape: []uint64{2, 3}},
+					{Name: "blk.1.ffn_up.weight", Offset: 128, Shape: []uint64{2, 3}},
+					{Name: "blk.2.ffn_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}},
--- a/fs/ggml/type.go
+++ b/fs/ggml/type.go
@@ -146,8 +146,6 @@ func (ftype FileType) ToTensorType() TensorType {
 		return TensorTypeQ4_0
 	case fileTypeQ4_1:
 		return TensorTypeQ4_1
-	case fileTypeMXFP4:
-		return TensorTypeMXFP4 // Formerly unused tensorTypeQ4_2
 	case FileTypeQ8_0:
 		return TensorTypeQ8_0
 	case fileTypeQ5_0:
@@ -176,6 +174,8 @@ func (ftype FileType) ToTensorType() TensorType {
 		return TensorTypeQ2_K
 	case FileTypeBF16:
 		return TensorTypeBF16
+	case fileTypeMXFP4:
+		return TensorTypeMXFP4
 	default:
 		slog.Warn("unsupported file type", "type", ftype)
 		return 0 // F32
@@ -191,8 +191,8 @@ const (
 	TensorTypeF16
 	TensorTypeQ4_0
 	TensorTypeQ4_1
-	TensorTypeMXFP4 // Formerly unused tensorTypeQ4_2
-	tensorTypeQ4_3  // unused by GGML
+	tensorTypeQ4_2
+	tensorTypeQ4_3 // unused by GGML
 	TensorTypeQ5_0
 	TensorTypeQ5_1
 	TensorTypeQ8_0
@@ -226,6 +226,7 @@ const (
 	tensorTypeIQ4_NL_4_4 // unused by GGML
 	tensorTypeIQ4_NL_4_8 // unused by GGML
 	tensorTypeIQ4_NL_8_8 // unused by GGML
+	TensorTypeMXFP4
 )

 // ParseFileType parses the provided GGUF file type
@@ -318,7 +319,7 @@ func (t TensorType) String() string {
 		return "F64"
 	case TensorTypeBF16:
 		return "BF16"
-	case TensorTypeMXFP4:
+	case 4, TensorTypeMXFP4:
 		return "MXFP4"
 	default:
 		return "unknown"
--- a/harmony/harmonyparser.go
+++ b/harmony/harmonyparser.go
@@ -1,9 +1,9 @@
-package server
+package harmony

 import (
-	"context"
+	"encoding/json"
+	"fmt"
 	"log/slog"
-	"slices"
 	"strings"
 	"unicode"

@@ -19,18 +19,6 @@ const (
 	harmonyParserState_ParsingContent
 )

-func shouldUseHarmony(model Model) bool {
-	if slices.Contains([]string{"gptoss", "gpt-oss"}, model.Config.ModelFamily) {
-		// heuristic to check whether the template expects to be parsed via harmony:
-		// search for harmony tags that are nearly always used
-		if model.Template.Contains("<|start|>") && model.Template.Contains("<|end|>") {
-			return true
-		}
-	}
-
-	return false
-}
-
 func (s harmonyParserState) String() string {
 	switch s {
 	// we're looking for the message start tag
@@ -275,19 +263,24 @@ const (
 // HarmonyMessageHandler processes harmony events and accumulates content appropriately.
 // This is a higher level interface that maps harmony concepts into ollama concepts
 type HarmonyMessageHandler struct {
-	state         harmonyMessageState
-	harmonyParser *HarmonyParser
+	state           harmonyMessageState
+	HarmonyParser   *HarmonyParser
+	FunctionNameMap *FunctionNameMap
+	toolAccumulator *HarmonyToolCallAccumulator
+	convertedTools  map[string]struct{}
 }

 // NewHarmonyMessageHandler creates a new message handler
 func NewHarmonyMessageHandler() *HarmonyMessageHandler {
 	return &HarmonyMessageHandler{
 		state: harmonyMessageState_Normal,
-		harmonyParser: &HarmonyParser{
+		HarmonyParser: &HarmonyParser{
 			MessageStartTag: "<|start|>",
 			MessageEndTag:   "<|end|>",
 			HeaderEndTag:    "<|message|>",
 		},
+		FunctionNameMap: NewFunctionNameMap(),
+		convertedTools:  make(map[string]struct{}),
 	}
 }

@@ -298,11 +291,11 @@ func (h *HarmonyMessageHandler) AddContent(content string, toolParser *HarmonyTo
 	thinkingSb := strings.Builder{}
 	toolContentSb := strings.Builder{}

-	events := h.harmonyParser.AddContent(content)
+	events := h.HarmonyParser.AddContent(content)
 	for _, event := range events {
 		switch event := event.(type) {
 		case HarmonyEventHeaderComplete:
-			slog.Log(context.TODO(), logutil.LevelTrace, "harmony event header complete", "header", event.Header)
+			logutil.Trace("harmony event header complete", "header", event.Header)
 			switch event.Header.Channel {
 			case "analysis":
 				if event.Header.Recipient != "" {
@@ -325,7 +318,7 @@ func (h *HarmonyMessageHandler) AddContent(content string, toolParser *HarmonyTo
 				h.state = harmonyMessageState_Normal
 			}
 		case HarmonyEventContentEmitted:
-			slog.Log(context.TODO(), logutil.LevelTrace, "harmony event content", "content", event.Content, "state", h.state)
+			logutil.Trace("harmony event content", "content", event.Content, "state", h.state)
 			if h.state == harmonyMessageState_Normal {
 				contentSb.WriteString(event.Content)
 			} else if h.state == harmonyMessageState_Thinking {
@@ -378,3 +371,174 @@ func (a *HarmonyToolCallAccumulator) Drain() (*string, string) {
 func (a *HarmonyToolCallAccumulator) Content() string {
 	return a.acc.String()
 }
+
+// FunctionNameMap maps a user-specified function name to a valid function
+// name for harmony (which look like TypeScript identifiers). This is needed to
+// transform user-specified function names, which might contain characters that
+// are not allowed in TypeScript identifiers
+type FunctionNameMap struct {
+	userToHarmony map[string]string
+	harmonyToUser map[string]string
+}
+
+func NewFunctionNameMap() *FunctionNameMap {
+	return &FunctionNameMap{
+		userToHarmony: make(map[string]string),
+		harmonyToUser: make(map[string]string),
+	}
+}
+
+// Init initializes the handler with tools and optional last message
+// Implements the Parser interface
+func (h *HarmonyMessageHandler) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
+	// Initialize the harmony parser
+	if h.HarmonyParser == nil {
+		h.HarmonyParser = &HarmonyParser{
+			MessageStartTag: "<|start|>",
+			MessageEndTag:   "<|end|>",
+			HeaderEndTag:    "<|message|>",
+		}
+	}
+
+	// Handle prefill for chat mode
+	if lastMessage != nil {
+		h.HarmonyParser.AddImplicitStartOrPrefill(lastMessage)
+	} else {
+		h.HarmonyParser.AddImplicitStart()
+	}
+
+	// Initialize tool accumulator
+	h.toolAccumulator = h.CreateToolParser()
+
+	// Process tools and return renamed versions
+	if len(tools) == 0 {
+		return tools
+	}
+
+	processedTools := make([]api.Tool, len(tools))
+	copy(processedTools, tools)
+	for i, tool := range processedTools {
+		if tool.Function.Name != "" {
+			processedTools[i].Function.Name = h.FunctionNameMap.ConvertAndAdd(tool.Function.Name)
+			h.convertedTools[tool.Function.Name] = struct{}{}
+		}
+	}
+	return processedTools
+}
+
+// Add implements the Parser interface - processes streamed content and extracts content, thinking, and tool calls
+func (h *HarmonyMessageHandler) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	content, thinking, toolContent := h.AddContent(s, h.toolAccumulator)
+	if toolContent != "" {
+		h.toolAccumulator.Add(toolContent)
+	}
+
+	// tool calls always happen one at a time, and always at the end of a message,
+	// so for simplicity we defer parsing them until we know we're done
+	if done {
+		toolName, raw := h.toolAccumulator.Drain()
+		if toolName != nil {
+			name := strings.TrimPrefix(*toolName, "functions.")
+			name = h.FunctionNameMap.OriginalFromConverted(name)
+			var args api.ToolCallFunctionArguments
+			if err := json.Unmarshal([]byte(raw), &args); err != nil {
+				return "", "", nil, fmt.Errorf("error parsing tool call: raw='%s', err=%w", raw, err)
+			}
+			calls = append(calls, api.ToolCall{Function: api.ToolCallFunction{Name: name, Arguments: args}})
+		}
+	}
+
+	return content, thinking, calls, nil
+}
+
+// HasToolSupport implements the Parser interface
+func (h *HarmonyMessageHandler) HasToolSupport() bool {
+	return true
+}
+
+// HasThinkingSupport implements the Parser interface
+func (h *HarmonyMessageHandler) HasThinkingSupport() bool {
+	return true
+}
+
+func (m *FunctionNameMap) ConvertAndAdd(userFunctionName string) string {
+	harmonyFunctionName := m.deriveName(userFunctionName)
+	// built-in functions should not be renamed
+	if userFunctionName == "browser.open" || userFunctionName == "browser.search" || userFunctionName == "browser.find" || userFunctionName == "python" {
+		harmonyFunctionName = userFunctionName
+	}
+	m.userToHarmony[userFunctionName] = harmonyFunctionName
+	m.harmonyToUser[harmonyFunctionName] = userFunctionName
+	return harmonyFunctionName
+}
+
+// OriginalFromConverted looks up the reverse-mapping of a previously-converted
+// user->harmony function name. To unmap reliably, the mapping must exist, as
+// the conversion process is not reversible without the appropriate state
+func (m *FunctionNameMap) OriginalFromConverted(harmonyFunctionName string) string {
+	if userFunctionName, ok := m.harmonyToUser[harmonyFunctionName]; ok {
+		return userFunctionName
+	}
+	slog.Warn("harmony parser: no reverse mapping found for function name", "harmonyFunctionName", harmonyFunctionName)
+	// fallback to the original function name if we can't find a mapping
+	return harmonyFunctionName
+}
+
+// convertToValidChars converts a user-specified function name to a valid
+// TypeScript identifier.
+//
+// Limitations:
+//
+//   - This doesn't restrict reserved TypeScript keywords.
+//   - We don't perform a real ID_Start/ID_Continue check, and instead use the more
+//     restrictive unicode.IsLetter/unicode.IsDigit check. Unclear what kind of
+//     identifiers these models were trained on, so in the end we might want to
+//     convert unicode-heavy identifiers to their closest ASCII equivalents.
+func (m *FunctionNameMap) convertToValidChars(userFunctionName string) string {
+	mapper := func(r rune) rune {
+		// first, replace certain characters with underscores
+		if r == ' ' || r == '-' || r == '.' {
+			return '_'
+		}
+
+		if unicode.IsLetter(r) || unicode.IsDigit(r) || r == '_' || r == '$' {
+			return r
+		}
+
+		// finally, remove any other characters
+		return -1
+	}
+	candidate := strings.Map(mapper, userFunctionName)
+
+	// set a default name if we end up with nothing left
+	if candidate == "" {
+		return "unnamed"
+	}
+
+	// if the candidate starts with a number, prepend an underscore to make it a
+	// valid identifier
+	if unicode.IsDigit(rune(candidate[0])) {
+		candidate = "_" + candidate
+	}
+
+	return candidate
+}
+
+func (m *FunctionNameMap) deriveName(userFunctionName string) string {
+	originalCandidate := m.convertToValidChars(userFunctionName)
+	candidate := originalCandidate
+
+	// Check for dupes, and if so, add a number to the end.
+	// We start at 2 because if we have dupes and the first is never renamed, it
+	// makes sense for them to be named, say, `f`, `f_2`, `f_3`
+	count := 2
+	for {
+		if _, exists := m.harmonyToUser[candidate]; !exists {
+			break
+		}
+		candidate = fmt.Sprintf("%s_%d", originalCandidate, count)
+		count++
+	}
+
+	return candidate
+}
--- a/harmony/harmonyparser_test.go
+++ b/harmony/harmonyparser_test.go
@@ -1,4 +1,4 @@
-package server
+package harmony

 import (
 	"fmt"
@@ -467,3 +467,72 @@ func TestHarmonyParserStreaming(t *testing.T) {
 		})
 	}
 }
+
+// TestFunctionConvertToValidChars tests only FunctionNameMap.convert(), which doesn't
+// handle any saving (and therefore no dupe handling)
+func TestFunctionConvertToValidChars(t *testing.T) {
+	tests := []struct {
+		name string
+		in   string
+		want string
+	}{
+		{name: "replace spaces with underscores", in: "get weather", want: "get_weather"},
+		{name: "replace hyphens with underscores", in: "get-weather", want: "get_weather"},
+		{name: "replace periods with underscores", in: "get.weather", want: "get_weather"},
+		{name: "disallow non-word characters", in: "get weather!", want: "get_weather"},
+		{name: "strip out invalid non-alphanumeric unicode characters", in: "a🫠bc", want: "abc"},
+		{name: "names that only contain invalid characters", in: "🫠", want: "unnamed"},
+		{name: "leading number", in: "123", want: "_123"},
+		{name: "$ allowed", in: "$", want: "$"},
+		// show that we allow weird unicode letter characters, though we might want
+		// to convert them to their closest ASCII equivalents in the future
+		{name: "allow weird unicode letter characters", in: "𝓸𝓵𝓵𝓪𝓶𝓪", want: "𝓸𝓵𝓵𝓪𝓶𝓪"},
+		// names that look like words but are invalid (i.e., not ID_Start/ID_Continue)
+		{name: "disallow non-word characters that look like words", in: "ⓞⓛⓛⓐⓜⓐ123", want: "_123"},
+	}
+
+	for i, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			parser := NewFunctionNameMap()
+			got := parser.convertToValidChars(tt.in)
+			if got != tt.want {
+				t.Errorf("case %d: got %q, want %q", i, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestFunctionConvertAndAdd(t *testing.T) {
+	// make a fresh map for each test, but within a test use the same map so we can test for dupe handling
+	tests := []struct {
+		name string
+		in   []string
+		want []string
+	}{
+		{name: "basic dupe handling", in: []string{"get weather", "get weather"}, want: []string{"get_weather", "get_weather_2"}},
+		{name: "dupes from different user-specified names", in: []string{"get weather", "get_weather", "get-weather"}, want: []string{"get_weather", "get_weather_2", "get_weather_3"}},
+		{name: "non dupes after dupes", in: []string{"get weather", "get_weather", "get-weather", "something-different"}, want: []string{"get_weather", "get_weather_2", "get_weather_3", "something_different"}},
+		{name: "multiple sets of dupes", in: []string{"a", "a", "b", "a", "a", "b", "a"}, want: []string{"a", "a_2", "b", "a_3", "a_4", "b_2", "a_5"}},
+		{name: "built-in functions should not be renamed", in: []string{"browser.open", "python", "not.a.built-in.function", "browser.not_a_real_built_in"}, want: []string{"browser.open", "python", "not_a_built_in_function", "browser_not_a_real_built_in"}},
+	}
+
+	for i, tt := range tests {
+		parser := NewFunctionNameMap()
+		t.Run(tt.name, func(t *testing.T) {
+			for j, in := range tt.in {
+				got := parser.ConvertAndAdd(in)
+				want := tt.want[j]
+				if got != want {
+					t.Errorf("case %d: got %q, want %q", i, got, want)
+				}
+				// check that the maps are correct
+				if parser.userToHarmony[in] != want {
+					t.Errorf("case %d: userToHarmony[%q] = %q, want %q", i, in, parser.userToHarmony[in], want)
+				}
+				if parser.harmonyToUser[want] != in {
+					t.Errorf("case %d: harmonyToUser[%q] = %q, want %q", i, want, parser.harmonyToUser[want], in)
+				}
+			}
+		})
+	}
+}
--- a/integration/README.md
+++ b/integration/README.md
@@ -2,10 +2,16 @@

 This directory contains integration tests to exercise Ollama end-to-end to verify behavior

-By default, these tests are disabled so `go test ./...` will exercise only unit tests.  To run integration tests you must pass the integration tag.  `go test -tags=integration ./...`
+By default, these tests are disabled so `go test ./...` will exercise only unit tests.  To run integration tests you must pass the integration tag.  `go test -tags=integration ./...` Some tests require additional tags to enable to allow scoped testing to keep the duration reasonable.  For example, testing a broad set of models requires `-tags=integration,models` and a longer timeout (~60m or more depending on the speed of your GPU.). To view the current set of tag combinations use `find integration -type f | xargs grep "go:build"`


 The integration tests have 2 modes of operating.

 1. By default, they will start the server on a random port, run the tests, and then shutdown the server.
-2. If `OLLAMA_TEST_EXISTING` is set to a non-empty string, the tests will run against an existing running server, which can be remote
+2. If `OLLAMA_TEST_EXISTING` is set to a non-empty string, the tests will run against an existing running server, which can be remote based on your `OLLAMA_HOST` environment variable
+
+> [!IMPORTANT]
+> Before running the tests locally without the "test existing" setting, compile ollama from the top of the source tree  `go build .` in addition to GPU support with cmake if applicable on your platform.  The integration tests expect to find an ollama binary at the top of the tree.
+
+
+Many tests use a default small model suitable to run on many systems.  You can override this default model by setting `OLLAMA_TEST_DEFAULT_MODEL`
--- a/integration/api_test.go
+++ b/integration/api_test.go
@@ -22,13 +22,12 @@ func TestAPIGenerate(t *testing.T) {
 	// Set up the test data
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: "why is the sky blue? be brief",
+		Prompt: blueSkyPrompt,
 		Options: map[string]interface{}{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
-	anyResp := []string{"rayleigh", "scattering"}

 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
@@ -120,14 +119,14 @@ func TestAPIGenerate(t *testing.T) {
 				// Verify the response contains the expected data
 				response := buf.String()
 				atLeastOne := false
-				for _, resp := range anyResp {
+				for _, resp := range blueSkyExpected {
 					if strings.Contains(strings.ToLower(response), resp) {
 						atLeastOne = true
 						break
 					}
 				}
 				if !atLeastOne {
-					t.Errorf("none of %v found in %s", anyResp, response)
+					t.Errorf("none of %v found in %s", blueSkyExpected, response)
 				}
 			case <-ctx.Done():
 				t.Error("outer test context done while waiting for generate")
@@ -181,7 +180,7 @@ func TestAPIChat(t *testing.T) {
 		Messages: []api.Message{
 			{
 				Role:    "user",
-				Content: "why is the sky blue?  be brief",
+				Content: blueSkyPrompt,
 			},
 		},
 		Options: map[string]interface{}{
@@ -189,7 +188,6 @@ func TestAPIChat(t *testing.T) {
 			"seed":        123,
 		},
 	}
-	anyResp := []string{"rayleigh", "scattering"}

 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
@@ -279,14 +277,14 @@ func TestAPIChat(t *testing.T) {
 				// Verify the response contains the expected data
 				response := buf.String()
 				atLeastOne := false
-				for _, resp := range anyResp {
+				for _, resp := range blueSkyExpected {
 					if strings.Contains(strings.ToLower(response), resp) {
 						atLeastOne = true
 						break
 					}
 				}
 				if !atLeastOne {
-					t.Errorf("none of %v found in %s", anyResp, response)
+					t.Errorf("none of %v found in %s", blueSkyExpected, response)
 				}
 			case <-ctx.Done():
 				t.Error("outer test context done while waiting for chat")
@@ -390,7 +388,7 @@ func TestAPIEmbeddings(t *testing.T) {
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
 	req := api.EmbeddingRequest{
-		Model:  "orca-mini",
+		Model:  libraryEmbedModels[0],
 		Prompt: "why is the sky blue?",
 		Options: map[string]interface{}{
 			"temperature": 0,
@@ -410,3 +408,99 @@ func TestAPIEmbeddings(t *testing.T) {
 		t.Errorf("zero length embedding response")
 	}
 }
+
+func TestAPIToolCalling(t *testing.T) {
+	initialTimeout := 60 * time.Second
+	streamTimeout := 30 * time.Second
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	modelName := "qwen3:0.6b"
+	if err := PullIfMissing(ctx, client, modelName); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	tools := []api.Tool{
+		{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name:        "get_weather",
+				Description: "Get the current weather in a given location",
+				Parameters: api.ToolFunctionParameters{
+					Type:     "object",
+					Required: []string{"location"},
+					Properties: map[string]api.ToolProperty{
+						"location": {
+							Type:        api.PropertyType{"string"},
+							Description: "The city and state, e.g. San Francisco, CA",
+						},
+					},
+				},
+			},
+		},
+	}
+
+	req := api.ChatRequest{
+		Model: modelName,
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: "Call get_weather with location set to San Francisco.",
+			},
+		},
+		Tools: tools,
+		Options: map[string]any{
+			"temperature": 0,
+		},
+	}
+
+	stallTimer := time.NewTimer(initialTimeout)
+	var gotToolCall bool
+	var lastToolCall api.ToolCall
+
+	fn := func(response api.ChatResponse) error {
+		if len(response.Message.ToolCalls) > 0 {
+			gotToolCall = true
+			lastToolCall = response.Message.ToolCalls[len(response.Message.ToolCalls)-1]
+		}
+		if !stallTimer.Reset(streamTimeout) {
+			return fmt.Errorf("stall was detected while streaming response, aborting")
+		}
+		return nil
+	}
+
+	stream := true
+	req.Stream = &stream
+	done := make(chan int)
+	var genErr error
+	go func() {
+		genErr = client.Chat(ctx, &req, fn)
+		done <- 0
+	}()
+
+	select {
+	case <-stallTimer.C:
+		t.Errorf("tool-calling chat never started. Timed out after: %s", initialTimeout.String())
+	case <-done:
+		if genErr != nil {
+			t.Fatalf("chat failed: %v", genErr)
+		}
+
+		if !gotToolCall {
+			t.Fatalf("expected at least one tool call, got none")
+		}
+
+		if lastToolCall.Function.Name != "get_weather" {
+			t.Errorf("unexpected tool called: got %q want %q", lastToolCall.Function.Name, "get_weather")
+		}
+
+		if _, ok := lastToolCall.Function.Arguments["location"]; !ok {
+			t.Errorf("expected tool arguments to include 'location', got: %s", lastToolCall.Function.Arguments.String())
+		}
+	case <-ctx.Done():
+		t.Error("outer test context done while waiting for tool-calling chat")
+	}
+}
--- a/integration/basic_test.go
+++ b/integration/basic_test.go
@@ -11,23 +11,27 @@ import (
 	"time"

 	"github.com/ollama/ollama/api"
-	"github.com/stretchr/testify/require"
 )

 func TestBlueSky(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 	// Set up the test data
-	req := api.GenerateRequest{
-		Model:  smol,
-		Prompt: "why is the sky blue?",
+	req := api.ChatRequest{
+		Model: smol,
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: blueSkyPrompt,
+			},
+		},
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
+	ChatTestHelper(ctx, t, req, blueSkyExpected)
 }

 func TestUnicode(t *testing.T) {
@@ -35,10 +39,15 @@ func TestUnicode(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
 	// Set up the test data
-	req := api.GenerateRequest{
+	req := api.ChatRequest{
 		// DeepSeek has a Unicode tokenizer regex, making it a unicode torture test
-		Model:  "deepseek-coder-v2:16b-lite-instruct-q2_K",
-		Prompt: "天空为什么是蓝色的?",
+		Model: "deepseek-coder-v2:16b-lite-instruct-q2_K", // TODO is there an ollama-engine model we can switch to and keep the coverage?
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: "天空为什么是蓝色的?", // Why is the sky blue?
+			},
+		},
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
@@ -50,17 +59,39 @@ func TestUnicode(t *testing.T) {
 	}
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, req.Model))
-	DoGenerate(ctx, t, client, req, []string{"散射", "频率"}, 120*time.Second, 120*time.Second)
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatal(err)
+	}
+	slog.Info("loading", "model", req.Model)
+	err := client.Generate(ctx, &api.GenerateRequest{Model: req.Model}, func(response api.GenerateResponse) error { return nil })
+	if err != nil {
+		t.Fatalf("failed to load model %s: %s", req.Model, err)
+	}
+	defer func() {
+		// best effort unload once we're done with the model
+		client.Generate(ctx, &api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 0}}, func(rsp api.GenerateResponse) error { return nil })
+	}()
+
+	skipIfNotGPULoaded(ctx, t, client, req.Model, 100)
+
+	DoChat(ctx, t, client, req, []string{
+		"散射", // scattering
+		"频率", // frequency
+	}, 120*time.Second, 120*time.Second)
 }

 func TestExtendedUnicodeOutput(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 	// Set up the test data
-	req := api.GenerateRequest{
-		Model:  "gemma2:2b",
-		Prompt: "Output some smily face emoji",
+	req := api.ChatRequest{
+		Model: "gemma2:2b",
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: "Output some smily face emoji",
+			},
+		},
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
@@ -69,8 +100,10 @@ func TestExtendedUnicodeOutput(t *testing.T) {
 	}
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, req.Model))
-	DoGenerate(ctx, t, client, req, []string{"😀", "😊", "😁", "😂", "😄", "😃"}, 120*time.Second, 120*time.Second)
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatal(err)
+	}
+	DoChat(ctx, t, client, req, []string{"😀", "😊", "😁", "😂", "😄", "😃"}, 120*time.Second, 120*time.Second)
 }

 func TestUnicodeModelDir(t *testing.T) {
@@ -84,7 +117,9 @@ func TestUnicodeModelDir(t *testing.T) {
 	}

 	modelDir, err := os.MkdirTemp("", "ollama_埃")
-	require.NoError(t, err)
+	if err != nil {
+		t.Fatal(err)
+	}
 	defer os.RemoveAll(modelDir)
 	slog.Info("unicode", "OLLAMA_MODELS", modelDir)

@@ -93,14 +128,19 @@ func TestUnicodeModelDir(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()

-	req := api.GenerateRequest{
-		Model:  smol,
-		Prompt: "why is the sky blue?",
+	req := api.ChatRequest{
+		Model: smol,
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: blueSkyPrompt,
+			},
+		},
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
+	ChatTestHelper(ctx, t, req, blueSkyExpected)
 }
--- a/integration/concurrency_test.go
+++ b/integration/concurrency_test.go
@@ -14,17 +14,15 @@ import (
 	"testing"
 	"time"

-	"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) {
+func TestConcurrentChat(t *testing.T) {
 	// Assumes all requests have the same model
-	req, resp := GenerateRequests()
+	req, resp := ChatRequests()
 	numParallel := int(envconfig.NumParallel() + 1)
 	iterLimit := 3

@@ -59,7 +57,7 @@ func TestConcurrentGenerate(t *testing.T) {
 				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
-				DoGenerate(ctx, t, client, req[k], resp[k], 120*time.Second, 20*time.Second)
+				DoChat(ctx, t, client, req[k], resp[k], 120*time.Second, 20*time.Second)
 			}
 		}(i)
 	}
@@ -79,21 +77,21 @@ func TestMultiModelStress(t *testing.T) {
 		t.Fatal(err)
 	}

+	// All models compatible with ollama-engine
 	smallModels := []string{
 		"llama3.2:1b",
 		"qwen3:0.6b",
-		"gemma:2b",
-		"deepseek-r1:1.5b",
-		"starcoder2:3b",
+		"gemma2:2b",
+		"deepseek-r1:1.5b", // qwen2 arch
+		"gemma3:270m",
 	}
 	mediumModels := []string{
-		"qwen3:8b",
-		"llama2",
-		"deepseek-r1:7b",
-		"mistral",
-		"dolphin-mistral",
-		"gemma:7b",
-		"codellama:7b",
+		"llama3.2:3b",    // ~3.4G
+		"qwen3:8b",       // ~6.6G
+		"gpt-oss:20b",    // ~15G
+		"deepseek-r1:7b", // ~5.6G
+		"gemma3:4b",      // ~5.8G
+		"gemma2:9b",      // ~8.1G
 	}

 	var chosenModels []string
@@ -114,13 +112,16 @@ func TestMultiModelStress(t *testing.T) {

 	// Make sure all the models are pulled before we get started
 	for _, model := range chosenModels {
-		require.NoError(t, PullIfMissing(ctx, client, model))
+		if err := PullIfMissing(ctx, client, model); err != nil {
+			t.Fatal(err)
+		}
 	}

 	// 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")
+chooseModels:
 	for i, model := range chosenModels {
 		req := &api.GenerateRequest{Model: model}
 		slog.Info("loading", "model", model)
@@ -142,6 +143,13 @@ func TestMultiModelStress(t *testing.T) {
 				slog.Info("found model load capacity", "target", targetLoadCount, "current", loaded, "chosen", chosenModels[:targetLoadCount])
 				break
 			}
+			// Effectively limit model count to 2 on CPU only systems to avoid thrashing and timeouts
+			for _, m := range models.Models {
+				if m.SizeVRAM == 0 {
+					slog.Info("model running on CPU", "name", m.Name, "target", targetLoadCount, "chosen", chosenModels[:targetLoadCount])
+					break chooseModels
+				}
+			}
 		}
 	}
 	if targetLoadCount == len(chosenModels) {
@@ -155,7 +163,7 @@ func TestMultiModelStress(t *testing.T) {
 		wg.Add(1)
 		go func(i int) {
 			defer wg.Done()
-			reqs, resps := GenerateRequests()
+			reqs, resps := ChatRequests()
 			for j := 0; j < 3; j++ {
 				if time.Now().Sub(started) > softTimeout {
 					slog.Info("exceeded soft timeout, winding down test")
@@ -163,8 +171,8 @@ func TestMultiModelStress(t *testing.T) {
 				}
 				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],
+				slog.Info("Starting", "model", reqs[k].Model, "iteration", j, "request", reqs[k].Messages[0].Content)
+				DoChat(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
 				)
--- a/integration/context_test.go
+++ b/integration/context_test.go
@@ -21,9 +21,14 @@ func TestLongInputContext(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 	defer cancel()
 	// Set up the test data
-	req := api.GenerateRequest{
-		Model:  "llama2",
-		Prompt: "Oh, don’t speak to me of Austria. Perhaps I don’t understand things, but Austria never has wished, and does not wish, for war. She is betraying us! Russia alone must save Europe. Our gracious sovereign recognizes his high vocation and will be true to it. That is the one thing I have faith in! Our good and wonderful sovereign has to perform the noblest role on earth, and he is so virtuous and noble that God will not forsake him. He will fulfill his vocation and crush the hydra of revolution, which has become more terrible than ever in the person of this murderer and villain! We alone must avenge the blood of the just one.... Whom, I ask you, can we rely on?... England with her commercial spirit will not and cannot understand the Emperor Alexander’s loftiness of soul. She has refused to evacuate Malta. She wanted to find, and still seeks, some secret motive in our actions. What answer did Novosíltsev get? None. The English have not understood and cannot understand the self-abnegation of our Emperor who wants nothing for himself, but only desires the good of mankind. And what have they promised? Nothing! And what little they have promised they will not perform! Prussia has always declared that Buonaparte is invincible, and that all Europe is powerless before him.... And I don’t believe a word that Hardenburg says, or Haugwitz either. This famous Prussian neutrality is just a trap. I have faith only in God and the lofty destiny of our adored monarch. He will save Europe! What country is this referring to?",
+	req := api.ChatRequest{
+		Model: smol,
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: "Oh, don’t speak to me of Austria. Perhaps I don’t understand things, but Austria never has wished, and does not wish, for war. She is betraying us! Russia alone must save Europe. Our gracious sovereign recognizes his high vocation and will be true to it. That is the one thing I have faith in! Our good and wonderful sovereign has to perform the noblest role on earth, and he is so virtuous and noble that God will not forsake him. He will fulfill his vocation and crush the hydra of revolution, which has become more terrible than ever in the person of this murderer and villain! We alone must avenge the blood of the just one.... Whom, I ask you, can we rely on?... England with her commercial spirit will not and cannot understand the Emperor Alexander’s loftiness of soul. She has refused to evacuate Malta. She wanted to find, and still seeks, some secret motive in our actions. What answer did Novosíltsev get? None. The English have not understood and cannot understand the self-abnegation of our Emperor who wants nothing for himself, but only desires the good of mankind. And what have they promised? Nothing! And what little they have promised they will not perform! Prussia has always declared that Buonaparte is invincible, and that all Europe is powerless before him.... And I don’t believe a word that Hardenburg says, or Haugwitz either. This famous Prussian neutrality is just a trap. I have faith only in God and the lofty destiny of our adored monarch. He will save Europe! What country is this referring to?",
+			},
+		},
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
@@ -36,7 +41,7 @@ func TestLongInputContext(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia"}, 120*time.Second, 10*time.Second)
+	DoChat(ctx, t, client, req, []string{"russia", "german", "france", "england", "austria", "prussia", "europe", "individuals", "coalition", "conflict"}, 120*time.Second, 10*time.Second)
 }

 func TestContextExhaustion(t *testing.T) {
@@ -48,9 +53,14 @@ func TestContextExhaustion(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 	defer cancel()
 	// Set up the test data
-	req := api.GenerateRequest{
-		Model:  "llama2",
-		Prompt: "Write me a story with a ton of emojis?",
+	req := api.ChatRequest{
+		Model: smol,
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: "Write me a story in english with a lot of emojis",
+			},
+		},
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
@@ -63,12 +73,12 @@ func TestContextExhaustion(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived"}, 120*time.Second, 10*time.Second)
+	DoChat(ctx, t, client, req, []string{"once", "upon", "lived", "sunny", "cloudy", "clear", "water", "time", "travel", "world"}, 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
+// Send multiple generate requests with prior context and ensure the response is coherant and expected
+func TestParallelGenerateWithHistory(t *testing.T) {
+	modelOverride := "gpt-oss:20b"
 	req, resp := GenerateRequests()
 	numParallel := 2
 	iterLimit := 2
@@ -111,5 +121,148 @@ func TestGenerateWithHistory(t *testing.T) {
 		}(i)
 	}
 	wg.Wait()
-
+}
+
+// Send generate requests with prior context and ensure the response is coherant and expected
+func TestGenerateWithHistory(t *testing.T) {
+	req := api.GenerateRequest{
+		Model:     smol,
+		Prompt:    rainbowPrompt,
+		Stream:    &stream,
+		KeepAlive: &api.Duration{Duration: 10 * time.Second},
+		Options: map[string]any{
+			"num_ctx": 16384,
+		},
+	}
+
+	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", req.Model)
+	err := client.Generate(ctx,
+		&api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 10 * time.Second}, Options: req.Options},
+		func(response api.GenerateResponse) error { return nil },
+	)
+	if err != nil {
+		t.Fatalf("failed to load model %s: %s", req.Model, err)
+	}
+
+	req.Context = DoGenerate(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
+
+	for i := 0; i < len(rainbowFollowups); i++ {
+		req.Prompt = rainbowFollowups[i]
+		if time.Now().Sub(started) > softTimeout {
+			slog.Info("exceeded soft timeout, winding down test")
+			return
+		}
+		req.Context = DoGenerate(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
+	}
+}
+
+// Send multiple chat requests with prior context and ensure the response is coherant and expected
+func TestParallelChatWithHistory(t *testing.T) {
+	modelOverride := "gpt-oss:20b"
+	req, resp := ChatRequests()
+	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 empty 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
+				assistant := DoChat(ctx, t, client, req[k], resp[k], 120*time.Second, 20*time.Second)
+				if assistant == nil {
+					t.Fatalf("didn't get an assistant response for context")
+				}
+				req[k].Messages = append(req[k].Messages,
+					*assistant,
+					api.Message{Role: "user", Content: "tell me more!"},
+				)
+			}
+		}(i)
+	}
+	wg.Wait()
+}
+
+// Send generate requests with prior context and ensure the response is coherant and expected
+func TestChatWithHistory(t *testing.T) {
+	req := api.ChatRequest{
+		Model:     smol,
+		Stream:    &stream,
+		KeepAlive: &api.Duration{Duration: 10 * time.Second},
+		Options: map[string]any{
+			"num_ctx": 16384,
+		},
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: rainbowPrompt,
+			},
+		},
+	}
+
+	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", req.Model)
+	err := client.Generate(ctx,
+		&api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 10 * time.Second}, Options: req.Options},
+		func(response api.GenerateResponse) error { return nil },
+	)
+	if err != nil {
+		t.Fatalf("failed to load model %s: %s", req.Model, err)
+	}
+
+	assistant := DoChat(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
+
+	for i := 0; i < len(rainbowFollowups); i++ {
+		if time.Now().Sub(started) > softTimeout {
+			slog.Info("exceeded soft timeout, winding down test")
+			return
+		}
+		req.Messages = append(req.Messages,
+			*assistant,
+			api.Message{Role: "user", Content: rainbowFollowups[i]},
+		)
+
+		assistant = DoChat(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
+		if assistant == nil {
+			t.Fatalf("didn't get an assistant response for context")
+		}
+	}
 }
--- a/integration/embed_test.go
+++ b/integration/embed_test.go
@@ -8,6 +8,7 @@ import (
 	"testing"
 	"time"

+	"github.com/google/go-cmp/cmp"
 	"github.com/ollama/ollama/api"
 )

@@ -38,14 +39,14 @@ func TestAllMiniLMEmbeddings(t *testing.T) {
 	defer cleanup()

 	req := api.EmbeddingRequest{
-		Model:  "all-minilm",
-		Prompt: "why is the sky blue?",
+		Model:     "all-minilm",
+		Prompt:    "why is the sky blue?",
+		KeepAlive: &api.Duration{Duration: 10 * time.Second},
 	}

 	res, err := embeddingTestHelper(ctx, client, t, req)
-
 	if err != nil {
-		t.Fatalf("error: %v", err)
+		t.Fatal(err)
 	}

 	if len(res.Embedding) != 384 {
@@ -73,9 +74,8 @@ func TestAllMiniLMEmbed(t *testing.T) {
 	}

 	res, err := embedTestHelper(ctx, client, t, req)
-
 	if err != nil {
-		t.Fatalf("error: %v", err)
+		t.Fatal(err)
 	}

 	if len(res.Embeddings) != 1 {
@@ -111,9 +111,8 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 	}

 	res, err := embedTestHelper(ctx, client, t, req)
-
 	if err != nil {
-		t.Fatalf("error: %v", err)
+		t.Fatal(err)
 	}

 	if len(res.Embeddings) != 2 {
@@ -155,93 +154,135 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {

 	truncTrue, truncFalse := true, false

-	type testReq struct {
-		Name    string
-		Request api.EmbedRequest
+	want, err := embedTestHelper(ctx, client, t, api.EmbedRequest{
+		Model: "all-minilm",
+		Input: "why",
+	})
+	if err != nil {
+		t.Fatal(err)
 	}

-	reqs := []testReq{
+	cases := []struct {
+		name    string
+		request api.EmbedRequest
+		check   func(*api.EmbedResponse, error)
+	}{
 		{
-			Name: "Target Truncation",
-			Request: api.EmbedRequest{
+			name: "target truncation",
+			request: api.EmbedRequest{
 				Model: "all-minilm",
 				Input: "why",
 			},
-		},
-		{
-			Name: "Default Truncate",
-			Request: api.EmbedRequest{
-				Model:   "all-minilm",
-				Input:   "why is the sky blue?",
-				Options: map[string]any{"num_ctx": 1},
+			check: func(got *api.EmbedResponse, err error) {
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
+					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
+				}
 			},
 		},
 		{
-			Name: "Explicit Truncate",
-			Request: api.EmbedRequest{
+			name: "default truncate",
+			request: api.EmbedRequest{
+				Model:   "all-minilm",
+				Input:   "why is the sky blue?",
+				Options: map[string]any{"num_ctx": 3},
+			},
+			check: func(got *api.EmbedResponse, err error) {
+				if err != nil {
+					t.Fatal(err)
+				}
+				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
+					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
+				}
+			},
+		},
+		{
+			name: "explicit truncate",
+			request: api.EmbedRequest{
+				Model:    "all-minilm",
+				Input:    "why is the sky blue?",
+				Truncate: &truncTrue,
+				Options:  map[string]any{"num_ctx": 3},
+			},
+			check: func(got *api.EmbedResponse, err error) {
+				if err != nil {
+					t.Fatal(err)
+				}
+				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
+					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
+				}
+			},
+		},
+		{
+			name: "truncate error",
+			request: api.EmbedRequest{
+				Model:    "all-minilm",
+				Input:    "why is the sky blue?",
+				Truncate: &truncFalse,
+				Options:  map[string]any{"num_ctx": 3},
+			},
+			check: func(res *api.EmbedResponse, err error) {
+				if err.Error() != "input exceeds maximum context length" {
+					t.Fatalf("expected truncation error, got: %v", err)
+				}
+			},
+		},
+		{
+			name: "input after truncate error",
+			request: api.EmbedRequest{
 				Model:    "all-minilm",
 				Input:    "why is the sky blue?",
 				Truncate: &truncTrue,
 				Options:  map[string]any{"num_ctx": 1},
 			},
+			check: func(res *api.EmbedResponse, err error) {
+				if err.Error() != "input after truncation exceeds maximum context length" {
+					t.Fatalf("expected truncation error, got: %v", err)
+				}
+			},
+		},
+		{
+			name: "input after truncate error",
+			request: api.EmbedRequest{
+				Model:    "all-minilm",
+				Input:    "why is the sky blue?",
+				Truncate: &truncTrue,
+				Options:  map[string]any{"num_ctx": 0},
+			},
+			check: func(res *api.EmbedResponse, err error) {
+				if err.Error() != "input after truncation exceeds maximum context length" {
+					t.Fatalf("expected truncation error, got: %v", err)
+				}
+			},
 		},
 	}

-	res := make(map[string]*api.EmbedResponse)
-
-	for _, req := range reqs {
-		response, err := embedTestHelper(ctx, client, t, req.Request)
-		if err != nil {
-			t.Fatalf("error: %v", err)
-		}
-		res[req.Name] = response
-	}
-
-	if res["Target Truncation"].Embeddings[0][0] != res["Default Truncate"].Embeddings[0][0] {
-		t.Fatal("expected default request to truncate correctly")
-	}
-
-	if res["Default Truncate"].Embeddings[0][0] != res["Explicit Truncate"].Embeddings[0][0] {
-		t.Fatal("expected default request and truncate true request to be the same")
-	}
-
-	// check that truncate set to false returns an error if context length is exceeded
-	_, err := embedTestHelper(ctx, client, t, api.EmbedRequest{
-		Model:    "all-minilm",
-		Input:    "why is the sky blue?",
-		Truncate: &truncFalse,
-		Options:  map[string]any{"num_ctx": 1},
-	})
-
-	if err == nil {
-		t.Fatal("expected error, got nil")
+	for _, req := range cases {
+		t.Run(req.name, func(t *testing.T) {
+			req.check(embedTestHelper(ctx, client, t, req.request))
+		})
 	}
 }

 func embeddingTestHelper(ctx context.Context, client *api.Client, t *testing.T, req api.EmbeddingRequest) (*api.EmbeddingResponse, error) {
+	t.Helper()
+
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatalf("failed to pull model %s: %v", req.Model, err)
+		t.Fatal(err)
 	}

-	response, err := client.Embeddings(ctx, &req)
-
-	if err != nil {
-		return nil, err
-	}
-
-	return response, nil
+	return client.Embeddings(ctx, &req)
 }

 func embedTestHelper(ctx context.Context, client *api.Client, t *testing.T, req api.EmbedRequest) (*api.EmbedResponse, error) {
+	t.Helper()
+
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatalf("failed to pull model %s: %v", req.Model, err)
+		t.Fatal(err)
 	}

-	response, err := client.Embed(ctx, &req)
-
-	if err != nil {
-		return nil, err
-	}
-
-	return response, nil
+	return client.Embed(ctx, &req)
 }
--- a/integration/library_models_test.go
+++ b/integration/library_models_test.go
@@ -4,7 +4,9 @@ package integration

 import (
 	"context"
+	"fmt"
 	"log/slog"
+	"os"
 	"testing"
 	"time"

@@ -13,13 +15,14 @@ import (

 // 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) {
+func TestLibraryModelsChat(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()
+	targetArch := os.Getenv("OLLAMA_TEST_ARCHITECTURE")

 	chatModels := libraryChatModels
 	for _, model := range chatModels {
@@ -30,28 +33,43 @@ func TestLibraryModelsGenerate(t *testing.T) {
 			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?",
+			if targetArch != "" {
+				resp, err := client.Show(ctx, &api.ShowRequest{Name: model})
+				if err != nil {
+					t.Fatalf("unable to show model: %s", err)
+				}
+				arch := resp.ModelInfo["general.architecture"].(string)
+				if arch != targetArch {
+					t.Skip(fmt.Sprintf("Skipping %s architecture %s != %s", model, arch, targetArch))
+				}
+			}
+			req := api.ChatRequest{
+				Model: model,
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: blueSkyPrompt,
+					},
+				},
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"temperature": 0.1,
 					"seed":        123,
 				},
 			}
-			anyResp := []string{"rayleigh", "scatter", "atmosphere", "nitrogen", "oxygen", "wavelength"}
+			anyResp := blueSkyExpected
 			// 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" {
+			} else if model == "openthinker" {
 				anyResp = []string{"plugin", "im_sep", "components", "function call"}
 			} else if model == "starcoder" || model == "starcoder2" || model == "magicoder" || model == "deepseek-coder" {
-				req.Prompt = "def fibonacci():"
+				req.Messages[0].Content = "def fibonacci():"
 				anyResp = []string{"f(n)", "sequence", "n-1", "main()", "__main__", "while"}
 			}
-			DoGenerate(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
+			DoChat(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
 		})
 	}
 }
--- a/integration/llm_image_test.go
+++ b/integration/llm_image_test.go
@@ -9,7 +9,6 @@ import (
 	"time"

 	"github.com/ollama/ollama/api"
-	"github.com/stretchr/testify/require"
 )

 func TestVisionModels(t *testing.T) {
@@ -32,18 +31,25 @@ func TestVisionModels(t *testing.T) {
 	for _, v := range testCases {
 		t.Run(v.model, func(t *testing.T) {
 			image, err := base64.StdEncoding.DecodeString(imageEncoding)
-			require.NoError(t, err)
-			req := api.GenerateRequest{
-				Model:  v.model,
-				Prompt: "what does the text in this image say?",
+			if err != nil {
+				t.Fatal(err)
+			}
+			req := api.ChatRequest{
+				Model: v.model,
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "what does the text in this image say?",
+						Images: []api.ImageData{
+							image,
+						},
+					},
+				},
 				Stream: &stream,
 				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
-				Images: []api.ImageData{
-					image,
-				},
 			}
 			ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 			defer cancel()
@@ -52,9 +58,18 @@ func TestVisionModels(t *testing.T) {
 			// Note: sometimes it returns "the ollamas" sometimes "the ollams"
 			resp := "the ollam"
 			defer cleanup()
-			require.NoError(t, PullIfMissing(ctx, client, req.Model))
+			if err := PullIfMissing(ctx, client, req.Model); err != nil {
+				t.Fatal(err)
+			}
+			// Preload to skip if we're less than 80% on GPU to avoid extremely slow tests
+			err = client.Generate(ctx, &api.GenerateRequest{Model: req.Model}, func(response api.GenerateResponse) error { return nil })
+			if err != nil {
+				t.Fatalf("failed to load model %s: %s", req.Model, err)
+			}
+			skipIfNotGPULoaded(ctx, t, client, req.Model, 80)
+
 			// llava models on CPU can be quite slow to start
-			DoGenerate(ctx, t, client, req, []string{resp}, 240*time.Second, 30*time.Second)
+			DoChat(ctx, t, client, req, []string{resp}, 240*time.Second, 30*time.Second)
 		})
 	}
 }
@@ -62,7 +77,9 @@ func TestVisionModels(t *testing.T) {
 func TestIntegrationSplitBatch(t *testing.T) {
 	skipUnderMinVRAM(t, 6)
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
-	require.NoError(t, err)
+	if err != nil {
+		t.Fatal(err)
+	}
 	req := api.GenerateRequest{
 		Model: "gemma3:4b",
 		// Fill up a chunk of the batch so the image will partially spill over into the next one
@@ -84,7 +101,9 @@ func TestIntegrationSplitBatch(t *testing.T) {
 	defer cancel()
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, req.Model))
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatal(err)
+	}
 	// llava models on CPU can be quite slow to start,
 	DoGenerate(ctx, t, client, req, []string{resp}, 120*time.Second, 30*time.Second)
 }
--- a/integration/llm_test.go
+++ b/integration/llm_test.go
@@ -1,47 +0,0 @@
-//go:build integration
-
-package integration
-
-import (
-	"context"
-	"testing"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-// TODO - this would ideally be in the llm package, but that would require some refactoring of interfaces in the server
-//        package to avoid circular dependencies
-
-var (
-	stream = false
-	req    = [2]api.GenerateRequest{
-		{
-			Model:  smol,
-			Prompt: "why is the ocean blue?",
-			Stream: &stream,
-			Options: map[string]any{
-				"seed":        42,
-				"temperature": 0.0,
-			},
-		}, {
-			Model:  smol,
-			Prompt: "what is the origin of the us thanksgiving holiday?",
-			Stream: &stream,
-			Options: map[string]any{
-				"seed":        42,
-				"temperature": 0.0,
-			},
-		},
-	}
-	resp = [2][]string{
-		{"sunlight", "scattering", "interact"},
-		{"england", "english", "massachusetts", "pilgrims"},
-	}
-)
-
-func TestIntegrationSimple(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
-	defer cancel()
-	GenerateTestHelper(ctx, t, req[0], resp[0])
-}
--- a/integration/max_queue_test.go
+++ b/integration/max_queue_test.go
@@ -13,12 +13,12 @@ import (
 	"testing"
 	"time"

-	"github.com/stretchr/testify/require"
-
 	"github.com/ollama/ollama/api"
 )

 func TestMaxQueue(t *testing.T) {
+	t.Skip("this test needs to be re-evaluated to use a proper embedding model")
+
 	if os.Getenv("OLLAMA_TEST_EXISTING") != "" {
 		t.Skip("Max Queue test requires spawning a local server so we can adjust the queue size")
 		return
@@ -45,7 +45,9 @@ func TestMaxQueue(t *testing.T) {
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()

-	require.NoError(t, PullIfMissing(ctx, client, req.Model))
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatal(err)
+	}

 	// Context for the worker threads so we can shut them down
 	// embedCtx, embedCancel := context.WithCancel(ctx)
@@ -89,7 +91,9 @@ func TestMaxQueue(t *testing.T) {
 			switch {
 			case genErr == nil:
 				successCount++
-				require.Greater(t, len(resp.Embedding), 5) // somewhat arbitrary, but sufficient to be reasonable
+				if len(resp.Embedding) < 5 { // somewhat arbitrary, but sufficient to be reasonable
+					t.Fatalf("embeddings shorter than expected: %d", len(resp.Embedding))
+				}
 			case errors.Is(genErr, context.Canceled):
 				canceledCount++
 			case strings.Contains(genErr.Error(), "busy"):
@@ -97,7 +101,9 @@ func TestMaxQueue(t *testing.T) {
 			case strings.Contains(genErr.Error(), "connection reset by peer"):
 				resetByPeerCount++
 			default:
-				require.NoError(t, genErr, "%d request failed", i)
+				if genErr != nil {
+					t.Fatalf("%d request failed", i)
+				}
 			}

 			slog.Info("embed finished", "id", i)
@@ -108,8 +114,13 @@ func TestMaxQueue(t *testing.T) {
 	embedwg.Wait()

 	slog.Info("embeds completed", "success", successCount, "busy", busyCount, "reset", resetByPeerCount, "canceled", canceledCount)
-	require.Equal(t, resetByPeerCount, 0, "Connections reset by peer, have you updated your fd and socket limits?")
-	require.True(t, busyCount > 0, "no requests hit busy error but some should have")
-	require.True(t, canceledCount == 0, "no requests should have been canceled due to timeout")
-
+	if resetByPeerCount != 0 {
+		t.Fatalf("Connections reset by peer, have you updated your fd and socket limits? %d", resetByPeerCount)
+	}
+	if busyCount == 0 {
+		t.Fatalf("no requests hit busy error but some should have")
+	}
+	if canceledCount > 0 {
+		t.Fatalf("no requests should have been canceled due to timeout %d", canceledCount)
+	}
 }
--- a/integration/model_arch_test.go
+++ b/integration/model_arch_test.go
@@ -19,7 +19,7 @@ import (
 	"github.com/ollama/ollama/format"
 )

-func TestModelsGenerate(t *testing.T) {
+func TestModelsChat(t *testing.T) {
 	softTimeout, hardTimeout := getTimeouts(t)
 	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
 	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
@@ -66,16 +66,23 @@ func TestModelsGenerate(t *testing.T) {
 				}
 			}
 			// TODO - fiddle with context size
-			req := api.GenerateRequest{
-				Model:  model,
-				Prompt: "why is the sky blue?",
+			req := api.ChatRequest{
+				Model: model,
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: blueSkyPrompt,
+					},
+				},
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"temperature": 0,
 					"seed":        123,
 				},
 			}
-			anyResp := []string{"rayleigh", "scattering", "atmosphere", "nitrogen", "oxygen"}
-			DoGenerate(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
+			DoChat(ctx, t, client, req, blueSkyExpected, 120*time.Second, 30*time.Second)
+			// best effort unload once we're done with the model
+			client.Generate(ctx, &api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 0}}, func(rsp api.GenerateResponse) error { return nil })
 		})
 	}
 }
@@ -129,8 +136,9 @@ func TestModelsEmbed(t *testing.T) {
 				}
 			}
 			req := api.EmbeddingRequest{
-				Model:  model,
-				Prompt: "why is the sky blue?",
+				Model:     model,
+				Prompt:    "why is the sky blue?",
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"temperature": 0,
 					"seed":        123,
@@ -140,6 +148,10 @@ func TestModelsEmbed(t *testing.T) {
 			if err != nil {
 				t.Fatalf("embeddings call failed %s", err)
 			}
+			defer func() {
+				// best effort unload once we're done with the model
+				client.Generate(ctx, &api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 0}}, func(rsp api.GenerateResponse) error { return nil })
+			}()
 			if len(resp.Embedding) == 0 {
 				t.Errorf("zero length embedding response")
 			}
--- a/integration/model_perf_test.go
+++ b/integration/model_perf_test.go
@@ -40,6 +40,18 @@ var (
 // 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) {
+	if s := os.Getenv("OLLAMA_NEW_ENGINE"); s != "" {
+		doModelPerfTest(t, ollamaEngineChatModels)
+	} else {
+		doModelPerfTest(t, append(ollamaEngineChatModels, llamaRunnerChatModels...))
+	}
+}
+
+func TestLibraryModelsPerf(t *testing.T) {
+	doModelPerfTest(t, libraryChatModels)
+}
+
+func doModelPerfTest(t *testing.T, chatModels []string) {
 	softTimeout, hardTimeout := getTimeouts(t)
 	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
 	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
@@ -65,14 +77,12 @@ func TestModelsPerf(t *testing.T) {
 	}
 	longPrompt := "summarize the following: " + string(data)

-	var chatModels []string
-	if s := os.Getenv("OLLAMA_NEW_ENGINE"); s != "" {
-		chatModels = ollamaEngineChatModels
-	} else {
-		chatModels = append(ollamaEngineChatModels, llamaRunnerChatModels...)
-	}
+	targetArch := os.Getenv("OLLAMA_TEST_ARCHITECTURE")

 	for _, model := range chatModels {
+		if !strings.Contains(model, ":") {
+			model = model + ":latest"
+		}
 		t.Run(model, func(t *testing.T) {
 			if time.Now().Sub(started) > softTimeout {
 				t.Skip("skipping remaining tests to avoid excessive runtime")
@@ -88,6 +98,9 @@ func TestModelsPerf(t *testing.T) {
 			}
 			arch := resp.ModelInfo["general.architecture"].(string)
 			maxContext = int(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))
+			if targetArch != "" && arch != targetArch {
+				t.Skip(fmt.Sprintf("Skipping %s architecture %s != %s", model, arch, targetArch))
+			}

 			if maxVram > 0 {
 				resp, err := client.List(ctx)
@@ -151,8 +164,8 @@ func TestModelsPerf(t *testing.T) {
 					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"}},
+					{blueSkyPrompt, blueSkyExpected},
+					{maxPrompt, []string{"shakespeare", "oppression", "sorrows", "gutenberg", "child", "license", "sonnet", "melancholy", "love", "sorrow", "beauty"}},
 				}
 				var gpuPercent int
 				for _, tc := range testCases {
@@ -160,9 +173,14 @@ func TestModelsPerf(t *testing.T) {
 						slog.Info("skipping long prompt", "model", model, "num_ctx", numCtx, "gpu_percent", gpuPercent)
 						continue
 					}
-					req := api.GenerateRequest{
-						Model:     model,
-						Prompt:    tc.prompt,
+					req := api.ChatRequest{
+						Model: model,
+						Messages: []api.Message{
+							{
+								Role:    "user",
+								Content: tc.prompt,
+							},
+						},
 						KeepAlive: &api.Duration{Duration: 20 * time.Second}, // long enough to ensure a ps returns
 						Options: map[string]interface{}{
 							"temperature": 0,
@@ -171,7 +189,7 @@ func TestModelsPerf(t *testing.T) {
 						},
 					}
 					atLeastOne := false
-					var resp api.GenerateResponse
+					var resp api.ChatResponse

 					stream := false
 					req.Stream = &stream
@@ -185,7 +203,7 @@ func TestModelsPerf(t *testing.T) {
 					)
 					defer cancel()

-					err = client.Generate(genCtx, &req, func(rsp api.GenerateResponse) error {
+					err = client.Chat(genCtx, &req, func(rsp api.ChatResponse) error {
 						resp = rsp
 						return nil
 					})
@@ -201,13 +219,13 @@ func TestModelsPerf(t *testing.T) {
 					}
 					loaded = true
 					for _, expResp := range tc.anyResp {
-						if strings.Contains(strings.ToLower(resp.Response), expResp) {
+						if strings.Contains(strings.ToLower(resp.Message.Content), 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)
+						t.Fatalf("response didn't contain expected values: ctx:%d  expected:%v response:%s ", numCtx, tc.anyResp, resp.Message.Content)
 					}
 					models, err := client.ListRunning(ctx)
 					if err != nil {
@@ -241,11 +259,12 @@ func TestModelsPerf(t *testing.T) {
 							}
 						}
 					}
+					// Round the logged prompt count for comparisons across versions/configurations which can vary slightly
 					fmt.Fprintf(os.Stderr, "MODEL_PERF_HEADER:%s,%s,%s,%s,%s,%s,%s\n",
 						"MODEL",
 						"CONTEXT",
 						"GPU PERCENT",
-						"PROMPT COUNT",
+						"APPROX PROMPT COUNT",
 						"LOAD TIME",
 						"PROMPT EVAL TPS",
 						"EVAL TPS",
@@ -254,7 +273,7 @@ func TestModelsPerf(t *testing.T) {
 						model,
 						numCtx,
 						gpuPercent,
-						resp.PromptEvalCount,
+						(resp.PromptEvalCount/10)*10,
 						float64(resp.LoadDuration)/1000000000.0,
 						float64(resp.PromptEvalCount)/(float64(resp.PromptEvalDuration)/1000000000.0),
 						float64(resp.EvalCount)/(float64(resp.EvalDuration)/1000000000.0),
--- a/integration/quantization_test.go
+++ b/integration/quantization_test.go
@@ -74,9 +74,14 @@ func TestQuantization(t *testing.T) {
 				}

 				stream := true
-				genReq := api.GenerateRequest{
-					Model:     newName,
-					Prompt:    "why is the sky blue?",
+				chatReq := api.ChatRequest{
+					Model: newName,
+					Messages: []api.Message{
+						{
+							Role:    "user",
+							Content: blueSkyPrompt,
+						},
+					},
 					KeepAlive: &api.Duration{Duration: 3 * time.Second},
 					Options: map[string]any{
 						"seed":        42,
@@ -88,14 +93,13 @@ func TestQuantization(t *testing.T) {

 				// Some smaller quantizations can cause models to have poor quality
 				// or get stuck in repetition loops, so we stop as soon as we have any matches
-				anyResp := []string{"rayleigh", "scattering", "day", "sun", "moon", "color", "nitrogen", "oxygen"}
 				reqCtx, reqCancel := context.WithCancel(ctx)
 				atLeastOne := false
 				var buf bytes.Buffer
-				genfn := func(response api.GenerateResponse) error {
-					buf.Write([]byte(response.Response))
+				chatfn := func(response api.ChatResponse) error {
+					buf.Write([]byte(response.Message.Content))
 					fullResp := strings.ToLower(buf.String())
-					for _, resp := range anyResp {
+					for _, resp := range blueSkyExpected {
 						if strings.Contains(fullResp, resp) {
 							atLeastOne = true
 							t.Log(fullResp)
@@ -109,14 +113,14 @@ func TestQuantization(t *testing.T) {
 				done := make(chan int)
 				var genErr error
 				go func() {
-					genErr = client.Generate(reqCtx, &genReq, genfn)
+					genErr = client.Chat(reqCtx, &chatReq, chatfn)
 					done <- 0
 				}()

 				select {
 				case <-done:
 					if genErr != nil && !atLeastOne {
-						t.Fatalf("failed with %s request prompt %s ", genReq.Model, genReq.Prompt)
+						t.Fatalf("failed with %s request prompt %s ", chatReq.Model, chatReq.Messages[0].Content)
 					}
 				case <-ctx.Done():
 					t.Error("outer test context done while waiting for generate")
--- a/integration/utils_test.go
+++ b/integration/utils_test.go
@@ -9,11 +9,13 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
+	"math"
 	"math/rand"
 	"net"
 	"net/http"
 	"net/url"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"runtime"
 	"strconv"
@@ -23,13 +25,12 @@ import (
 	"time"

 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/app/lifecycle"
 	"github.com/ollama/ollama/format"
-	"github.com/stretchr/testify/require"
 )

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

 var (
@@ -37,6 +38,7 @@ var (

 	// Note: add newer models at the top of the list to test them first
 	ollamaEngineChatModels = []string{
+		"qwen3-coder:30b",
 		"gpt-oss:20b",
 		"gemma3n:e2b",
 		"mistral-small3.2:latest",
@@ -45,6 +47,7 @@ var (
 		"qwen2.5-coder:latest",
 		"qwen2.5vl:3b",
 		"qwen3:0.6b", // dense
+		"qwen3:1.7b", // dense
 		"qwen3:30b",  // MOE
 		"gemma3:1b",
 		"llama3.1:latest",
@@ -255,13 +258,28 @@ var (
 		"snowflake-arctic-embed",
 		"snowflake-arctic-embed2",
 	}
+
+	blueSkyPrompt   = "why is the sky blue? Be brief but factual in your reply"
+	blueSkyExpected = []string{"rayleigh", "scatter", "atmosphere", "nitrogen", "oxygen", "wavelength", "interact"}
+
+	rainbowPrompt    = "how do rainbows form? Be brief but factual in your reply"
+	rainbowFollowups = []string{
+		"Explain the physics involved in them.  Be breif in your reply",
+		"Explain the chemistry involved in them.  Be breif in your reply",
+		"What are common myths related to them? Be brief in your reply",
+		"Can they form if there is no rain?  Be breif in your reply",
+		"Can they form if there are no clouds?  Be breif in your reply",
+		"Do they happen on other planets? Be brief in your reply",
+	}
+	rainbowExpected = []string{"water", "droplet", "mist", "glow", "refract", "reflect", "scatter", "particles", "wave", "color", "spectrum", "raindrop", "atmosphere", "frequency", "shower", "sky", "shimmer", "light", "storm", "sunny", "sunburst", "phenomenon", "mars", "venus", "jupiter"}
 )

 func init() {
-	lifecycle.InitLogging()
-	custom := os.Getenv("OLLAMA_TEST_SMOL_MODEL")
+	logger := slog.New(slog.NewTextHandler(os.Stdout, &slog.HandlerOptions{Level: slog.LevelDebug}))
+	slog.SetDefault(logger)
+	custom := os.Getenv("OLLAMA_TEST_DEFAULT_MODEL")
 	if custom != "" {
-		slog.Info("setting smol test model to " + custom)
+		slog.Info("setting default test model to " + custom)
 		smol = custom
 	}
 }
@@ -319,6 +337,7 @@ func GetTestEndpoint() (*api.Client, string) {

 var serverMutex sync.Mutex
 var serverReady bool
+var serverLogFile string

 func startServer(t *testing.T, ctx context.Context, ollamaHost string) error {
 	// Make sure the server has been built
@@ -345,8 +364,9 @@ func startServer(t *testing.T, ctx context.Context, ollamaHost string) error {
 		t.Setenv("OLLAMA_HOST", ollamaHost)
 	}

+	logDir := t.TempDir()
 	slog.Info("starting server", "url", ollamaHost)
-	done, err := lifecycle.SpawnServer(ctx, "../ollama")
+	done, err := SpawnServer(ctx, "../ollama", logDir)
 	if err != nil {
 		return fmt.Errorf("failed to start server: %w", err)
 	}
@@ -369,6 +389,36 @@ func startServer(t *testing.T, ctx context.Context, ollamaHost string) error {
 	return nil
 }

+func SpawnServer(ctx context.Context, command, logDir string) (chan int, error) {
+	done := make(chan int)
+	fp, err := os.CreateTemp(logDir, "ollama-server-*.log")
+	if err != nil {
+		return nil, fmt.Errorf("failed to create log file: %w", err)
+	}
+	serverLogFile = fp.Name()
+
+	cmd := exec.CommandContext(ctx, command, "serve")
+	cmd.Stderr = fp
+	cmd.Stdout = fp
+
+	go func() {
+		slog.Info("starting server...")
+		if err := cmd.Run(); err != nil {
+			// "signal: killed" expected
+			if !strings.Contains(err.Error(), "signal") {
+				slog.Info("failed to run server", "error", err)
+			}
+		}
+		var code int
+		if cmd.ProcessState != nil {
+			code = cmd.ProcessState.ExitCode()
+		}
+		slog.Info("server exited")
+		done <- code
+	}()
+	return done, nil
+}
+
 func PullIfMissing(ctx context.Context, client *api.Client, modelName string) error {
 	slog.Info("checking status of model", "model", modelName)
 	showReq := &api.ShowRequest{Name: modelName}
@@ -429,47 +479,59 @@ func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, strin
 	client, testEndpoint := GetTestEndpoint()
 	if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
 		serverProcMutex.Lock()
-		fp, err := os.CreateTemp("", "ollama-server-*.log")
-		if err != nil {
-			t.Fatalf("failed to generate log file: %s", err)
+		if err := startServer(t, ctx, testEndpoint); err != nil {
+			t.Fatal(err)
 		}
-		lifecycle.ServerLogFile = fp.Name()
-		fp.Close()
-		require.NoError(t, startServer(t, ctx, testEndpoint))
+	}
+	// Make sure server is online and healthy before returning
+	listCtx, cancel := context.WithDeadlineCause(
+		ctx,
+		time.Now().Add(120*time.Second),
+		fmt.Errorf("list models took too long"),
+	)
+	defer cancel()
+	models, err := client.ListRunning(listCtx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(models.Models) > 0 {
+		names := make([]string, len(models.Models))
+		for i, m := range models.Models {
+			names[i] = m.Name
+		}
+		slog.Info("currently loaded", "models", names)
 	}

 	return client, testEndpoint, func() {
 		if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
 			defer serverProcMutex.Unlock()
 			if t.Failed() {
-				fp, err := os.Open(lifecycle.ServerLogFile)
+				fp, err := os.Open(serverLogFile)
 				if err != nil {
-					slog.Error("failed to open server log", "logfile", lifecycle.ServerLogFile, "error", err)
+					slog.Error("failed to open server log", "logfile", serverLogFile, "error", err)
 					return
 				}
 				defer fp.Close()
 				data, err := io.ReadAll(fp)
 				if err != nil {
-					slog.Error("failed to read server log", "logfile", lifecycle.ServerLogFile, "error", err)
+					slog.Error("failed to read server log", "logfile", serverLogFile, "error", err)
 					return
 				}
 				slog.Warn("SERVER LOG FOLLOWS")
 				os.Stderr.Write(data)
 				slog.Warn("END OF SERVER")
 			}
-			err := os.Remove(lifecycle.ServerLogFile)
-			if err != nil && !os.IsNotExist(err) {
-				slog.Warn("failed to cleanup", "logfile", lifecycle.ServerLogFile, "error", err)
-			}
 		}
 	}
 }

-func GenerateTestHelper(ctx context.Context, t *testing.T, genReq api.GenerateRequest, anyResp []string) {
+func ChatTestHelper(ctx context.Context, t *testing.T, req api.ChatRequest, anyResp []string) {
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, genReq.Model))
-	DoGenerate(ctx, t, client, genReq, anyResp, 30*time.Second, 10*time.Second)
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatal(err)
+	}
+	DoChat(ctx, t, client, req, 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 {
@@ -497,6 +559,22 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 		done <- 0
 	}()

+	var response string
+	verify := func() {
+		// Verify the response contains the expected data
+		response = buf.String()
+		atLeastOne := false
+		for _, resp := range anyResp {
+			if strings.Contains(strings.ToLower(response), resp) {
+				atLeastOne = true
+				break
+			}
+		}
+		if !atLeastOne {
+			t.Fatalf("%s: none of %v found in %s", genReq.Model, anyResp, response)
+		}
+	}
+
 	select {
 	case <-stallTimer.C:
 		if buf.Len() == 0 {
@@ -509,20 +587,17 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 			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()
-		atLeastOne := false
-		for _, resp := range anyResp {
-			if strings.Contains(strings.ToLower(response), resp) {
-				atLeastOne = true
-				break
-			}
+		if genErr != nil {
+			t.Fatalf("%s failed with %s request prompt %s", genErr, genReq.Model, genReq.Prompt)
 		}
-		require.True(t, atLeastOne, "%s: none of %v found in %s", genReq.Model, anyResp, response)
+		verify()
 		slog.Info("test pass", "model", genReq.Model, "prompt", genReq.Prompt, "contains", anyResp, "response", response)
 	case <-ctx.Done():
-		t.Error("outer test context done while waiting for generate")
+		// On slow systems, we might timeout before some models finish rambling, so check what we have so far to see
+		// if it's considered a pass - the stallTimer will detect hangs, but we want to consider slow systems a pass
+		// if they are still generating valid responses
+		slog.Warn("outer test context done while waiting for generate")
+		verify()
 	}
 	return context
 }
@@ -543,7 +618,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 			}, {
 				Model:     smol,
-				Prompt:    "what is the origin of the US thanksgiving holiday? Be brief but factual in your reply",
+				Prompt:    rainbowPrompt,
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 			}, {
@@ -559,19 +634,106 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 			},
 		},
 		[][]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", "scatter", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorb", "wavelength", "water", "molecule"},
+			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigment", "particle", "iron oxide", "rust", "air", "water", "wet", "mixture", "mixing", "mineral", "element", "decomposed", "matter", "wavelength"},
+			rainbowExpected,
 			{"fourth", "july", "declaration", "independence"},
-			{"nitrogen", "oxygen", "carbon", "dioxide"},
+			{"nitrogen", "oxygen", "carbon", "dioxide", "water", "vapor", "fluid", "particles", "gas"},
 		}
 }

+func DoChat(ctx context.Context, t *testing.T, client *api.Client, req api.ChatRequest, anyResp []string, initialTimeout, streamTimeout time.Duration) *api.Message {
+	stallTimer := time.NewTimer(initialTimeout)
+	var buf bytes.Buffer
+	role := "assistant"
+	fn := func(response api.ChatResponse) error {
+		// fmt.Print(".")
+		role = response.Message.Role
+		buf.Write([]byte(response.Message.Content))
+		if !stallTimer.Reset(streamTimeout) {
+			return errors.New("stall was detected while streaming response, aborting")
+		}
+		return nil
+	}
+
+	stream := true
+	req.Stream = &stream
+	done := make(chan int)
+	var genErr error
+	go func() {
+		genErr = client.Chat(ctx, &req, fn)
+		done <- 0
+	}()
+
+	var response string
+	verify := func() {
+		// Verify the response contains the expected data
+		response = buf.String()
+		atLeastOne := false
+		for _, resp := range anyResp {
+			if strings.Contains(strings.ToLower(response), resp) {
+				atLeastOne = true
+				break
+			}
+		}
+		if !atLeastOne {
+			t.Fatalf("%s: none of %v found in \"%s\" -- request was:%v", req.Model, anyResp, response, req.Messages)
+		}
+	}
+
+	select {
+	case <-stallTimer.C:
+		if buf.Len() == 0 {
+			t.Errorf("generate never started.  Timed out after :%s", initialTimeout.String())
+		} else {
+			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", req.Model, "error", genErr)
+			return nil
+		}
+		if genErr != nil {
+			t.Fatalf("%s failed with %s request prompt %v", genErr, req.Model, req.Messages)
+		}
+		verify()
+		slog.Info("test pass", "model", req.Model, "messages", req.Messages, "contains", anyResp, "response", response)
+	case <-ctx.Done():
+		// On slow systems, we might timeout before some models finish rambling, so check what we have so far to see
+		// if it's considered a pass - the stallTimer will detect hangs, but we want to consider slow systems a pass
+		// if they are still generating valid responses
+		slog.Warn("outer test context done while waiting for chat")
+		verify()
+	}
+	return &api.Message{Role: role, Content: buf.String()}
+}
+
+func ChatRequests() ([]api.ChatRequest, [][]string) {
+	genReqs, results := GenerateRequests()
+	reqs := make([]api.ChatRequest, len(genReqs))
+	// think := api.ThinkValue{Value: "low"}
+	for i := range reqs {
+		reqs[i].Model = genReqs[i].Model
+		reqs[i].Stream = genReqs[i].Stream
+		reqs[i].KeepAlive = genReqs[i].KeepAlive
+		// reqs[i].Think = &think
+		reqs[i].Messages = []api.Message{
+			{
+				Role:    "user",
+				Content: genReqs[i].Prompt,
+			},
+		}
+	}
+	return reqs, results
+}
+
 func skipUnderMinVRAM(t *testing.T, gb uint64) {
 	// TODO use info API in the future
 	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
 		maxVram, err := strconv.ParseUint(s, 10, 64)
-		require.NoError(t, err)
+		if err != nil {
+			t.Fatal(err)
+		}
 		// Don't hammer on small VRAM cards...
 		if maxVram < gb*format.GibiByte {
 			t.Skip("skipping with small VRAM to avoid timeouts")
@@ -579,6 +741,45 @@ func skipUnderMinVRAM(t *testing.T, gb uint64) {
 	}
 }

+// Skip if the target model isn't X% GPU loaded to avoid excessive runtime
+func skipIfNotGPULoaded(ctx context.Context, t *testing.T, client *api.Client, model string, minPercent int) {
+	models, err := client.ListRunning(ctx)
+	if err != nil {
+		t.Fatalf("failed to list running models: %s", err)
+	}
+	loaded := []string{}
+	for _, m := range models.Models {
+		loaded = append(loaded, m.Name)
+		if strings.Contains(model, ":") {
+			if m.Name != model {
+				continue
+			}
+		} else if strings.Contains(m.Name, ":") {
+			if !strings.HasPrefix(m.Name, model+":") {
+				continue
+			}
+		}
+		gpuPercent := 0
+		switch {
+		case m.SizeVRAM == 0:
+			gpuPercent = 0
+		case m.SizeVRAM == m.Size:
+			gpuPercent = 100
+		case m.SizeVRAM > m.Size || m.Size == 0:
+			t.Logf("unexpected size detected: %d", m.SizeVRAM)
+		default:
+			sizeCPU := m.Size - m.SizeVRAM
+			cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 110)
+			gpuPercent = int(100 - cpuPercent)
+		}
+		if gpuPercent < minPercent {
+			t.Skip(fmt.Sprintf("test requires minimum %d%% GPU load, but model %s only has %d%%", minPercent, model, gpuPercent))
+		}
+		return
+	}
+	t.Skip(fmt.Sprintf("model %s not loaded - actually loaded: %v", model, loaded))
+}
+
 func getTimeouts(t *testing.T) (soft time.Duration, hard time.Duration) {
 	deadline, hasDeadline := t.Deadline()
 	if !hasDeadline {
--- a/kvcache/causal.go
+++ b/kvcache/causal.go
@@ -160,7 +160,15 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
 	if c.swaMemorySize == 0 {
 		c.swaMemorySize = c.swaWindowSize
 	}
-	if int(c.swaMemorySize) > capacity {
+	// We will allocate space in the cache for the stop token, which won't be part of a follow on
+	// sequence, so allocate an extra token of storage to ensure that we can jump back without
+	// causing a cache break. As an optimization, only do this when we have parallel sequences
+	// because the extra token will live in the batch buffer and won't get overwritten if we
+	// only have a single sequence.
+	if c.swaMemorySize != math.MaxInt32 && maxSequences > 1 {
+		c.swaMemorySize = max(c.swaMemorySize, c.swaWindowSize+1)
+	}
+	if int(c.swaMemorySize) >= capacity {
 		c.swaMemorySize = math.MaxInt32
 	}

@@ -214,7 +222,6 @@ 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
 		}

@@ -288,23 +295,44 @@ func (c *Causal) updateSlidingWindow() {
 		return
 	}

+	type lowestPosition struct {
+		pos      int32
+		curBatch bool
+	}
+
 	// create a map of unique sequences to the lowest position in that sequence
-	lowestPos := make(map[int]int32)
+	lowestPos := make(map[int]lowestPosition)
 	for i := range c.curPositions {
 		seq := c.curSequences[i]

-		pos, ok := lowestPos[seq]
+		lowest, ok := lowestPos[seq]
 		if !ok {
-			pos = c.curPositions[i]
-		} else if c.curPositions[i] < pos {
-			pos = c.curPositions[i]
+			lowest = lowestPosition{pos: c.curPositions[i], curBatch: true}
+		} else if c.curPositions[i] < lowest.pos {
+			lowest.pos = c.curPositions[i]
 		}

-		lowestPos[seq] = pos
+		lowestPos[seq] = lowest
+	}
+
+	// for any sequences are not part of this batch, clean up any tokens
+	// that are no longer needed after the processing of the previous
+	// batch
+	for seq, seqRange := range c.cellRanges {
+		if _, ok := lowestPos[seq]; !ok {
+			var last int32
+			for i := seqRange.min; i <= seqRange.max; i++ {
+				if slices.Contains(c.cells[i].sequences, seq) {
+					last = max(last, c.cells[i].pos)
+				}
+			}
+
+			lowestPos[seq] = lowestPosition{pos: last + 1, curBatch: false}
+		}
 	}

 	// delete any entries that are beyond the window of the oldest position in the sequence
-	for seq, pos := range lowestPos {
+	for seq, lowest := range lowestPos {
 		oldRange, ok := c.cellRanges[seq]
 		if !ok {
 			continue
@@ -314,13 +342,13 @@ 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 < lowest.pos-c.swaMemorySize {
 					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 {
+				if lowest.curBatch && c.cells[i].pos >= lowest.pos-c.swaWindowSize {
 					c.curCellRange.min = min(c.curCellRange.min, i)
 					c.curCellRange.max = max(c.curCellRange.max, i)
 				}
@@ -378,9 +406,7 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 	maskTensor := ctx.Input().FromFloatSlice(mask, length, batchSize)

 	if c.config.MaskDType != ml.DTypeF32 {
-		out := ctx.Input().Empty(c.config.MaskDType, maskTensor.Shape()...)
-		ctx.Forward(maskTensor.Copy(ctx, out))
-		maskTensor = out
+		maskTensor = maskTensor.Cast(ctx, c.config.MaskDType)
 	}

 	return maskTensor
@@ -659,9 +685,11 @@ func (c *Causal) CanResume(seq int, pos int32) bool {

 	// for sliding window, check that the window of the new sequence is contained in
 	// the window of what we are storing
+	var first int32 = math.MaxInt32
 	var last int32 = -1
 	for i := seqRange.min; i <= seqRange.max; i++ {
 		if slices.Contains(c.cells[i].sequences, seq) {
+			first = min(first, c.cells[i].pos)
 			last = max(last, c.cells[i].pos)
 		}
 	}
@@ -670,10 +698,8 @@ func (c *Causal) CanResume(seq int, pos int32) bool {
 		return false
 	}

-	lastWindowStart := max(0, last-c.swaMemorySize)
 	posWindowStart := max(0, pos-c.swaWindowSize)
-
-	return posWindowStart >= lastWindowStart
+	return posWindowStart >= first && pos <= last+1
 }

 func (c *Causal) shift(seq int, beginIndex, offset int32) error {
--- a/kvcache/causal_test.go
+++ b/kvcache/causal_test.go
@@ -96,6 +96,86 @@ func TestSWA(t *testing.T) {
 	testCache(t, backend, cache, tests)
 }

+func TestSWASeparateBatches(t *testing.T) {
+	backend := &testBackend{}
+	cache := NewSWACache(1, nil)
+	defer cache.Close()
+
+	cache.Init(backend, ml.DTypeF16, 2, 16, 2)
+
+	x := float32(math.Inf(-1))
+
+	tests := []testCase{
+		{
+			name:          "First seq 0",
+			in:            []float32{1, 2},
+			inShape:       []int{1, 1, 2},
+			seqs:          []int{0, 0},
+			pos:           []int32{0, 1},
+			expected:      []float32{1, 2},
+			expectedShape: []int{1, 1, 2},
+			expectedMask: []float32{
+				0, x,
+				0, 0,
+			},
+		},
+		{
+			name:          "Second seq 0",
+			in:            []float32{3, 4},
+			inShape:       []int{1, 1, 2},
+			seqs:          []int{0, 0},
+			pos:           []int32{2, 3},
+			expected:      []float32{2, 3, 4},
+			expectedShape: []int{1, 1, 3},
+			expectedMask: []float32{
+				0, 0, x,
+				x, 0, 0,
+			},
+		},
+		{
+			name:          "First seq 1",
+			in:            []float32{5, 6},
+			inShape:       []int{1, 1, 2},
+			seqs:          []int{1, 1},
+			pos:           []int32{0, 1},
+			expected:      []float32{5, 6},
+			expectedShape: []int{1, 1, 2},
+			expectedMask: []float32{
+				0, x,
+				0, 0,
+			},
+		},
+		{
+			name:          "Second seq 1",
+			in:            []float32{7, 8},
+			inShape:       []int{1, 1, 2},
+			seqs:          []int{1, 1},
+			pos:           []int32{2, 3},
+			expected:      []float32{6, 3, 4, 7, 8},
+			expectedShape: []int{1, 1, 5},
+			expectedMask: []float32{
+				0, x, x, 0, x,
+				x, x, x, 0, 0,
+			},
+		},
+		{
+			name:          "Third seq 0",
+			in:            []float32{9, 10},
+			inShape:       []int{1, 1, 2},
+			seqs:          []int{0, 0},
+			pos:           []int32{4, 5},
+			expected:      []float32{9, 10, 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)
@@ -431,15 +511,15 @@ func TestCanResume(t *testing.T) {
 	defer context.Close()

 	err := cache.StartForward(context, input.Batch{
-		Positions: []int32{0, 1, 2, 3},
-		Sequences: []int{0, 0, 0, 0},
+		Positions: []int32{0, 1, 2, 3, 4},
+		Sequences: []int{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}, 1, 1, 4)
+	tensor := context.FromFloatSlice([]float32{1, 2, 3, 4, 5}, 1, 1, 5)
 	cache.Put(context, tensor, tensor)

 	// with window size 4, nothing has slid out of the window yet
@@ -455,18 +535,21 @@ func TestCanResume(t *testing.T) {
 	if !cache.CanResume(0, 3) {
 		t.Errorf("CanResume(0, 3) = false, want true (latest position)")
 	}
+	if !cache.CanResume(0, 4) {
+		t.Errorf("CanResume(0, 4) = false, want true (latest position)")
+	}

-	// shift window by adding position 4
+	// shift window by adding position 5
 	err = cache.StartForward(context, input.Batch{
-		Positions: []int32{4, 5},
-		Sequences: []int{0, 0},
+		Positions: []int32{5},
+		Sequences: []int{0},
 	}, false)
 	if err != nil {
 		t.Fatalf("StartForward failed: %v", err)
 	}

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

 	// only the latest position has overlapping windows
@@ -503,28 +586,28 @@ func TestCanResumeSWAMem(t *testing.T) {
 	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},
+		Positions: []int32{0, 1, 2, 3, 4, 5, 6},
+		Sequences: []int{0, 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)
+	tensor := context.FromFloatSlice([]float32{1, 2, 3, 4, 5, 6, 7}, 1, 1, 7)
 	cache.Put(context, tensor, tensor)

-	// shift window by adding position 6
+	// shift window by adding position 7
 	err = cache.StartForward(context, input.Batch{
-		Positions: []int32{6, 7},
-		Sequences: []int{0, 0},
+		Positions: []int32{7},
+		Sequences: []int{0},
 	}, false)
 	if err != nil {
 		t.Fatalf("StartForward failed: %v", err)
 	}

 	cache.SetLayer(0)
-	tensor = context.FromFloatSlice([]float32{7, 8}, 1, 1, 2)
+	tensor = context.FromFloatSlice([]float32{8}, 1, 1, 1)
 	cache.Put(context, tensor, tensor)

 	// only the latest position has overlapping windows
--- 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 = "364a7a6d4a786e98947c8a90430ea581213c0ba9";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";
--- a/llama/llama.cpp/common/common.cpp
+++ b/llama/llama.cpp/common/common.cpp
@@ -14,6 +14,7 @@
 #include <climits>
 #include <cmath>
 #include <codecvt>
+#include <chrono>
 #include <cstdarg>
 #include <cstring>
 #include <ctime>
@@ -41,6 +42,7 @@
 #endif
 #include <locale>
 #include <windows.h>
+#include <string.h>
 #include <fcntl.h>
 #include <io.h>
 #else
@@ -49,6 +51,11 @@
 #include <unistd.h>
 #endif

+#if defined(__linux__)
+#include <sys/types.h>
+#include <pwd.h>
+#endif
+
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
@@ -557,13 +564,6 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam

        auto detokenized = common_token_to_piece(ctx, token);

-        detokenized.erase(
-            std::remove_if(
-                detokenized.begin(),
-                detokenized.end(),
-                [](const unsigned char c) { return !std::isprint(c); }),
-            detokenized.end());
-
        buf << "'" << detokenized << "'"
            << ":" << std::to_string(token);
    }
@@ -588,13 +588,6 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat

        auto detokenized = common_token_to_piece(ctx, batch.token[i]);

-        detokenized.erase(
-                std::remove_if(
-                    detokenized.begin(),
-                    detokenized.end(),
-                    [](const unsigned char c) { return !std::isprint(c); }),
-                detokenized.end());
-
        buf << "\n"          << std::to_string(i)
            << ", token '"   << detokenized << "'"
            << ", pos "      << std::to_string(batch.pos[i])
@@ -877,8 +870,20 @@ std::string fs_get_cache_directory() {
 #if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
        if (std::getenv("XDG_CACHE_HOME")) {
            cache_directory = std::getenv("XDG_CACHE_HOME");
-        } else {
+        } else if (std::getenv("HOME")) {
            cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        } else {
+#if defined(__linux__)
+            /* no $HOME is defined, fallback to getpwuid */
+            struct passwd *pw = getpwuid(getuid());
+            if ((!pw) || (!pw->pw_dir)) {
+                throw std::runtime_error("Failed to find $HOME directory");
+            }
+
+            cache_directory = std::string(pw->pw_dir) + std::string("/.cache/");
+#else /* defined(__linux__) */
+            throw std::runtime_error("Failed to find $HOME directory");
+#endif /* defined(__linux__) */
        }
 #elif defined(__APPLE__)
        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
@@ -914,7 +919,8 @@ struct common_init_result common_init_from_params(common_params & params) {

    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
    if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to load model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
        return iparams;
    }

@@ -924,7 +930,8 @@ struct common_init_result common_init_from_params(common_params & params) {

    llama_context * lctx = llama_init_from_model(model, cparams);
    if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
        llama_model_free(model);
        return iparams;
    }
@@ -971,15 +978,13 @@ struct common_init_result common_init_from_params(common_params & params) {

        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+        bool has_rerank_prompt = llama_model_chat_template(model, "rerank") != 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__);
+        if (!has_eos && !has_sep && !has_rerank_prompt) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, SEP token, or rerank prompt. 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) {
@@ -1001,7 +1006,12 @@ struct common_init_result common_init_from_params(common_params & params) {
            return iparams;
        }

+        char buf[1024];
        la.ptr = lora.get();
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
+        la.task_name = buf;
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
+        la.prompt_prefix = buf;
        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
    }

@@ -1165,11 +1175,10 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
    cparams.pooling_type      = params.pooling_type;
    cparams.attention_type    = params.attention_type;
-    cparams.defrag_thold      = params.defrag_thold;
+    cparams.flash_attn_type   = params.flash_attn_type;
    cparams.cb_eval           = params.cb_eval;
    cparams.cb_eval_user_data = params.cb_eval_user_data;
    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;
@@ -1565,3 +1574,56 @@ ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std

    return result;
 }
+
+ggml_opt_optimizer_params common_opt_lr_pars(void * userdata) {
+    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
+    const lr_opt &            d      = *(lr_opt *) userdata;
+    result.adamw.alpha = result.sgd.alpha = d.get_lr(d.epoch);
+    result.sgd.wd = result.adamw.wd = d.wd;
+    return result;
+}
+
+// TODO make all command line args case-insensitive
+static inline bool eq_case_insensitive(char const* a, char const* b) {
+    return !
+#if defined(_MSC_VER)
+        _stricmp
+#else
+        strcasecmp
+#endif // defined(_MSC_VER)
+        (a, b);
+}
+
+enum ggml_opt_optimizer_type common_opt_get_optimizer(const char * n) {
+    if (eq_case_insensitive("adamw", n)) {
+        return GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    }
+    if (eq_case_insensitive("sgd", n)) {
+        return GGML_OPT_OPTIMIZER_TYPE_SGD;
+    }
+    return GGML_OPT_OPTIMIZER_TYPE_COUNT;
+}
+
+// TODO simplify to use just log and exp
+static float const k_log_2 = std::log(2.f);
+
+void lr_opt::init() {
+    if (lr_min > 0 && lr_min < lr0) {
+        float nhalf = std::log(lr0 / lr_min) / k_log_2;
+        float e     = epochs;
+        if (decay_epochs > 0 && decay_epochs < e) {
+            e = decay_epochs;
+        } else {
+            decay_epochs = e;
+        }
+        scale_epoch = nhalf / e;
+    }
+}
+
+float lr_opt::get_lr(float epoch) const {
+    float r = lr_min <= 0 ? lr0 :
+        epoch >= decay_epochs ? lr_min :
+        lr0 * std::pow(0.5f, epoch * scale_epoch);
+    LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
+    return r;
+}
--- a/llama/llama.cpp/common/common.h
+++ b/llama/llama.cpp/common/common.h
@@ -2,14 +2,17 @@

 #pragma once

-#include "llama-cpp.h"
-
 #include <set>
+#include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
 #include <sstream>
+#include <cmath>
+
+#include "ggml-opt.h"
+#include "llama-cpp.h"

 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -31,6 +34,9 @@ struct common_adapter_lora_info {
    std::string path;
    float scale;

+    std::string task_name;
+    std::string prompt_prefix;
+
    struct llama_adapter_lora * ptr;
 };

@@ -82,6 +88,7 @@ enum llama_example {
    LLAMA_EXAMPLE_PARALLEL,
    LLAMA_EXAMPLE_TTS,
    LLAMA_EXAMPLE_DIFFUSION,
+    LLAMA_EXAMPLE_FINETUNE,

    LLAMA_EXAMPLE_COUNT,
 };
@@ -186,10 +193,11 @@ struct common_params_sampling {
 };

 struct common_params_model {
-    std::string path    = ""; // model local path                                           // NOLINT
-    std::string url     = ""; // model url to download                                      // NOLINT
-    std::string hf_repo = ""; // HF repo                                                    // NOLINT
-    std::string hf_file = ""; // HF file                                                    // NOLINT
+    std::string path        = ""; // model local path                                       // NOLINT
+    std::string url         = ""; // model url to download                                  // NOLINT
+    std::string hf_repo     = ""; // HF repo                                                // NOLINT
+    std::string hf_file     = ""; // HF file                                                // NOLINT
+    std::string docker_repo = ""; // Docker repo                                            // NOLINT
 };

 struct common_params_speculative {
@@ -202,6 +210,7 @@ struct common_params_speculative {
    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
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;

    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
@@ -234,14 +243,36 @@ struct common_params_diffusion {
    bool    add_gumbel_noise = false; // add gumbel noise to the logits if temp > 0.0
 };

+// reasoning API response format (not to be confused as chat template's reasoning format)
 enum common_reasoning_format {
    COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_AUTO,
+    COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`
    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.
+    // do not extend this enum unless you absolutely have to
+    // in most cases, use COMMON_REASONING_FORMAT_AUTO
+    // see: https://github.com/ggml-org/llama.cpp/pull/15408
 };

+
+struct lr_opt {
+    float    lr0          = 1e-5; // learning rate at first epoch
+    float    lr_min       = -1;
+    float    decay_epochs = -1;   // if >0, the learning rate starts at lr0 and decays to lr_min after this many epochs
+    float    scale_epoch  = 0;
+    float    wd           = 0;
+    unsigned epochs       = 2;
+
+    unsigned epoch; // set by optimizer outer (epochs) loop
+    // learning rate decay - constant LR per epoch only for now
+    float get_lr(float e) const;
+    float get_lr() const { return get_lr(epoch); }
+    // must call after arg parse, before get_lr
+    void init();
+};
+
+struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
+
 struct common_params {
    int32_t n_predict             =    -1; // new tokens to predict
    int32_t n_ctx                 =  4096; // context size
@@ -257,11 +288,10 @@ struct common_params {
    float   rope_freq_base        =  0.0f; // RoPE base frequency
    float   rope_freq_scale       =  0.0f; // RoPE frequency scaling factor
    float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor      =  1.0f; // YaRN magnitude scaling factor
-    float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
+    float   yarn_attn_factor      = -1.0f; // YaRN magnitude scaling factor
+    float   yarn_beta_fast        = -1.0f; // YaRN low correction dim
+    float   yarn_beta_slow        = -1.0f; // YaRN high correction dim
    int32_t yarn_orig_ctx         =     0; // YaRN original context length
-    float   defrag_thold          =  0.1f; // KV cache defragmentation threshold

    // offload params
    std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
@@ -283,6 +313,7 @@ struct common_params {
    enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
    enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
    enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
+    enum llama_flash_attn_type   flash_attn_type   = LLAMA_FLASH_ATTN_TYPE_AUTO; // whether to use Flash Attention

    struct common_params_sampling    sampling;
    struct common_params_speculative speculative;
@@ -346,9 +377,8 @@ struct common_params {
    bool multiline_input   = false; // reverse the usage of `\`
    bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
    bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
-    bool flash_attn        = false; // flash attention
    bool no_perf           = false; // disable performance metrics
-    bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool ctx_shift         = false;  // context shift on infinite 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

@@ -376,6 +406,11 @@ struct common_params {
    bool no_mmproj = false;         // explicitly disable multimodal model
    std::vector<std::string> image; // path to image file(s)

+    // finetune
+    struct lr_opt lr;
+    enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    float val_split = 0.05f; // fraction of the data used for the validation set
+
    // embedding
    bool embedding         = false; // get only sentence embedding
    int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
@@ -384,11 +419,12 @@ struct common_params {
    std::string cls_sep    = "\t";  // separator of classification sequences

    // server params
-    int32_t port           = 8080;         // server listens on this network port
-    int32_t timeout_read   = 600;          // http read timeout in seconds
-    int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
-    int32_t n_cache_reuse  = 0;            // min chunk size to reuse from the cache via KV shifting
+    int32_t port              = 8080;         // server listens on this network port
+    int32_t timeout_read      = 600;          // http read timeout in seconds
+    int32_t timeout_write     = timeout_read; // http write timeout in seconds
+    int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
+    int32_t n_swa_checkpoints = 3;            // max number of SWA checkpoints per slot

    std::string hostname      = "127.0.0.1";
    std::string public_path   = "";                                                                         // NOLINT
@@ -409,7 +445,7 @@ struct common_params {

    // "advanced" endpoints are disabled by default for better security
    bool webui            = true;
-    bool endpoint_slots   = false;
+    bool endpoint_slots   = true;
    bool endpoint_props   = false; // only control POST requests, not GET
    bool endpoint_metrics = false;

@@ -417,7 +453,7 @@ struct common_params {

    std::string slot_save_path;

-    float slot_prompt_similarity = 0.5f;
+    float slot_prompt_similarity = 0.1f;

    // batched-bench params
    bool is_pp_shared = false;
@@ -698,8 +734,25 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

 }

+//
+// MoE utils
+//
+
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";
+
+static std::string llm_ffn_exps_block_regex(int idx) {
+    return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
+}
+
+static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
+    return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
+}
+
 //
 // training utils
 //

 ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
+
+// "adamw" or "sgd" (case insensitive)
+enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);
--- a/llama/llama.cpp/common/json-schema-to-grammar.cpp
+++ b/llama/llama.cpp/common/json-schema-to-grammar.cpp
@@ -257,12 +257,13 @@ std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
 };

 static bool is_reserved_name(const std::string & name) {
-    static std::unordered_set<std::string> RESERVED_NAMES;
-    if (RESERVED_NAMES.empty()) {
-        RESERVED_NAMES.insert("root");
-        for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
-        for (const auto &p : STRING_FORMAT_RULES) RESERVED_NAMES.insert(p.first);
-    }
+    static const std::unordered_set<std::string> RESERVED_NAMES = [] {
+        std::unordered_set<std::string> s;
+        s.insert("root");
+        for (const auto & p : PRIMITIVE_RULES) s.insert(p.first);
+        for (const auto & p : STRING_FORMAT_RULES) s.insert(p.first);
+        return s;
+    }();
    return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
 }

@@ -843,9 +844,10 @@ public:
                _build_object_rule(
                    properties, required, name,
                    schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
-        } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) {
+        } else if ((schema_type.is_null() || schema_type == "object" || schema_type == "string") && schema.contains("allOf")) {
            std::unordered_set<std::string> required;
            std::vector<std::pair<std::string, json>> properties;
+            std::map<std::string, size_t> enum_values;
            std::string hybrid_name = name;
            std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
                if (comp_schema.contains("$ref")) {
@@ -857,6 +859,14 @@ public:
                            required.insert(prop.key());
                        }
                    }
+                } else if (comp_schema.contains("enum")) {
+                    for (const auto & v : comp_schema["enum"]) {
+                        const auto rule = _generate_constant_rule(v);
+                        if (enum_values.find(rule) == enum_values.end()) {
+                            enum_values[rule] = 0;
+                        }
+                        enum_values[rule] += 1;
+                    }
                } else {
                  // todo warning
                }
@@ -870,6 +880,17 @@ public:
                    add_component(t, true);
                }
            }
+            if (!enum_values.empty()) {
+                std::vector<std::string> enum_intersection;
+                for (const auto & p : enum_values) {
+                    if (p.second == schema["allOf"].size()) {
+                        enum_intersection.push_back(p.first);
+                    }
+                }
+                if (!enum_intersection.empty()) {
+                    return _add_rule(rule_name, "(" + string_join(enum_intersection, " | ") + ") space");
+                }
+            }
            return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
        } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
            json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];
--- a/llama/llama.cpp/common/log.cpp
+++ b/llama/llama.cpp/common/log.cpp
@@ -4,17 +4,52 @@
 #include <condition_variable>
 #include <cstdarg>
 #include <cstdio>
+#include <cstdlib>
+#include <cstring>
 #include <mutex>
 #include <sstream>
 #include <thread>
 #include <vector>

+#if defined(_WIN32)
+#    include <io.h>
+#    include <windows.h>
+#    define isatty _isatty
+#    define fileno _fileno
+#else
+#    include <unistd.h>
+#endif // defined(_WIN32)
+
 int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;

 void common_log_set_verbosity_thold(int verbosity) {
    common_log_verbosity_thold = verbosity;
 }

+// Auto-detect if colors should be enabled based on terminal and environment
+static bool common_log_should_use_colors_auto() {
+    // Check NO_COLOR environment variable (https://no-color.org/)
+    if (const char * no_color = std::getenv("NO_COLOR")) {
+        if (no_color[0] != '\0') {
+            return false;
+        }
+    }
+
+    // Check TERM environment variable
+    if (const char * term = std::getenv("TERM")) {
+        if (std::strcmp(term, "dumb") == 0) {
+            return false;
+        }
+    }
+
+    // Check if stdout and stderr are connected to a terminal
+    // We check both because log messages can go to either
+    bool stdout_is_tty = isatty(fileno(stdout));
+    bool stderr_is_tty = isatty(fileno(stderr));
+
+    return stdout_is_tty || stderr_is_tty;
+}
+
 static int64_t t_us() {
    return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
 }
@@ -353,6 +388,11 @@ struct common_log * common_log_init() {

 struct common_log * common_log_main() {
    static struct common_log log;
+    static std::once_flag    init_flag;
+    std::call_once(init_flag, [&]() {
+        // Set default to auto-detect colors
+        log.set_colors(common_log_should_use_colors_auto());
+    });

    return &log;
 }
@@ -380,8 +420,19 @@ void common_log_set_file(struct common_log * log, const char * file) {
    log->set_file(file);
 }

-void common_log_set_colors(struct common_log * log, bool colors) {
-    log->set_colors(colors);
+void common_log_set_colors(struct common_log * log, log_colors colors) {
+    if (colors == LOG_COLORS_AUTO) {
+        log->set_colors(common_log_should_use_colors_auto());
+        return;
+    }
+
+    if (colors == LOG_COLORS_DISABLED) {
+        log->set_colors(false);
+        return;
+    }
+
+    GGML_ASSERT(colors == LOG_COLORS_ENABLED);
+    log->set_colors(true);
 }

 void common_log_set_prefix(struct common_log * log, bool prefix) {
--- a/llama/llama.cpp/common/log.h
+++ b/llama/llama.cpp/common/log.h
@@ -24,6 +24,12 @@
 #define LOG_DEFAULT_DEBUG 1
 #define LOG_DEFAULT_LLAMA 0

+enum log_colors {
+    LOG_COLORS_AUTO     = -1,
+    LOG_COLORS_DISABLED = 0,
+    LOG_COLORS_ENABLED  = 1,
+};
+
 // needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
 // set via common_log_set_verbosity()
 extern int common_log_verbosity_thold;
@@ -65,10 +71,10 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
 // D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
 //

-void common_log_set_file      (struct common_log * log, const char * file);       // not thread-safe
-void common_log_set_colors    (struct common_log * log,       bool   colors);     // not thread-safe
-void common_log_set_prefix    (struct common_log * log,       bool   prefix);     // whether to output prefix to each log
-void common_log_set_timestamps(struct common_log * log,       bool   timestamps); // whether to output timestamps in the prefix
+void common_log_set_file      (struct common_log * log, const char * file); // not thread-safe
+void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
+void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
+void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix

 // helper macros for logging
 // use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
--- a/llama/llama.cpp/common/sampling.cpp
+++ b/llama/llama.cpp/common/sampling.cpp
@@ -332,6 +332,7 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
    }
    if (ctx) {
        llama_perf_context_print(ctx);
+        llama_memory_breakdown_print(ctx);
    }
 }

@@ -426,8 +427,29 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {

 // helpers

-llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl) {
-    return &gsmpl->cur_p;
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
+    auto * res = &gsmpl->cur_p;
+
+    if (do_sort && !res->sorted) {
+        // remember the selected token before sorting
+        const llama_token id = res->data[res->selected].id;
+
+        std::sort(res->data, res->data + res->size, [](const llama_token_data & a, const llama_token_data & b) {
+            return a.p > b.p;
+        });
+
+        // restore the selected token after sorting
+        for (size_t i = 0; i < res->size; ++i) {
+            if (res->data[i].id == id) {
+                res->selected = i;
+                break;
+            }
+        }
+
+        res->sorted = true;
+    }
+
+    return res;
 }

 llama_token common_sampler_last(const struct common_sampler * gsmpl) {
--- a/llama/llama.cpp/common/sampling.h
+++ b/llama/llama.cpp/common/sampling.h
@@ -86,7 +86,9 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 // helpers

 // access the internal list of current candidate tokens
-llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl);
+// if do_sort == true, the candidates are guaranteed to be sorted afterwards (in descending order of probability)
+// the .sorted flag of the result indicates whether the returned candidates are sorted
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort);

 // get the last accepted token
 llama_token common_sampler_last(const struct common_sampler * gsmpl);
--- a/llama/llama.cpp/include/llama.h
+++ b/llama/llama.cpp/include/llama.h
@@ -64,8 +64,6 @@ extern "C" {

    typedef struct llama_memory_i * llama_memory_t;

-    struct llama_kv_cache; // DEPRECATED (use llama_memory instead)
-
    typedef int32_t llama_pos;
    typedef int32_t llama_token;
    typedef int32_t llama_seq_id;
@@ -181,6 +179,14 @@ extern "C" {
        LLAMA_ATTENTION_TYPE_NON_CAUSAL  = 1,
    };

+    enum llama_flash_attn_type {
+        LLAMA_FLASH_ATTN_TYPE_AUTO     = -1,
+        LLAMA_FLASH_ATTN_TYPE_DISABLED = 0,
+        LLAMA_FLASH_ATTN_TYPE_ENABLED  = 1,
+    };
+
+    LLAMA_API const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_type);
+
    enum llama_split_mode {
        LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
        LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
@@ -200,7 +206,7 @@ extern "C" {
        llama_token_data * data;
        size_t size;
        int64_t selected; // this is the index in the data array (i.e. not the token id)
-        bool sorted;
+        bool sorted;      // note: do not assume the data is sorted - always check this flag
    } llama_token_data_array;

    typedef bool (*llama_progress_callback)(float progress, void * user_data);
@@ -305,6 +311,7 @@ extern "C" {
        enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
        enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
        enum llama_attention_type    attention_type;    // attention type to use for embeddings
+        enum llama_flash_attn_type   flash_attn_type;   // when to enable Flash Attention

        // ref: https://github.com/ggml-org/llama.cpp/pull/2054
        float    rope_freq_base;   // RoPE base frequency, 0 = from model
@@ -314,7 +321,7 @@ extern "C" {
        float    yarn_beta_fast;   // YaRN low correction dim
        float    yarn_beta_slow;   // YaRN high correction dim
        uint32_t yarn_orig_ctx;    // YaRN original context size
-        float    defrag_thold;     // defragment the KV cache if holes/size > thold, <= 0 disabled (default)
+        float    defrag_thold;     // [DEPRECATED] defragment the KV cache if holes/size > thold, <= 0 disabled (default)

        ggml_backend_sched_eval_callback cb_eval;
        void * cb_eval_user_data;
@@ -331,7 +338,6 @@ extern "C" {
        // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
        bool embeddings;  // if true, extract embeddings (together with logits)
        bool offload_kqv; // offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // use flash attention [EXPERIMENTAL]
        bool no_perf;     // measure performance timings
        bool op_offload;  // offload host tensor operations to device
        bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
@@ -469,8 +475,6 @@ extern "C" {
    LLAMA_API           llama_memory_t   llama_get_memory  (const struct llama_context * ctx);
    LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type

-    DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead");
-
    LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
    LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);

@@ -557,10 +561,32 @@ extern "C" {
            struct llama_model * model,
            const char * path_lora);

+    // Functions to access the adapter's GGUF metadata scalar values
+    // - The functions return the length of the string on success, or -1 on failure
+    // - The output string is always null-terminated and cleared on failure
+    // - When retrieving a string, an extra byte must be allocated to account for the null terminator
+    // - GGUF array values are not supported by these functions
+
+    // Get metadata value as a string by key name
+    LLAMA_API int32_t llama_adapter_meta_val_str(const struct llama_adapter_lora * adapter, const char * key, char * buf, size_t buf_size);
+
+    // Get the number of metadata key/value pairs
+    LLAMA_API int32_t llama_adapter_meta_count(const struct llama_adapter_lora * adapter);
+
+    // Get metadata key name by index
+    LLAMA_API int32_t llama_adapter_meta_key_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size);
+
+    // Get metadata value as a string by index
+    LLAMA_API int32_t llama_adapter_meta_val_str_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size);
+
    // Manually free a LoRA adapter
    // Note: loaded adapters will be free when the associated model is deleted
    LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);

+    // Get the invocation tokens if the current lora is an alora
+    LLAMA_API uint64_t            llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter);
+    LLAMA_API const llama_token * llama_adapter_get_alora_invocation_tokens  (const struct llama_adapter_lora * adapter);
+
    // The following functions operate on a llama_context, hence the naming: llama_verb_...

    // Add a loaded LoRA adapter to given context
@@ -667,111 +693,6 @@ extern "C" {
    // Check if the memory supports shifting
    LLAMA_API bool llama_memory_can_shift(llama_memory_t mem);

-    //
-    // KV cache for self-attention (TODO: deprecate in favor of llama_memory)
-    //
-
-    // Returns the number of tokens in the KV cache (slow, use only for debug)
-    // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
-    DEPRECATED(LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx),
-               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
-
-    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    DEPRECATED(LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx),
-               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
-
-    // Clear the KV cache - both cell info is erased and KV data is zeroed
-    DEPRECATED(LLAMA_API void llama_kv_self_clear(
-                struct llama_context * ctx),
-            "Use llama_memory_clear() instead");
-
-    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
-    // seq_id < 0 : match any sequence
-    // p0 < 0     : [0,  p1]
-    // p1 < 0     : [p0, inf)
-    DEPRECATED(LLAMA_API bool llama_kv_self_seq_rm(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1),
-            "Use llama_memory_seq_rm() instead");
-
-    // Copy all tokens that belong to the specified sequence to another sequence
-    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_cp(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id_src,
-                    llama_seq_id   seq_id_dst,
-                       llama_pos   p0,
-                       llama_pos   p1),
-            "Use llama_memory_seq_cp() instead");
-
-    // Removes all tokens that do not belong to the specified sequence
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_keep(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_keep() instead");
-
-    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_add(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                       llama_pos   delta),
-            "Use llama_memory_seq_add() instead");
-
-    // Integer division of the positions by factor of `d > 1`
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_div(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                             int   d),
-            "Use llama_memory_seq_div() instead");
-
-    // Returns the smallest position present in the KV cache for the specified sequence
-    // This is typically non-zero only for SWA caches
-    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
-    // Return -1 if the sequence is empty
-    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_min(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_pos_min() instead");
-
-    // Returns the largest position present in the KV cache for the specified sequence
-    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
-    // Return -1 if the sequence is empty
-    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_max(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_pos_max() instead");
-
-    // Defragment the KV cache
-    // This will be applied:
-    //   - lazily on next llama_decode()
-    DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx),
-            "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");
-
-    // Check if the context supports KV cache shifting
-    DEPRECATED(LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx),
-            "use llama_memory_can_shift() instead");
-
-    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx),
-            "simply remove this call, updates are applied lazily on the next llama_decode()");
-
    //
    // State / sessions
    //
@@ -870,6 +791,29 @@ extern "C" {
                          size_t   n_token_capacity,
                          size_t * n_token_count_out);

+#define LLAMA_STATE_SEQ_FLAGS_SWA_ONLY 1
+
+    typedef uint32_t llama_state_seq_flags;
+
+    LLAMA_API size_t llama_state_seq_get_size_ext(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_get_data_ext(
+            struct llama_context * ctx,
+                         uint8_t * dst,
+                          size_t   size,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_set_data_ext(
+            struct llama_context * ctx,
+                   const uint8_t * src,
+                          size_t   size,
+                    llama_seq_id   dest_seq_id,
+           llama_state_seq_flags   flags);
+
    //
    // Decoding
    //
@@ -1216,11 +1160,6 @@ extern "C" {
    LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);

-    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
-    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
-        "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
-
    /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
    /// Setting k <= 0 makes this a noop
    LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1390,24 +1329,25 @@ extern "C" {
    //
    // Performance utils
    //
-    // NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements.
+    // NOTE: Used by llama.cpp examples/tools, avoid using in third-party apps. Instead, do your own performance measurements.
    //

    struct llama_perf_context_data {
-        double t_start_ms;
-        double t_load_ms;
-        double t_p_eval_ms;
-        double t_eval_ms;
+        // ms == milliseconds
+        double t_start_ms;  // absolute start time
+        double t_load_ms;   // time needed for loading the model
+        double t_p_eval_ms; // time needed for processing the prompt
+        double t_eval_ms;   // time needed for generating tokens

-        int32_t n_p_eval;
-        int32_t n_eval;
-        int32_t n_reused; // number of times a ggml compute graph had been reused
+        int32_t n_p_eval;   // number of prompt tokens
+        int32_t n_eval;     // number of generated tokens
+        int32_t n_reused;   // number of times a ggml compute graph had been reused
    };

    struct llama_perf_sampler_data {
-        double t_sample_ms;
+        double t_sample_ms; // time needed for sampling in ms

-        int32_t n_sample;
+        int32_t n_sample;   // number of sampled tokens
    };

    LLAMA_API struct llama_perf_context_data llama_perf_context      (const struct llama_context * ctx);
@@ -1419,6 +1359,9 @@ extern "C" {
    LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
    LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);

+    // print a breakdown of per-device memory use via LLAMA_LOG:
+    LLAMA_API void llama_memory_breakdown_print(const struct llama_context * ctx);
+
    //
    // training
    //
@@ -1437,6 +1380,8 @@ extern "C" {

        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+
+        enum ggml_opt_optimizer_type optimizer_type;
    };

    LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);
--- a/llama/llama.cpp/src/llama-adapter.cpp
+++ b/llama/llama.cpp/src/llama-adapter.cpp
@@ -6,6 +6,7 @@

 #include <map>
 #include <cassert>
+#include <sstream>
 #include <stdexcept>

 // vec
@@ -163,13 +164,38 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_

    // check metadata
    {
+        const gguf_context * gguf_ctx = ctx_gguf.get();
+
+        LLAMA_LOG_INFO("%s: Dumping metadata keys/values.\n", __func__);
+
+        // get metadata as string
+        for (int i = 0; i < gguf_get_n_kv(gguf_ctx); i++) {
+            gguf_type type = gguf_get_kv_type(gguf_ctx, i);
+            const std::string type_name =
+                type == GGUF_TYPE_ARRAY
+                ? format("%s[%s,%zu]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(gguf_ctx, i)), gguf_get_arr_n(gguf_ctx, i))
+                : gguf_type_name(type);
+            const char * name = gguf_get_key(gguf_ctx, i);
+            const std::string value = gguf_kv_to_str(gguf_ctx, i);
+
+            if (type != GGUF_TYPE_ARRAY) {
+                adapter.gguf_kv.emplace(name, value);
+            }
+
+            const size_t MAX_VALUE_LEN = 40;
+            std::string print_value = value.size() > MAX_VALUE_LEN ? format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()) : value;
+            replace_all(print_value, "\n", "\\n");
+
+            LLAMA_LOG_INFO("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), print_value.c_str());
+        }
+
        auto get_kv_str = [&](const std::string & key) -> std::string {
-            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
-            return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf.get(), id));
+            int id = gguf_find_key(gguf_ctx, key.c_str());
+            return id < 0 ? "" : std::string(gguf_get_val_str(gguf_ctx, id));
        };
        auto get_kv_f32 = [&](const std::string & key) -> float {
-            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
-            return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf.get(), id);
+            int id = gguf_find_key(gguf_ctx, key.c_str());
+            return id < 0 ? 0.0f : gguf_get_val_f32(gguf_ctx, id);
        };
        LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);

@@ -190,6 +216,26 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
        }

        adapter.alpha = get_kv_f32(llm_kv(LLM_KV_ADAPTER_LORA_ALPHA));
+
+        // parse alora invocation sequence vector
+        const auto & key = llm_kv(LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS);
+        const int kid = gguf_find_key(ctx_gguf.get(), key.c_str());
+        if (kid >= 0) {
+            if (gguf_get_kv_type(ctx_gguf.get(), kid) != GGUF_TYPE_ARRAY) {
+                throw std::runtime_error("invalid gguf type for " + key);
+            }
+            const auto arr_type = gguf_get_arr_type(ctx_gguf.get(), kid);
+            if (arr_type != GGUF_TYPE_UINT32) {
+                throw std::runtime_error("invalid gguf element type for " + key);
+            }
+            const size_t seq_len = gguf_get_arr_n(ctx_gguf.get(), kid);
+            const void * data = gguf_get_arr_data(ctx_gguf.get(), kid);
+            adapter.alora_invocation_tokens.resize(seq_len);
+            std::copy(
+                (const llama_token *)data,
+                (const llama_token *)data + seq_len,
+                adapter.alora_invocation_tokens.begin());
+        }
    }

    int n_tensors = gguf_get_n_tensors(ctx_gguf.get());
@@ -383,6 +429,57 @@ llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * p
    return nullptr;
 }

+int32_t llama_adapter_meta_val_str(const llama_adapter_lora * adapter, const char * key, char * buf, size_t buf_size) {
+    const auto & it = adapter->gguf_kv.find(key);
+    if (it == adapter->gguf_kv.end()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    return snprintf(buf, buf_size, "%s", it->second.c_str());
+}
+
+int32_t llama_adapter_meta_count(const llama_adapter_lora * adapter) {
+    return (int)adapter->gguf_kv.size();
+}
+
+int32_t llama_adapter_meta_key_by_index(const llama_adapter_lora * adapter, int i, char * buf, size_t buf_size) {
+    if (i < 0 || i >= (int)adapter->gguf_kv.size()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    auto it = adapter->gguf_kv.begin();
+    std::advance(it, i);
+    return snprintf(buf, buf_size, "%s", it->first.c_str());
+}
+
+int32_t llama_adapter_meta_val_str_by_index(const llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size) {
+    if (i < 0 || i >= (int)adapter->gguf_kv.size()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    auto it = adapter->gguf_kv.begin();
+    std::advance(it, i);
+    return snprintf(buf, buf_size, "%s", it->second.c_str());
+}
+
 void llama_adapter_lora_free(llama_adapter_lora * adapter) {
    delete adapter;
 }
+
+uint64_t llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter) {
+    if (!adapter) {
+        return 0;
+    }
+    return adapter->alora_invocation_tokens.size();
+}
+
+const llama_token * llama_adapter_get_alora_invocation_tokens(const llama_adapter_lora * adapter) {
+    GGML_ASSERT(adapter);
+    return adapter->alora_invocation_tokens.data();
+}
--- a/llama/llama.cpp/src/llama-adapter.h
+++ b/llama/llama.cpp/src/llama-adapter.h
@@ -67,6 +67,12 @@ struct llama_adapter_lora {

    float alpha;

+    // gguf metadata
+    std::unordered_map<std::string, std::string> gguf_kv;
+
+    // activated lora (aLoRA)
+    std::vector<llama_token> alora_invocation_tokens;
+
    llama_adapter_lora() = default;
    ~llama_adapter_lora() = default;

--- a/llama/llama.cpp/src/llama-arch.cpp
+++ b/llama/llama.cpp/src/llama-arch.cpp
@@ -22,6 +22,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
    { LLM_ARCH_NEO_BERT,         "neo-bert"         },
    { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
+    { LLM_ARCH_JINA_BERT_V3,     "jina-bert-v3"     },
    { LLM_ARCH_BLOOM,            "bloom"            },
    { LLM_ARCH_STABLELM,         "stablelm"         },
    { LLM_ARCH_QWEN,             "qwen"             },
@@ -44,6 +45,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_GEMMA2,           "gemma2"           },
    { LLM_ARCH_GEMMA3,           "gemma3"           },
    { LLM_ARCH_GEMMA3N,          "gemma3n"          },
+    { LLM_ARCH_GEMMA_EMBEDDING,  "gemma-embedding"  },
    { LLM_ARCH_STARCODER2,       "starcoder2"       },
    { LLM_ARCH_MAMBA,            "mamba"            },
    { LLM_ARCH_MAMBA2,           "mamba2"           },
@@ -68,6 +70,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_T5ENCODER,        "t5encoder"        },
    { LLM_ARCH_JAIS,             "jais"             },
    { LLM_ARCH_NEMOTRON,         "nemotron"         },
+    { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
    { LLM_ARCH_EXAONE,           "exaone"           },
    { LLM_ARCH_EXAONE4,          "exaone4"          },
    { LLM_ARCH_RWKV6,            "rwkv6"            },
@@ -94,6 +97,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_DREAM,            "dream"            },
    { LLM_ARCH_SMALLTHINKER,     "smallthinker"     },
    { LLM_ARCH_LLADA,            "llada"            },
+    { LLM_ARCH_LLADA_MOE,        "llada-moe"        },
+    { LLM_ARCH_SEED_OSS,         "seed_oss"         },
+    { LLM_ARCH_GROVEMOE,         "grovemoe"         },
    { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };

@@ -121,6 +127,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_FEED_FORWARD_LENGTH,               "%s.feed_forward_length"               },
    { LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        "%s.expert_feed_forward_length"        },
    { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
+    { LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,  "%s.expert_chunk_feed_forward_length"  },
    { LLM_KV_USE_PARALLEL_RESIDUAL,             "%s.use_parallel_residual"             },
    { LLM_KV_TENSOR_DATA_LAYOUT,                "%s.tensor_data_layout"                },
    { LLM_KV_EXPERT_COUNT,                      "%s.expert_count"                      },
@@ -129,12 +136,16 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
    { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
    { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
+    { LLM_KV_EXPERT_GROUP_SCALE,                "%s.expert_group_scale"                },
+    { LLM_KV_EXPERTS_PER_GROUP,                 "%s.experts_per_group"                 },
    { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
    { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
    { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
    { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
    { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
+    { LLM_KV_DECODER_BLOCK_COUNT,               "%s.decoder_block_count"               },
    { LLM_KV_ATTN_LOGIT_SOFTCAPPING,            "%s.attn_logit_softcapping"            },
+    { LLM_KV_ROUTER_LOGIT_SOFTCAPPING,          "%s.router_logit_softcapping"          },
    { LLM_KV_FINAL_LOGIT_SOFTCAPPING,           "%s.final_logit_softcapping"           },
    { LLM_KV_SWIN_NORM,                         "%s.swin_norm"                         },
    { LLM_KV_RESCALE_EVERY_N_LAYERS,            "%s.rescale_every_n_layers"            },
@@ -165,20 +176,26 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
    { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
    { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
+    { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
+    { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
    { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
    { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },

-    { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
-    { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
-    { LLM_KV_ROPE_FREQ_BASE,            "%s.rope.freq_base"                       },
-    { LLM_KV_ROPE_SCALE_LINEAR,         "%s.rope.scale_linear"                    },
-    { LLM_KV_ROPE_SCALING_TYPE,         "%s.rope.scaling.type"                    },
-    { LLM_KV_ROPE_SCALING_FACTOR,       "%s.rope.scaling.factor"                  },
-    { LLM_KV_ROPE_SCALING_ATTN_FACTOR,  "%s.rope.scaling.attn_factor"             },
-    { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
-    { LLM_KV_ROPE_SCALING_FINETUNED,    "%s.rope.scaling.finetuned"               },
-    { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier"     },
+    { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
+    { LLM_KV_ROPE_DIMENSION_SECTIONS,       "%s.rope.dimension_sections"              },
+    { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+    { LLM_KV_ROPE_SCALE_LINEAR,             "%s.rope.scale_linear"                    },
+    { LLM_KV_ROPE_SCALING_TYPE,             "%s.rope.scaling.type"                    },
+    { LLM_KV_ROPE_SCALING_FACTOR,           "%s.rope.scaling.factor"                  },
+    { LLM_KV_ROPE_SCALING_ATTN_FACTOR,      "%s.rope.scaling.attn_factor"             },
+    { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,     "%s.rope.scaling.original_context_length" },
+    { LLM_KV_ROPE_SCALING_FINETUNED,        "%s.rope.scaling.finetuned"               },
+    { LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     "%s.rope.scaling.yarn_log_multiplier"     },
+    { LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,  "%s.rope.scaling.yarn_ext_factor"         },
+    { LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor"        },
+    { LLM_KV_ROPE_SCALING_YARN_BETA_FAST,   "%s.rope.scaling.yarn_beta_fast"          },
+    { LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,   "%s.rope.scaling.yarn_beta_slow"          },

    { LLM_KV_SPLIT_NO,            "split.no"            },
    { LLM_KV_SPLIT_COUNT,         "split.count"         },
@@ -235,8 +252,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_TOKENIZER_FIM_REP_ID,           "tokenizer.ggml.fim_rep_token_id"         },
    { LLM_KV_TOKENIZER_FIM_SEP_ID,           "tokenizer.ggml.fim_sep_token_id"         },

-    { LLM_KV_ADAPTER_TYPE,       "adapter.type"       },
-    { LLM_KV_ADAPTER_LORA_ALPHA, "adapter.lora.alpha" },
+    { LLM_KV_ADAPTER_TYPE,                    "adapter.type"               },
+    { LLM_KV_ADAPTER_LORA_ALPHA,              "adapter.lora.alpha"         },
+    { LLM_KV_ADAPTER_LORA_TASK_NAME,          "adapter.lora.task_name"     },
+    { LLM_KV_ADAPTER_LORA_PROMPT_PREFIX,      "adapter.lora.prompt_prefix" },
+    { LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS, "adapter.alora.invocation_tokens" },

    // deprecated
    { LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" },
@@ -392,12 +412,16 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
        },
@@ -576,6 +600,20 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_CLS,             "cls" },
        },
    },
+    {
+        LLM_ARCH_JINA_BERT_V3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
+        },
+    },
    {
        LLM_ARCH_BLOOM,
        {
@@ -689,6 +727,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_CLS_OUT,         "cls.output" },
            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
@@ -1021,6 +1060,27 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_LAUREL_POST_NORM,     "blk.%d.laurel_post_norm" },
        },
    },
+    {
+        LLM_ARCH_GEMMA_EMBEDDING,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
    {
        LLM_ARCH_STARCODER2,
        {
@@ -1534,6 +1594,31 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
        },
    },
+    {
+        LLM_ARCH_NEMOTRON_H,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            // mamba(2) ssm layers
+            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
+            // attention layers
+            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
+            // dense FFN
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+        },
+    },
    {
        LLM_ARCH_EXAONE,
        {
@@ -2030,6 +2115,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
            { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT,            "output" },
        }
    },
    {
@@ -2087,6 +2173,66 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
        },
    },
+    {
+        LLM_ARCH_LLADA_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
+    {
+        LLM_ARCH_SEED_OSS,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_GROVEMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_CHEXPS,    "blk.%d.ffn_gate_chexps" },
+            { LLM_TENSOR_FFN_DOWN_CHEXPS,    "blk.%d.ffn_down_chexps" },
+            { LLM_TENSOR_FFN_UP_CHEXPS,      "blk.%d.ffn_up_chexps" },
+        },
+    },
    {
        LLM_ARCH_UNKNOWN,
        {
@@ -2219,6 +2365,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
    {LLM_TENSOR_FFN_DOWN_EXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
    {LLM_TENSOR_FFN_GATE_EXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
    {LLM_TENSOR_FFN_UP_EXPS,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
+    {LLM_TENSOR_FFN_DOWN_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
+    {LLM_TENSOR_FFN_GATE_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
+    {LLM_TENSOR_FFN_UP_CHEXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
    {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
    // altup / laurel (gemma 3n)
    {LLM_TENSOR_PER_LAYER_TOKEN_EMBD,       {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_GET_ROWS}},
@@ -2340,6 +2489,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
        case LLM_ARCH_PLAMO2:
        case LLM_ARCH_GRANITE_HYBRID:
        case LLM_ARCH_LFM2:
+        case LLM_ARCH_NEMOTRON_H:
            return true;
        default:
            return false;
@@ -2350,6 +2500,7 @@ bool llm_arch_is_diffusion(const llm_arch & arch) {
    switch (arch) {
        case LLM_ARCH_DREAM:
        case LLM_ARCH_LLADA:
+        case LLM_ARCH_LLADA_MOE:
            return true;
        default:
            return false;
--- a/llama/llama.cpp/src/llama-arch.h
+++ b/llama/llama.cpp/src/llama-arch.h
@@ -26,6 +26,7 @@ enum llm_arch {
    LLM_ARCH_NOMIC_BERT_MOE,
    LLM_ARCH_NEO_BERT,
    LLM_ARCH_JINA_BERT_V2,
+    LLM_ARCH_JINA_BERT_V3,
    LLM_ARCH_BLOOM,
    LLM_ARCH_STABLELM,
    LLM_ARCH_QWEN,
@@ -48,6 +49,7 @@ enum llm_arch {
    LLM_ARCH_GEMMA2,
    LLM_ARCH_GEMMA3,
    LLM_ARCH_GEMMA3N,
+    LLM_ARCH_GEMMA_EMBEDDING,
    LLM_ARCH_STARCODER2,
    LLM_ARCH_MAMBA,
    LLM_ARCH_MAMBA2,
@@ -72,6 +74,7 @@ enum llm_arch {
    LLM_ARCH_T5ENCODER,
    LLM_ARCH_JAIS,
    LLM_ARCH_NEMOTRON,
+    LLM_ARCH_NEMOTRON_H,
    LLM_ARCH_EXAONE,
    LLM_ARCH_EXAONE4,
    LLM_ARCH_RWKV6,
@@ -98,6 +101,9 @@ enum llm_arch {
    LLM_ARCH_DREAM,
    LLM_ARCH_SMALLTHINKER,
    LLM_ARCH_LLADA,
+    LLM_ARCH_LLADA_MOE,
+    LLM_ARCH_SEED_OSS,
+    LLM_ARCH_GROVEMOE,
    LLM_ARCH_UNKNOWN,
 };

@@ -125,6 +131,7 @@ enum llm_kv {
    LLM_KV_FEED_FORWARD_LENGTH,
    LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
    LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,
    LLM_KV_USE_PARALLEL_RESIDUAL,
    LLM_KV_TENSOR_DATA_LAYOUT,
    LLM_KV_EXPERT_COUNT,
@@ -133,12 +140,16 @@ enum llm_kv {
    LLM_KV_EXPERT_WEIGHTS_SCALE,
    LLM_KV_EXPERT_WEIGHTS_NORM,
    LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_EXPERT_GROUP_SCALE,
+    LLM_KV_EXPERTS_PER_GROUP,
    LLM_KV_MOE_EVERY_N_LAYERS,
    LLM_KV_NEXTN_PREDICT_LAYERS,
    LLM_KV_POOLING_TYPE,
    LLM_KV_LOGIT_SCALE,
    LLM_KV_DECODER_START_TOKEN_ID,
+    LLM_KV_DECODER_BLOCK_COUNT,
    LLM_KV_ATTN_LOGIT_SOFTCAPPING,
+    LLM_KV_ROUTER_LOGIT_SOFTCAPPING,
    LLM_KV_FINAL_LOGIT_SOFTCAPPING,
    LLM_KV_SWIN_NORM,
    LLM_KV_RESCALE_EVERY_N_LAYERS,
@@ -169,6 +180,8 @@ enum llm_kv {
    LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
    LLM_KV_ATTENTION_SLIDING_WINDOW,
    LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_OUTPUT_SCALE,
+    LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
    LLM_KV_ATTENTION_KEY_LENGTH_MLA,
    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
@@ -183,6 +196,10 @@ enum llm_kv {
    LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
    LLM_KV_ROPE_SCALING_FINETUNED,
    LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
+    LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,
+    LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR,
+    LLM_KV_ROPE_SCALING_YARN_BETA_FAST,
+    LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,

    LLM_KV_SPLIT_NO,
    LLM_KV_SPLIT_COUNT,
@@ -231,6 +248,9 @@ enum llm_kv {

    LLM_KV_ADAPTER_TYPE,
    LLM_KV_ADAPTER_LORA_ALPHA,
+    LLM_KV_ADAPTER_LORA_TASK_NAME,
+    LLM_KV_ADAPTER_LORA_PROMPT_PREFIX,
+    LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS,

    LLM_KV_POSNET_EMBEDDING_LENGTH,
    LLM_KV_POSNET_BLOCK_COUNT,
@@ -287,6 +307,9 @@ enum llm_tensor {
    LLM_TENSOR_FFN_DOWN_SHEXP,
    LLM_TENSOR_FFN_GATE_SHEXP,
    LLM_TENSOR_FFN_UP_SHEXP,
+    LLM_TENSOR_FFN_DOWN_CHEXPS,
+    LLM_TENSOR_FFN_GATE_CHEXPS,
+    LLM_TENSOR_FFN_UP_CHEXPS,
    LLM_TENSOR_FFN_EXP_PROBS_B,
    LLM_TENSOR_ATTN_Q_NORM,
    LLM_TENSOR_ATTN_K_NORM,
--- a/llama/llama.cpp/src/llama-batch.cpp
+++ b/llama/llama.cpp/src/llama-batch.cpp
@@ -477,7 +477,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {

 llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
    if (sequential && has_cpl) {
-        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag)\n", __func__);

        return {};
    }
--- a/llama/llama.cpp/src/llama-chat.cpp
+++ b/llama/llama.cpp/src/llama-chat.cpp
@@ -16,10 +16,10 @@
 static std::string trim(const std::string & str) {
    size_t start = 0;
    size_t end = str.size();
-    while (start < end && isspace(str[start])) {
+    while (start < end && isspace(static_cast<unsigned char>(str[start]))) {
        start += 1;
    }
-    while (end > start && isspace(str[end - 1])) {
+    while (end > start && isspace(static_cast<unsigned char>(str[end - 1]))) {
        end -= 1;
    }
    return str.substr(start, end - start);
@@ -69,6 +69,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
    { "gpt-oss",           LLM_CHAT_TEMPLATE_OPENAI_MOE        },
    { "hunyuan-dense",     LLM_CHAT_TEMPLATE_HUNYUAN_DENSE     },
    { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
+    { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
+    { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
 };

 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -201,6 +203,10 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
        return LLM_CHAT_TEMPLATE_HUNYUAN_DENSE;
    } else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) {
        return LLM_CHAT_TEMPLATE_KIMI_K2;
+    } else if (tmpl_contains("<seed:bos>")) {
+        return LLM_CHAT_TEMPLATE_SEED_OSS;
+    } else if (tmpl_contains("'Assistant: '  + message['content'] + '<|separator|>")) {
+        return LLM_CHAT_TEMPLATE_GROK_2;
    }
    return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -752,6 +758,28 @@ int32_t llm_chat_apply_template(
        if (add_ass) {
            ss << "<|im_assistant|>assistant<|im_middle|>";
        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SEED_OSS) {
+        for (auto message: chat) {
+            std::string role(message->role);
+            ss << "<seed:bos>" << role << "\n" << (role == "assistant" ? trim(message->content) : message->content) << "<seed:eos>";
+        }
+        if (add_ass) {
+            ss << "<seed:bos>assistant\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GROK_2) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "System: " << trim(message->content) << "<|separator|>\n\n";
+            } else if (role == "user") {
+                ss << "Human: " << trim(message->content) << "<|separator|>\n\n";
+            } else if (role == "assistant") {
+                ss << "Assistant: " << message->content << "<|separator|>\n\n";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
    } else {
        // template not supported
        return -1;
--- a/llama/llama.cpp/src/llama-chat.h
+++ b/llama/llama.cpp/src/llama-chat.h
@@ -49,6 +49,8 @@ enum llm_chat_template {
    LLM_CHAT_TEMPLATE_OPENAI_MOE,
    LLM_CHAT_TEMPLATE_HUNYUAN_DENSE,
    LLM_CHAT_TEMPLATE_KIMI_K2,
+    LLM_CHAT_TEMPLATE_SEED_OSS,
+    LLM_CHAT_TEMPLATE_GROK_2,
    LLM_CHAT_TEMPLATE_UNKNOWN,
 };

--- a/llama/llama.cpp/src/llama-context.cpp
+++ b/llama/llama.cpp/src/llama-context.cpp
@@ -35,14 +35,12 @@ llama_context::llama_context(

    cparams.n_threads        = params.n_threads;
    cparams.n_threads_batch  = params.n_threads_batch;
-    cparams.yarn_ext_factor  = params.yarn_ext_factor;
-    cparams.yarn_attn_factor = params.yarn_attn_factor;
-    cparams.yarn_beta_fast   = params.yarn_beta_fast;
-    cparams.yarn_beta_slow   = params.yarn_beta_slow;
-    cparams.defrag_thold     = params.defrag_thold;
+    cparams.yarn_ext_factor  = params.yarn_ext_factor  >= 0.0f ? params.yarn_ext_factor  : hparams.yarn_ext_factor;
+    cparams.yarn_attn_factor = params.yarn_attn_factor >= 0.0f ? params.yarn_attn_factor : hparams.yarn_attn_factor;
+    cparams.yarn_beta_fast   = params.yarn_beta_fast   >= 0.0f ? params.yarn_beta_fast   : hparams.yarn_beta_fast;
+    cparams.yarn_beta_slow   = params.yarn_beta_slow   >= 0.0f ? params.yarn_beta_slow   : hparams.yarn_beta_slow;
    cparams.embeddings       = params.embeddings;
    cparams.offload_kqv      = params.offload_kqv;
-    cparams.flash_attn       = params.flash_attn;
    cparams.no_perf          = params.no_perf;
    cparams.pooling_type     = params.pooling_type;
    cparams.warmup           = false;
@@ -87,13 +85,15 @@ llama_context::llama_context(
        cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL;
    }

+    cparams.flash_attn = params.flash_attn_type != LLAMA_FLASH_ATTN_TYPE_DISABLED;
+
    // with causal attention, the batch size is limited by the context size
    cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;

    // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
    // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
    // ref: https://github.com/ggerganov/llama.cpp/pull/5021
-    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_context_kv_self
+    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_memory
    if (cparams.n_batch < GGML_KQ_MASK_PAD) {
        LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
        cparams.n_batch = GGML_KQ_MASK_PAD;
@@ -103,16 +103,6 @@ llama_context::llama_context(
    cparams.op_offload = params.op_offload;
    cparams.kv_unified = params.kv_unified;

-    {
-        const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-        supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : supports_set_rows;
-
-        if (!supports_set_rows && !cparams.kv_unified) {
-            LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache\n", __func__);
-            cparams.kv_unified = true;
-        }
-    }
-
    {
        const char * LLAMA_GRAPH_REUSE_DISABLE = getenv("LLAMA_GRAPH_REUSE_DISABLE");
        graph_reuse_disable = LLAMA_GRAPH_REUSE_DISABLE ? (atoi(LLAMA_GRAPH_REUSE_DISABLE) != 0) : graph_reuse_disable;
@@ -130,7 +120,7 @@ llama_context::llama_context(
    LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
    LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
    LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
-    LLAMA_LOG_INFO("%s: flash_attn    = %d\n",   __func__, cparams.flash_attn);
+    LLAMA_LOG_INFO("%s: flash_attn    = %s\n",   __func__, llama_flash_attn_type_name(params.flash_attn_type));
    LLAMA_LOG_INFO("%s: kv_unified    = %s\n",   __func__, cparams.kv_unified ? "true" : "false");
    LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
    LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
@@ -145,11 +135,6 @@ llama_context::llama_context(
                __func__, n_ctx_per_seq, hparams.n_ctx_train);
    }

-    if (!params.swa_full && cparams.n_seq_max > 1 && hparams.is_swa_any()) {
-        LLAMA_LOG_WARN("%s: requested n_seq_max (%u) > 1, but swa_full is not enabled -- performance may be degraded: %s\n",
-                __func__, cparams.n_seq_max, "https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573");
-    }
-
    if (!hparams.vocab_only) {
        // GPU backends
        for (auto * dev : model.devices) {
@@ -196,7 +181,7 @@ llama_context::llama_context(
        // graph outputs buffer
        {
            // resized during inference when a batch uses more outputs
-            if ((uint32_t) output_reserve(params.n_seq_max) < params.n_seq_max) {
+            if (output_reserve(params.n_seq_max) < params.n_seq_max) {
                throw std::runtime_error("failed to reserve initial output buffer");
            }

@@ -285,28 +270,75 @@ llama_context::llama_context(
        }
    }

-    // reserve worst-case graph
-    if (!hparams.vocab_only && memory) {
+    if (!hparams.vocab_only) {
+        llama_memory_context_ptr mctx;
+        if (memory) {
+            LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
+            mctx = memory->init_full();
+            if (!mctx) {
+                throw std::runtime_error("failed to initialize memory module");
+            }
+        }
+
+        cross.v_embd.clear();
+
        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);

+        // avoid reserving graphs with zero outputs - assume one output per sequence
+        n_outputs = n_seqs;
+
        LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);

+        // resolve automatic Flash Attention use
+        if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO) {
+            auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
+            if (!gf) {
+                throw std::runtime_error("failed to split graph for Flash Attention check");
+            }
+
+            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
+            bool fa_device_mismatch = false;
+            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
+                ggml_tensor * n = ggml_graph_node(gf, i);
+                if (n->op != GGML_OP_FLASH_ATTN_EXT) {
+                    continue;
+                }
+                ggml_backend_dev_t device_fa = ggml_backend_get_device(
+                    ggml_backend_sched_get_tensor_backend(sched.get(), n));
+
+                // TODO: instead of the tensor names, use a map to keep track of which (FA) tensors belong to which layer
+                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FATTN "-", prefix_len) == 0);
+                const int il = std::stoi(n->name + prefix_len);
+                ggml_backend_dev_t device_kv = model.dev_layer(il);
+                if (device_fa != device_kv) {
+                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the Flash Attention tensor "
+                        "is assigned to device %s (usually due to missing support)\n",
+                        __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_fa));
+                    // FIXME: fa_device_mismatch logic is wrong for --no-kv-offload, but this is broken anyways
+                    fa_device_mismatch = true;
+                    break;
+                }
+            }
+            if (fa_device_mismatch) {
+                cparams.flash_attn = false;
+                LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
+                if (ggml_is_quantized(params.type_v)) {
+                    throw std::runtime_error("quantized V cache was requested, but this requires Flash Attention");
+                }
+            } else {
+                cparams.flash_attn = true;
+                LLAMA_LOG_INFO("%s: Flash Attention was auto, set to enabled\n", __func__);
+            }
+        }
+
+        // reserve worst-case graph
        int n_splits_pp = -1;
        int n_nodes_pp  = -1;

        int n_splits_tg = -1;
        int n_nodes_tg  = -1;

-        // simulate full KV cache
-
-        const auto mctx = memory->init_full();
-        if (!mctx) {
-            throw std::runtime_error("failed to initialize KV cache");
-        }
-
-        cross.v_embd.clear();
-
        // reserve pp (prompt processing) graph first so that buffers are only allocated once
        {
            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -444,26 +476,12 @@ llama_memory_t llama_context::get_memory() const {
    return memory.get();
 }

-// deprecated
-void llama_context::kv_self_defrag_sched() {
-    if (!memory) {
-        return;
-    }
-
-    memory_force_optimize = true;
-}
-
-// deprecated
-bool llama_context::kv_self_update(bool optimize) {
+bool llama_context::memory_update(bool optimize) {
    if (!memory) {
        return false;
    }

    {
-        // TODO: remove in the future
-        optimize |= memory_force_optimize;
-        memory_force_optimize = false;
-
        const auto mctx = memory->init_update(this, optimize);
        switch (mctx->get_status()) {
            case LLAMA_MEMORY_STATUS_SUCCESS:
@@ -908,12 +926,6 @@ int llama_context::encode(const llama_batch & batch_inp) {
        }
    }

-    if (!supports_set_rows) {
-        // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-        // overlap with device computation.
-        ggml_backend_sched_reset(sched.get());
-    }
-
    // TODO: hacky solution
    if (model.arch == LLM_ARCH_T5 && t_embd) {
        //cross.t_embd = t_embd;
@@ -962,8 +974,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
    const int64_t n_vocab = vocab.n_tokens();
    const int64_t n_embd  = hparams.n_embd;

-    // when computing embeddings, all tokens are output
-    const bool output_all = cparams.embeddings;
+    const bool output_all = false;

    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, output_all)) {
        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
@@ -997,8 +1008,8 @@ int llama_context::decode(const llama_batch & batch_inp) {

    bool did_optimize = false;

-    // handle any pending defrags/shifts
-    kv_self_update(false);
+    // handle any pending shifts/copies
+    memory_update(false);

    llama_memory_context_ptr mctx;

@@ -1023,7 +1034,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
                    if (!did_optimize) {
                        did_optimize = true;

-                        if (kv_self_update(true)) {
+                        if (memory_update(true)) {
                            LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, balloc->get_n_tokens());

                            continue;
@@ -1076,7 +1087,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
        const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);

        if (!res) {
-            // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
+            // the last ubatch failed or was aborted -> remove all positions of that ubatch from the memory module
            llama_pos pos_min[LLAMA_MAX_SEQ];
            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
                pos_min[s] = std::numeric_limits<llama_pos>::max();
@@ -1093,7 +1104,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
                    continue;
                }

-                LLAMA_LOG_WARN("%s: removing KV cache entries for seq_id = %d, pos = [%d, +inf)\n", __func__, s, pos_min[s]);
+                LLAMA_LOG_WARN("%s: removing memory module entries for seq_id = %d, pos = [%d, +inf)\n", __func__, s, pos_min[s]);

                memory->seq_rm(s, pos_min[s], -1);
            }
@@ -1244,12 +1255,6 @@ int llama_context::decode(const llama_batch & batch_inp) {
    // wait for the computation to finish (automatically done when obtaining the model output)
    //synchronize();

-    if (!supports_set_rows) {
-        // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-        // overlap with device computation.
-        ggml_backend_sched_reset(sched.get());
-    }
-
    return 0;
 }

@@ -1363,8 +1368,9 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
    return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }

-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) {
+ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
    LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
+    GGML_ASSERT(n_outputs >= 1);

    if (n_tokens % n_seqs != 0) {
        n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
@@ -1398,7 +1404,9 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
    this->n_outputs = save_n_outputs;

    // initialize scheduler with the specified graph
-    if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+    if (split_only) {
+        ggml_backend_sched_split_graph(sched.get(), gf);
+    } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
        LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
        return nullptr;
    }
@@ -1438,7 +1446,9 @@ ggml_status llama_context::graph_compute(
    if (backend_cpu != nullptr) {
        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu));
        auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
-        set_threadpool_fn(backend_cpu, tp);
+        if (set_threadpool_fn) {
+            set_threadpool_fn(backend_cpu, tp);
+        }
    }

    // set the number of threads for all the backends
@@ -1657,30 +1667,30 @@ size_t llama_context::state_set_data(const uint8_t * src, size_t size) {
    }
 }

-size_t llama_context::state_seq_get_size(llama_seq_id seq_id) {
+size_t llama_context::state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags) {
    llama_io_write_dummy io;
    try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
        return 0;
    }
 }

-size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size) {
+size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size, llama_state_seq_flags flags) {
    llama_io_write_buffer io(dst, size);
    try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
        return 0;
    }
 }

-size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size) {
+size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags) {
    llama_io_read_buffer io(src, size);
    try {
-        return state_seq_read_data(io, seq_id);
+        return state_seq_read_data(io, seq_id, flags);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
        return 0;
@@ -1778,7 +1788,7 @@ size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * file
    {
        const size_t state_size = file.size() - file.tell();
        llama_io_read_file io(&file);
-        const size_t nread = state_seq_read_data(io, seq_id);
+        const size_t nread = state_seq_read_data(io, seq_id, 0);
        if (!nread) {
            LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
            return 0;
@@ -1802,7 +1812,7 @@ size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * file

    // save the context state using stream saving
    llama_io_write_file io(&file);
-    state_seq_write_data(io, seq_id);
+    state_seq_write_data(io, seq_id, 0);

    const size_t res = file.tell();
    GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + io.n_bytes());
@@ -1877,7 +1887,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
    }

    if (memory != nullptr) {
-        LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__);
+        LLAMA_LOG_DEBUG("%s: - writing memory module\n", __func__);
        memory->state_write(io);
    }

@@ -1963,7 +1973,7 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
    }

    if (memory) {
-        LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
+        LLAMA_LOG_DEBUG("%s: - reading memory module\n", __func__);

        memory->state_read(io);
    }
@@ -1971,21 +1981,21 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
    return io.n_bytes();
 }

-size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
    GGML_UNUSED(seq_id);

    if (memory) {
-        memory->state_write(io, seq_id);
+        memory->state_write(io, seq_id, flags);
    }

    return io.n_bytes();
 }

-size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
    GGML_UNUSED(seq_id);

    if (memory) {
-        memory->state_read(io, seq_id);
+        memory->state_read(io, seq_id, flags);
    }

    return io.n_bytes();
@@ -2016,6 +2026,21 @@ void llama_context::perf_reset() {
    n_reused    = 0;
 }

+std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> llama_context::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> ret;
+    for (const auto & buft_size : model.memory_breakdown()) {
+        ret[buft_size.first].model += buft_size.second;
+    }
+    for (const auto & buft_size : memory->memory_breakdown()) {
+        ret[buft_size.first].context += buft_size.second;
+    }
+    for (const auto & backend_ptr : backends) {
+        ggml_backend_t backend = backend_ptr.get();
+        ret[ggml_backend_sched_get_buffer_type(sched.get(), backend)].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    }
+    return ret;
+}
+
 //
 // training
 //
@@ -2048,7 +2073,7 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
    opt_params.opt_period      = n_batch / n_ubatch;
    opt_params.get_opt_pars    = lopt_params.get_opt_pars;
    opt_params.get_opt_pars_ud = lopt_params.get_opt_pars_ud;
-
+    opt_params.optimizer       = lopt_params.optimizer_type;
    opt_ctx = ggml_opt_init(opt_params);

    llama_opt_param_filter param_filter = lopt_params.param_filter;
@@ -2248,12 +2273,13 @@ llama_context_params llama_context_default_params() {
        /*.rope_scaling_type           =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED,
        /*.pooling_type                =*/ LLAMA_POOLING_TYPE_UNSPECIFIED,
        /*.attention_type              =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED,
+        /*.flash_attn_type             =*/ LLAMA_FLASH_ATTN_TYPE_AUTO,
        /*.rope_freq_base              =*/ 0.0f,
        /*.rope_freq_scale             =*/ 0.0f,
        /*.yarn_ext_factor             =*/ -1.0f,
-        /*.yarn_attn_factor            =*/ 1.0f,
-        /*.yarn_beta_fast              =*/ 32.0f,
-        /*.yarn_beta_slow              =*/ 1.0f,
+        /*.yarn_attn_factor            =*/ -1.0f,
+        /*.yarn_beta_fast              =*/ -1.0f,
+        /*.yarn_beta_slow              =*/ -1.0f,
        /*.yarn_orig_ctx               =*/ 0,
        /*.defrag_thold                =*/ -1.0f,
        /*.cb_eval                     =*/ nullptr,
@@ -2264,7 +2290,6 @@ llama_context_params llama_context_default_params() {
        /*.abort_callback_data         =*/ nullptr,
        /*.embeddings                  =*/ false,
        /*.offload_kqv                 =*/ true,
-        /*.flash_attn                  =*/ false,
        /*.no_perf                     =*/ true,
        /*.op_offload                  =*/ true,
        /*.swa_full                    =*/ true,
@@ -2292,12 +2317,30 @@ llama_context * llama_init_from_model(
        return nullptr;
    }

-    if (params.flash_attn && model->arch == LLM_ARCH_GROK) {
+    if (params.flash_attn_type != LLAMA_FLASH_ATTN_TYPE_DISABLED && model->arch == LLM_ARCH_GROK) {
        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
-        params.flash_attn = false;
+        params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
    }

-    if (ggml_is_quantized(params.type_v) && !params.flash_attn) {
+    if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_k)) {
+        const uint32_t blck_size = ggml_blck_size(params.type_k);
+        if (model->hparams.n_embd_head_k % blck_size != 0) {
+            LLAMA_LOG_ERROR("%s: K cache type %s with block size %u does not divide n_embd_head_k=%u\n",
+                __func__, ggml_type_name(params.type_k), blck_size, model->hparams.n_embd_head_k);
+            return nullptr;
+        }
+    }
+
+    if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_v)) {
+        const uint32_t blck_size = ggml_blck_size(params.type_v);
+        if (model->hparams.n_embd_head_v % blck_size != 0) {
+            LLAMA_LOG_ERROR("%s: V cache type %s with block size %u does not divide n_embd_head_k=%u\n",
+                __func__, ggml_type_name(params.type_v), blck_size, model->hparams.n_embd_head_v);
+            return nullptr;
+        }
+    }
+
+    if (ggml_is_quantized(params.type_v) && params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_DISABLED) {
        LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
        return nullptr;
    }
@@ -2343,16 +2386,6 @@ const llama_model * llama_get_model(const llama_context * ctx) {
    return &ctx->get_model();
 }

-// deprecated
-llama_kv_cache * llama_get_kv_self(llama_context * ctx) {
-    return dynamic_cast<llama_kv_cache *>(ctx->get_memory());
-}
-
-// deprecated
-void llama_kv_self_update(llama_context * ctx) {
-    ctx->kv_self_update(false);
-}
-
 enum llama_pooling_type llama_pooling_type(const llama_context * ctx) {
    return ctx->pooling_type();
 }
@@ -2570,168 +2603,6 @@ bool llama_memory_can_shift(llama_memory_t mem) {
    return mem->get_can_shift();
 }

-//
-// kv cache
-//
-
-// deprecated
-int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
-    const auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return 0;
-    }
-
-    int32_t res = 0;
-
-    for (uint32_t s = 0; s < ctx->get_cparams().n_seq_max; s++) {
-        const llama_pos p0 = kv->seq_pos_min(s);
-        const llama_pos p1 = kv->seq_pos_max(s);
-
-        if (p0 >= 0) {
-            res += (p1 - p0) + 1;
-        }
-    }
-
-    return res;
-}
-
-// deprecated
-// note: this is the same as above - will be removed anyway, so it's ok
-int32_t llama_kv_self_used_cells(const llama_context * ctx) {
-    const auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return 0;
-    }
-
-    int32_t res = 0;
-
-    for (uint32_t s = 0; s < ctx->get_cparams().n_seq_max; s++) {
-        const llama_pos p0 = kv->seq_pos_min(s);
-        const llama_pos p1 = kv->seq_pos_max(s);
-
-        if (p0 >= 0) {
-            res += (p1 - p0) + 1;
-        }
-    }
-
-    return res;
-}
-
-// deprecated
-void llama_kv_self_clear(llama_context * ctx) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_clear(kv, true);
-}
-
-// deprecated
-bool llama_kv_self_seq_rm(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return true;
-    }
-
-    return llama_memory_seq_rm(kv, seq_id, p0, p1);
-}
-
-// deprecated
-void llama_kv_self_seq_cp(
-        llama_context * ctx,
-         llama_seq_id   seq_id_src,
-         llama_seq_id   seq_id_dst,
-            llama_pos   p0,
-            llama_pos   p1) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_cp(kv, seq_id_src, seq_id_dst, p0, p1);
-}
-
-// deprecated
-void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_keep(kv, seq_id);
-}
-
-// deprecated
-void llama_kv_self_seq_add(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1,
-            llama_pos   delta) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_add(kv, seq_id, p0, p1, delta);
-}
-
-// deprecated
-void llama_kv_self_seq_div(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1,
-                  int   d) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_div(kv, seq_id, p0, p1, d);
-}
-
-// deprecated
-llama_pos llama_kv_self_seq_pos_min(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return -1;
-    }
-
-    return llama_memory_seq_pos_min(kv, seq_id);
-}
-
-// deprecated
-llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return -1;
-    }
-
-    return llama_memory_seq_pos_max(kv, seq_id);
-}
-
-// deprecated
-void llama_kv_self_defrag(llama_context * ctx) {
-    // force defrag
-    ctx->kv_self_defrag_sched();
-}
-
-// deprecated
-bool llama_kv_self_can_shift(const llama_context * ctx) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return false;
-    }
-
-    return llama_memory_can_shift(kv);
-}
-
 // llama state API

 // deprecated
@@ -2801,19 +2672,31 @@ bool llama_state_save_file(llama_context * ctx, const char * path_session, const
 }

 size_t llama_state_seq_get_size(llama_context * ctx, llama_seq_id seq_id) {
-    return ctx->state_seq_get_size(seq_id);
+    return llama_state_seq_get_size_ext(ctx, seq_id, 0);
 }

 size_t llama_state_seq_get_data(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) {
-    ctx->synchronize();
-
-    return ctx->state_seq_get_data(seq_id, dst, size);
+    return llama_state_seq_get_data_ext(ctx, dst, size, seq_id, 0);
 }

 size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) {
+    return llama_state_seq_set_data_ext(ctx, src, size, seq_id, 0);
+}
+
+size_t llama_state_seq_get_size_ext(llama_context * ctx, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    return ctx->state_seq_get_size(seq_id, flags);
+}
+
+size_t llama_state_seq_get_data_ext(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
    ctx->synchronize();

-    return ctx->state_seq_set_data(seq_id, src, size);
+    return ctx->state_seq_get_data(seq_id, dst, size, flags);
+}
+
+size_t llama_state_seq_set_data_ext(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    ctx->synchronize();
+
+    return ctx->state_seq_set_data(seq_id, src, size, flags);
 }

 size_t llama_state_seq_save_file(llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {
@@ -2896,6 +2779,142 @@ void llama_perf_context_reset(llama_context * ctx) {
    ctx->perf_reset();
 }

+void llama_memory_breakdown_print(const struct llama_context * ctx) {
+    const std::vector<ggml_backend_dev_t> & devices = ctx->get_model().devices;
+
+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> memory_breakdown = ctx->memory_breakdown();
+
+    std::vector<std::array<std::string, 9>> table_data;
+    table_data.reserve(devices.size());
+    const std::string template_header = "%s: | %s | %s   %s    %s   %s   %s   %s    %s |\n";
+    const std::string template_gpu    = "%s: | %s | %s = %s + (%s = %s + %s + %s) + %s |\n";
+    const std::string template_other  = "%s: | %s | %s   %s    %s = %s + %s + %s    %s |\n";
+
+    table_data.push_back({template_header, "memory breakdown [MiB]", "total", "free", "self", "model", "context", "compute", "unaccounted"});
+
+    constexpr size_t MiB = 1024 * 1024;
+    const std::vector<std::string> desc_prefixes_strip = {"NVIDIA ", "GeForce ", "Tesla ", "AMD ", "Radeon ", "Instinct "};
+
+    // track seen buffer types to avoid double counting:
+    std::set<ggml_backend_buffer_type_t> seen_buffer_types;
+
+    // accumulative memory breakdown for each device and for host:
+    std::vector<llama_memory_breakdown_data> mb_dev(devices.size());
+    llama_memory_breakdown_data              mb_host;
+
+    for (const auto & buft_mb : memory_breakdown) {
+        ggml_backend_buffer_type_t          buft = buft_mb.first;
+        const llama_memory_breakdown_data & mb   = buft_mb.second;
+        if (ggml_backend_buft_is_host(buft)) {
+            mb_host.model   += mb.model;
+            mb_host.context += mb.context;
+            mb_host.compute += mb.compute;
+            seen_buffer_types.insert(buft);
+            continue;
+        }
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (dev) {
+            int i_dev = -1;
+            for (size_t i = 0; i < devices.size(); i++) {
+                if (devices[i] == dev) {
+                    i_dev = i;
+                    break;
+                }
+            }
+            if (i_dev != -1) {
+                mb_dev[i_dev].model   += mb.model;
+                mb_dev[i_dev].context += mb.context;
+                mb_dev[i_dev].compute += mb.compute;
+                seen_buffer_types.insert(buft);
+                continue;
+            }
+        }
+    }
+
+    // print memory breakdown for each device:
+    for (size_t i = 0; i < devices.size(); i++) {
+        ggml_backend_dev_t          dev = devices[i];
+        llama_memory_breakdown_data mb  = mb_dev[i];
+
+        const std::string name = ggml_backend_dev_name(dev);
+        std::string desc = ggml_backend_dev_description(dev);
+        for (const std::string & prefix : desc_prefixes_strip) {
+            if (desc.length() >= prefix.length() && desc.substr(0, prefix.length()) == prefix) {
+                desc = desc.substr(prefix.length());
+            }
+        }
+
+        size_t free, total;
+        ggml_backend_dev_memory(dev, &free, &total);
+
+        const size_t self = mb.model + mb.context + mb.compute;
+        const size_t unaccounted = total - self - free;
+
+        table_data.push_back({
+            template_gpu,
+            "  - " + name + " (" + desc + ")",
+            std::to_string(total / MiB),
+            std::to_string(free / MiB),
+            std::to_string(self / MiB),
+            std::to_string(mb.model / MiB),
+            std::to_string(mb.context / MiB),
+            std::to_string(mb.compute / MiB),
+            std::to_string(unaccounted / MiB)});
+    }
+
+    // print memory breakdown for host:
+    {
+        const size_t self = mb_host.model + mb_host.context + mb_host.compute;
+        table_data.push_back({
+            template_other,
+            "  - Host",
+            "", // total
+            "", // free
+            std::to_string(self / MiB),
+            std::to_string(mb_host.model / MiB),
+            std::to_string(mb_host.context / MiB),
+            std::to_string(mb_host.compute / MiB),
+            ""}); // unaccounted
+    }
+
+    // print memory breakdown for all remaining buffer types:
+    for (const auto & buft_mb : memory_breakdown) {
+        ggml_backend_buffer_type_t          buft = buft_mb.first;
+        const llama_memory_breakdown_data & mb   = buft_mb.second;
+        if (seen_buffer_types.count(buft) == 1) {
+            continue;
+        }
+        const std::string name = ggml_backend_buft_name(buft);
+        const size_t self = mb.model + mb.context + mb.compute;
+        table_data.push_back({
+            template_other,
+            "  - " + name,
+            "", // total
+            "", // free
+            std::to_string(self / MiB),
+            std::to_string(mb.model / MiB),
+            std::to_string(mb.context / MiB),
+            std::to_string(mb.compute / MiB),
+            ""}); // unaccounted
+        seen_buffer_types.insert(buft);
+    }
+
+    for (size_t j = 1; j < table_data[0].size(); j++) {
+        size_t max_len = 0;
+        for (const auto & td : table_data) {
+            max_len = std::max(max_len, td[j].length());
+        }
+        for (auto & td : table_data) {
+            td[j].insert(j == 1 ? td[j].length() : 0, max_len - td[j].length(), ' ');
+        }
+    }
+    for (const auto & td : table_data) {
+        LLAMA_LOG_INFO(td[0].c_str(),
+            __func__, td[1].c_str(), td[2].c_str(), td[3].c_str(), td[4].c_str(), td[5].c_str(),
+            td[6].c_str(), td[7].c_str(), td[8].c_str());
+    }
+}
+
 //
 // training
 //
--- a/llama/llama.cpp/src/llama-context.h
+++ b/llama/llama.cpp/src/llama-context.h
@@ -17,9 +17,17 @@ class llama_batch_allocr;
 class llama_io_read_i;
 class llama_io_write_i;

+// "memory" as in abstract memory for the context
 struct llama_memory_i;
 struct llama_memory_context_i;

+// "memory" as in physical memory for a buffer type, in bytes
+struct llama_memory_breakdown_data {
+    size_t model   = 0; // memory allocated for the model
+    size_t context = 0; // memory allocated for the context
+    size_t compute = 0; // memory allocated for temporary compute buffers
+};
+
 struct llama_context {
    // init scheduler and compute buffers, reserve worst-case graphs
    llama_context(
@@ -46,10 +54,8 @@ struct llama_context {

    llama_memory_t get_memory() const;

-    // return true of the KV cache was updated
-    // TODO: remove
-    bool kv_self_update(bool optimize);
-    void kv_self_defrag_sched();
+    // return true if the memory was updated
+    bool memory_update(bool optimize);

    enum llama_pooling_type pooling_type() const;

@@ -111,9 +117,9 @@ struct llama_context {
    size_t state_get_data(      uint8_t * dst, size_t size);
    size_t state_set_data(const uint8_t * src, size_t size);

-    size_t state_seq_get_size(llama_seq_id seq_id);
-    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
-    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+    size_t state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size, llama_state_seq_flags flags);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags);

    bool state_load_file(
            const char * filepath,
@@ -146,12 +152,15 @@ struct llama_context {
    llama_perf_context_data perf_get_data() const;
    void perf_reset();

+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> memory_breakdown() const;
+
    //
    // training
    //

    void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);

+    // TODO: more flexible combinations of logical/physical batch size and context size
    void opt_epoch(
            ggml_opt_dataset_t      dataset,
            ggml_opt_result_t       result_train,
@@ -197,7 +206,7 @@ public:
    ggml_status graph_compute(ggml_cgraph * gf, bool batched);

    // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx);
+    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false);

 private:
    llm_graph_params graph_params(
@@ -212,8 +221,8 @@ private:
    size_t state_write_data(llama_io_write_i & io);
    size_t state_read_data (llama_io_read_i  & io);

-    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
-    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id, llama_state_seq_flags flags);

    //
    // members
@@ -229,9 +238,6 @@ private:

    std::unique_ptr<llama_memory_i> memory;

-    // TODO: temporary, until the llama_kv_self_defrag() API is removed
-    bool memory_force_optimize = false;
-
    // decode output (2-dimensional array: [n_outputs][n_vocab])
    size_t  logits_size = 0; // capacity (of floats) for logits
    float * logits      = nullptr;
@@ -287,10 +293,6 @@ private:

    bool has_evaluated_once = false;

-    // env: LLAMA_SET_ROWS (temporary)
-    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    bool supports_set_rows = true;
-
    // env: LLAMA_GRAPH_REUSE_DISABLE
    bool graph_reuse_disable = false;

--- a/llama/llama.cpp/src/llama-cparams.h
+++ b/llama/llama.cpp/src/llama-cparams.h
@@ -4,7 +4,7 @@

 #include <cstdint>

-#define LLAMA_MAX_SEQ 64
+#define LLAMA_MAX_SEQ 256

 struct llama_cparams {
    uint32_t n_ctx;           // context size used during inference
@@ -24,7 +24,6 @@ struct llama_cparams {
    float yarn_attn_factor;
    float yarn_beta_fast;
    float yarn_beta_slow;
-    float defrag_thold;

    bool embeddings;
    bool causal_attn;
--- a/llama/llama.cpp/src/llama-graph.cpp
+++ b/llama/llama.cpp/src/llama-graph.cpp
@@ -4,8 +4,8 @@
 #include "llama-batch.h"
 #include "llama-cparams.h"

-#include "llama-kv-cache-unified.h"
-#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
 #include "llama-memory-recurrent.h"

@@ -204,7 +204,10 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
        std::vector<int> target_pos(n_seqs_unq, -1);
        std::vector<int> target_row(n_seqs_unq, -1);

-        bool last = cparams.pooling_type == LLAMA_POOLING_TYPE_LAST;
+        const bool last = (
+             cparams.pooling_type == LLAMA_POOLING_TYPE_LAST ||
+            (cparams.pooling_type == LLAMA_POOLING_TYPE_RANK && arch == LLM_ARCH_QWEN3) // qwen3 reranking & embedding models use last token
+        );

        for (int i = 0; i < n_tokens; ++i) {
            const llama_pos pos = ubatch->pos[i];
@@ -258,6 +261,36 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
    }
 }

+static void print_mask(float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
+    LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
+    const char * swa_type_str = (swa_type == LLAMA_SWA_TYPE_NONE) ? "LLAMA_SWA_TYPE_NONE" :
+                          (swa_type == LLAMA_SWA_TYPE_STANDARD) ? "LLAMA_SWA_TYPE_STANDARD" :
+                          (swa_type == LLAMA_SWA_TYPE_CHUNKED) ? "LLAMA_SWA_TYPE_CHUNKED" :
+                          (swa_type == LLAMA_SWA_TYPE_SYMMETRIC) ? "LLAMA_SWA_TYPE_SYMMETRIC" : "unknown";
+    LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
+    LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
+    LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);
+
+    LLAMA_LOG_DEBUG("    ");
+    for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
+        LLAMA_LOG_DEBUG("%2d", j);
+    }
+    LLAMA_LOG_DEBUG("\n");
+
+    for (int i = 0; i < std::min((int64_t)20, n_tokens); ++i) {
+        LLAMA_LOG_DEBUG(" %2d ", i);
+        for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
+            float val = data[i * n_kv + j];
+            if (val == -INFINITY) {
+                LLAMA_LOG_DEBUG(" ∞");
+            } else {
+                LLAMA_LOG_DEBUG(" 0");
+            }
+        }
+        LLAMA_LOG_DEBUG("\n");
+    }
+}
+
 void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
    const int64_t n_kv     = ubatch->n_tokens;
    const int64_t n_tokens = ubatch->n_tokens;
@@ -267,6 +300,9 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {

    float * data = (float *) kq_mask->data;

+    // [TAG_NO_CACHE_ISWA]
+    GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "TODO: implement");
+
    for (int h = 0; h < 1; ++h) {
        for (int i1 = 0; i1 < n_tokens; ++i1) {
            const llama_seq_id s1 = ubatch->seq_id[i1][0];
@@ -277,32 +313,44 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
                    const llama_seq_id s0 = ubatch->seq_id[i0][0];

+                    if (s0 != s1) {
+                        continue; // skip different sequences
+                    }
+
+                    if (cparams.causal_attn && ubatch->pos[i0] > ubatch->pos[i1]) {
+                        continue; // skip future tokens for causal attention
+                    }
+
+                    // TODO: this does not take into account that some layers are SWA and others are note (i.e. iSWA) [TAG_NO_CACHE_ISWA]
+                    //if (hparams.is_masked_swa(ubatch->pos[i0], ubatch->pos[i1])) {
+                    //    continue; // skip masked tokens for SWA
+                    //}
+
                    // TODO: reimplement this like in llama_kv_cache_unified
-                    if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) {
-                        if (hparams.use_alibi) {
-                            f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
-                        } else {
-                            f = 0.0f;
-                        }
-                        break;
+                    if (hparams.use_alibi) {
+                        f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
+                    } else {
+                        f = 0.0f;
                    }
                }
-
                data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
            }
        }
    }
+    if (debug) {
+        print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
+    }
 }

-void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {
    mctx->set_input_k_idxs(self_k_idxs, ubatch);
    mctx->set_input_v_idxs(self_v_idxs, ubatch);

    mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
 }

-bool llm_graph_input_attn_kv_unified::can_reuse(const llm_graph_params & params) {
-    const auto * mctx = static_cast<const llama_kv_cache_unified_context *>(params.mctx);
+bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);

    this->mctx = mctx;

@@ -314,12 +362,10 @@ bool llm_graph_input_attn_kv_unified::can_reuse(const llm_graph_params & params)
    res &= self_kq_mask->ne[0] == mctx->get_n_kv();
    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);

-    res &= mctx->get_supports_set_rows(); // TODO: tmp
-
    return res;
 }

-void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
    mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
    mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);

@@ -331,8 +377,8 @@ void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch
    mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
 }

-bool llm_graph_input_attn_kv_unified_iswa::can_reuse(const llm_graph_params & params) {
-    const auto * mctx = static_cast<const llama_kv_cache_unified_iswa_context *>(params.mctx);
+bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_iswa_context *>(params.mctx);

    this->mctx = mctx;

@@ -350,8 +396,6 @@ bool llm_graph_input_attn_kv_unified_iswa::can_reuse(const llm_graph_params & pa
    res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
    res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);

-    res &= mctx->get_base()->get_supports_set_rows(); // TODO: tmp
-
    return res;
 }

@@ -879,15 +923,29 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
        selection_probs = logits;
    }

+    if (arch == LLM_ARCH_GROVEMOE) {
+        selection_probs = ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens]
+        cb(selection_probs, "ffn_moe_probs_biased", il);
+    }
+
    // select experts
    ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
    cb(selected_experts->src[0], "ffn_moe_argsort", il);
    cb(selected_experts, "ffn_moe_topk", il);

-    ggml_tensor * weights = ggml_get_rows(ctx0,
-            ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
+    if (arch == LLM_ARCH_GROVEMOE && n_expert != hparams.n_expert) {
+        // TODO: Use scalar div instead when/if implemented
+        ggml_tensor * f_sel = ggml_cast(ctx0, selected_experts, GGML_TYPE_F32);
+        selected_experts = ggml_cast(ctx0, ggml_scale(ctx0, f_sel, 1.0f / float(hparams.n_group_experts)), GGML_TYPE_I32);
+        probs = ggml_reshape_3d(ctx0, probs, 1, hparams.n_expert, n_tokens);
+    } else {
+        probs = ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens);
+    }
+
+    ggml_tensor * weights = ggml_get_rows(ctx0, probs, selected_experts); // [1, n_expert_used, n_tokens]
    cb(weights, "ffn_moe_weights", il);

+
    if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT) {
        weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
        weights = ggml_soft_max(ctx0, weights); // [n_expert_used, n_tokens]
@@ -911,6 +969,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
        cb(weights, "ffn_moe_weights_scaled", il);
    }

+    //call early so that topk-moe can be used
+    ggml_build_forward_expand(gf, weights);
+
    cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);

    if (weight_before_ffn) {
@@ -1136,7 +1197,7 @@ ggml_tensor * llm_graph_context::build_inp_mean() const {
 }

 ggml_tensor * llm_graph_context::build_inp_cls() const {
-    auto inp = std::make_unique<llm_graph_input_cls>(cparams);
+    auto inp = std::make_unique<llm_graph_input_cls>(cparams, arch);

    auto & cur = inp->cls;

@@ -1186,7 +1247,7 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
 }

 ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);

    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, mctx_cur);

@@ -1223,15 +1284,16 @@ ggml_tensor * llm_graph_context::build_attn_mha(
         ggml_tensor * v,
         ggml_tensor * kq_b,
         ggml_tensor * kq_mask,
-         ggml_tensor * v_mla,
         ggml_tensor * sinks,
-             float     kq_scale) const {
+         ggml_tensor * v_mla,
+               float   kq_scale,
+                 int   il) const {
    const bool v_trans = v->nb[1] > v->nb[2];

    // split the batch into streams if needed
    const auto n_stream = k->ne[3];

-    q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream);
+    q = ggml_view_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream, q->nb[1], q->nb[2], q->nb[3]/n_stream, 0);

    q = ggml_permute(ctx0, q, 0, 2, 1, 3);
    k = ggml_permute(ctx0, k, 0, 2, 1, 3);
@@ -1260,6 +1322,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(

        cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
                                  hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
+        cb(cur, LLAMA_TENSOR_NAME_FATTN, il);

        ggml_flash_attn_ext_add_sinks(cur, sinks);
        ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);
@@ -1275,6 +1338,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
            // The permutations are noops and only change how the tensor data is interpreted.
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            cur = ggml_mul_mat(ctx0, v_mla, cur);
+            cb(cur, "fattn_mla", il);
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs.
 #endif
@@ -1283,6 +1347,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
    } else {
        ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+        cb(kq, "kq", il);

        // note: this op tends to require high floating point range
        //       while for some models F16 is enough, for others it is not, so we default to F32 here
@@ -1290,38 +1355,48 @@ ggml_tensor * llm_graph_context::build_attn_mha(

        if (arch == LLM_ARCH_GROK) {
            // need to do the following:
-            // multiply by attn_output_multiplyer of 0.08838834764831845
+            // multiply by attn_output_multiplier
            // and then :
            // kq = 30 * tanh(kq / 30)
            // before the softmax below

-            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f));
-            kq = ggml_scale(ctx0, kq, 30);
+            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, hparams.f_attn_out_scale / hparams.f_attn_logit_softcapping));
+            cb(kq, "kq_tanh", il);
+            kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
+            cb(kq, "kq_scaled", il);
        }

        if (hparams.attn_soft_cap) {
            kq = ggml_scale(ctx0, kq, 1.0f / hparams.f_attn_logit_softcapping);
+            cb(kq, "kq_scaled_1", il);
            kq = ggml_tanh (ctx0, kq);
+            cb(kq, "kq_tanh", il);
            kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
+            cb(kq, "kq_scaled_2", il);
        }

        if (kq_b) {
            kq = ggml_add(ctx0, kq, kq_b);
+            cb(kq, "kq_plus_kq_b", il);
        }

        kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
        ggml_soft_max_add_sinks(kq, sinks);
+        cb(kq, "kq_soft_max", il);

        if (!v_trans) {
            // note: avoid this branch
            v = ggml_cont(ctx0, ggml_transpose(ctx0, v));
+            cb(v, "v_cont", il);
        }

        ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
+        cb(kqv, "kqv", il);

        // for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA
        if (v_mla) {
            kqv = ggml_mul_mat(ctx0, v_mla, kqv);
+            cb(kqv, "kqv_mla", il);
        }

        cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
@@ -1360,6 +1435,7 @@ ggml_tensor * llm_graph_context::build_attn(
        ggml_tensor * k_cur,
        ggml_tensor * v_cur,
        ggml_tensor * kq_b,
+        ggml_tensor * sinks,
        ggml_tensor * v_mla,
            float     kq_scale,
            int       il) const {
@@ -1375,13 +1451,14 @@ ggml_tensor * llm_graph_context::build_attn(

    // [TAG_NO_CACHE_PAD]
    // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
-    assert(!ubatch.equal_seqs());
+    //       but it might not be worth it: https://github.com/ggml-org/llama.cpp/pull/15636
+    //assert(!ubatch.equal_seqs() || (k_cur->ne[3] == 1 && k_cur->ne[3] == ubatch.n_seqs_unq));

    ggml_tensor * q = q_cur;
    ggml_tensor * k = k_cur;
    ggml_tensor * v = v_cur;

-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);

    if (wo) {
@@ -1399,17 +1476,17 @@ ggml_tensor * llm_graph_context::build_attn(
    return cur;
 }

-static std::unique_ptr<llm_graph_input_attn_kv_unified> build_attn_inp_kv_unified_impl(
+static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
           ggml_context * ctx0,
     const llama_ubatch & ubatch,
    const llama_hparams & hparams,
    const llama_cparams & cparams,
-    const llama_kv_cache_unified_context * mctx_cur) {
+    const llama_kv_cache_context * mctx_cur) {

-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, mctx_cur);
+    auto inp = std::make_unique<llm_graph_input_attn_kv>(hparams, cparams, mctx_cur);

    {
-        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");

        const auto n_kv     = mctx_cur->get_n_kv();
        const auto n_tokens = ubatch.n_tokens;
@@ -1427,22 +1504,23 @@ static std::unique_ptr<llm_graph_input_attn_kv_unified> build_attn_inp_kv_unifie
    return inp;
 }

-llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+llm_graph_input_attn_kv * llm_graph_context::build_attn_inp_kv() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);

-    auto inp = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
+    auto inp = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur);

-    return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
+    return (llm_graph_input_attn_kv *) res->add_input(std::move(inp));
 }

 ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_attn_kv_unified * inp,
+        llm_graph_input_attn_kv * inp,
        ggml_tensor * wo,
        ggml_tensor * wo_b,
        ggml_tensor * q_cur,
        ggml_tensor * k_cur,
        ggml_tensor * v_cur,
        ggml_tensor * kq_b,
+        ggml_tensor * sinks,
        ggml_tensor * v_mla,
            float     kq_scale,
            int       il) const {
@@ -1469,7 +1547,7 @@ ggml_tensor * llm_graph_context::build_attn(
    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
    ggml_tensor * v = mctx_cur->get_v(ctx0, il);

-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);

    if (wo) {
@@ -1488,40 +1566,15 @@ ggml_tensor * llm_graph_context::build_attn(
 }

 ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_attn_kv_unified_iswa * inp,
+        llm_graph_input_attn_kv_iswa * inp,
        ggml_tensor * wo,
        ggml_tensor * wo_b,
        ggml_tensor * q_cur,
        ggml_tensor * k_cur,
        ggml_tensor * v_cur,
        ggml_tensor * kq_b,
-        ggml_tensor * v_mla,
-            float     kq_scale,
-            int       il) const {
-    return build_attn_with_sinks(
-            inp,
-            wo,
-            wo_b,
-            q_cur,
-            k_cur,
-            v_cur,
-            kq_b,
-            v_mla,
-            nullptr,
-            kq_scale,
-            il);
-}
-
-ggml_tensor * llm_graph_context::build_attn_with_sinks(
-        llm_graph_input_attn_kv_unified_iswa * inp,
-        ggml_tensor * wo,
-        ggml_tensor * wo_b,
-        ggml_tensor * q_cur,
-        ggml_tensor * k_cur,
-        ggml_tensor * v_cur,
-        ggml_tensor * kq_b,
-        ggml_tensor * v_mla,
        ggml_tensor * sinks,
+        ggml_tensor * v_mla,
            float     kq_scale,
            int       il) const {
    // these nodes are added to the graph together so that they are not reordered
@@ -1561,7 +1614,7 @@ ggml_tensor * llm_graph_context::build_attn_with_sinks(
    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
    ggml_tensor * v = mctx_cur->get_v(ctx0, il);

-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, sinks, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);

    if (wo) {
@@ -1600,6 +1653,7 @@ ggml_tensor * llm_graph_context::build_attn(
        ggml_tensor * k_cur,
        ggml_tensor * v_cur,
        ggml_tensor * kq_b,
+        ggml_tensor * sinks,
        ggml_tensor * v_mla,
            float     kq_scale,
            int       il) const {
@@ -1615,7 +1669,7 @@ ggml_tensor * llm_graph_context::build_attn(
    ggml_tensor * k = k_cur;
    ggml_tensor * v = v_cur;

-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);

    if (wo) {
@@ -1636,10 +1690,10 @@ ggml_tensor * llm_graph_context::build_attn(
 // TODO: maybe separate the inner implementation into a separate function
 //       like with the non-sliding window equivalent
 //       once sliding-window hybrid caches are a thing.
-llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
+llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_iswa_context *>(mctx);

-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
+    auto inp = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, mctx_cur);

    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;

@@ -1656,7 +1710,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
    }

    {
-        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
+        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache for non-SWA");

        const auto n_kv = mctx_cur->get_swa()->get_n_kv();

@@ -1669,7 +1723,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
    }

-    return (llm_graph_input_attn_kv_unified_iswa *) res->add_input(std::move(inp));
+    return (llm_graph_input_attn_kv_iswa *) res->add_input(std::move(inp));
 }

 ggml_tensor * llm_graph_context::build_rs(
@@ -1792,7 +1846,7 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
    const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx);

    auto inp_rs   = build_rs_inp_impl(ctx0, ubatch, mctx_cur->get_recr());
-    auto inp_attn = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
+    auto inp_attn = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());

    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(std::move(inp_attn), std::move(inp_rs), mctx_cur);

@@ -1843,34 +1897,32 @@ void llm_graph_context::build_pooling(
        case LLAMA_POOLING_TYPE_RANK:
            {
                ggml_tensor * inp_cls = build_inp_cls();
-                inp = ggml_get_rows(ctx0, inp, inp_cls);
+                cur = ggml_get_rows(ctx0, inp, inp_cls);

+                // classification head
+                // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
                if (cls) {
-                    // classification head
-                    // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                    cur = ggml_mul_mat(ctx0, cls, inp);
+                    cur = ggml_mul_mat(ctx0, cls, cur);
                    if (cls_b) {
                        cur = ggml_add(ctx0, cur, cls_b);
                    }
                    cur = ggml_tanh(ctx0, cur);
+                }

-                    // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
-                    // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
-                    if (cls_out) {
-                        cur = ggml_mul_mat(ctx0, cls_out, cur);
-                        if (cls_out_b) {
-                            cur = ggml_add(ctx0, cur, cls_out_b);
-                        }
-                    }
-                } else if (cls_out) {
-                    // Single layer classification head (direct projection)
-                    // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
-                    cur = ggml_mul_mat(ctx0, cls_out, inp);
+                // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+                // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
+                // Single layer classification head (direct projection)
+                // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
+                if (cls_out) {
+                    cur = ggml_mul_mat(ctx0, cls_out, cur);
                    if (cls_out_b) {
                        cur = ggml_add(ctx0, cur, cls_out_b);
                    }
-                } else {
-                    GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
+                }
+
+                // softmax for qwen3 reranker
+                if (arch == LLM_ARCH_QWEN3) {
+                    cur = ggml_soft_max(ctx0, cur);
                }
            } break;
        default:
--- a/llama/llama.cpp/src/llama-graph.h
+++ b/llama/llama.cpp/src/llama-graph.h
@@ -19,8 +19,8 @@ struct llama_cparams;

 struct llama_memory_context_i;

-class llama_kv_cache_unified_context;
-class llama_kv_cache_unified_iswa_context;
+class llama_kv_cache_context;
+class llama_kv_cache_iswa_context;
 class llama_memory_recurrent_context;
 class llama_memory_hybrid_context;

@@ -78,6 +78,11 @@ struct llm_graph_params;

 class llm_graph_input_i {
 public:
+    llm_graph_input_i() {
+        const char * LLAMA_GRAPH_INPUT_DEBUG = getenv("LLAMA_GRAPH_INPUT_DEBUG");
+        debug = LLAMA_GRAPH_INPUT_DEBUG ? atoi(LLAMA_GRAPH_INPUT_DEBUG) : 0;
+    }
+
    virtual ~llm_graph_input_i() = default;

    virtual void set_input(const llama_ubatch * ubatch) = 0;
@@ -90,6 +95,9 @@ public:
        GGML_UNUSED(params);
        return false;
    }
+protected:
+    // env: LLAMA_GRAPH_INPUT_DEBUG
+    int debug = 0;
 };

 using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
@@ -152,7 +160,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
 public:
    llm_graph_input_pos_bucket_kv(
            const llama_hparams & hparams,
-            const llama_kv_cache_unified_context * mctx) : hparams(hparams), mctx(mctx) {}
+            const llama_kv_cache_context * mctx) : hparams(hparams), mctx(mctx) {}
    virtual ~llm_graph_input_pos_bucket_kv() = default;

    void set_input(const llama_ubatch * ubatch) override;
@@ -161,7 +169,7 @@ public:

    const llama_hparams hparams;

-    const llama_kv_cache_unified_context * mctx;
+    const llama_kv_cache_context * mctx;
 };

 class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -198,7 +206,7 @@ public:

 class llm_graph_input_cls : public llm_graph_input_i {
 public:
-    llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {}
+    llm_graph_input_cls(const llama_cparams & cparams, const llm_arch arch) : cparams(cparams), arch(arch) {}
    virtual ~llm_graph_input_cls() = default;

    void set_input(const llama_ubatch * ubatch) override;
@@ -206,6 +214,7 @@ public:
    ggml_tensor * cls; // I32 [n_batch]

    const llama_cparams cparams;
+    const llm_arch arch;
 };

 class llm_graph_input_rs : public llm_graph_input_i {
@@ -257,17 +266,17 @@ public:
    const llama_cparams cparams;
 };

-class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
+class llm_graph_input_attn_kv : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_kv_unified(
+    llm_graph_input_attn_kv(
            const llama_hparams & hparams,
            const llama_cparams & cparams,
-            const llama_kv_cache_unified_context * mctx) :
+            const llama_kv_cache_context * mctx) :
        hparams(hparams),
        cparams(cparams),
        mctx(mctx) {
    }
-    ~llm_graph_input_attn_kv_unified() = default;
+    ~llm_graph_input_attn_kv() = default;

    void set_input(const llama_ubatch * ubatch) override;

@@ -290,20 +299,20 @@ public:
    const llama_hparams hparams;
    const llama_cparams cparams;

-    const llama_kv_cache_unified_context * mctx;
+    const llama_kv_cache_context * mctx;
 };

-class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
+class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_kv_unified_iswa(
+    llm_graph_input_attn_kv_iswa(
            const llama_hparams & hparams,
            const llama_cparams & cparams,
-            const llama_kv_cache_unified_iswa_context * mctx) :
+            const llama_kv_cache_iswa_context * mctx) :
        hparams(hparams),
        cparams(cparams),
        mctx(mctx) {
    }
-    ~llm_graph_input_attn_kv_unified_iswa() = default;
+    ~llm_graph_input_attn_kv_iswa() = default;

    void set_input(const llama_ubatch * ubatch) override;

@@ -330,7 +339,7 @@ public:
    const llama_hparams hparams;
    const llama_cparams cparams;

-    const llama_kv_cache_unified_iswa_context * mctx;
+    const llama_kv_cache_iswa_context * mctx;
 };

 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -351,7 +360,7 @@ public:
 class llm_graph_input_mem_hybrid : public llm_graph_input_i {
 public:
    llm_graph_input_mem_hybrid(
-            std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn,
+            std::unique_ptr<llm_graph_input_attn_kv> inp_attn,
            std::unique_ptr<llm_graph_input_rs>              inp_rs,
            const llama_memory_hybrid_context *              mctx) :
        inp_attn(std::move(inp_attn)),
@@ -361,11 +370,11 @@ public:

    void set_input(const llama_ubatch * ubatch) override;

-    std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn;
-    std::unique_ptr<llm_graph_input_rs>              inp_rs;
+    std::unique_ptr<llm_graph_input_attn_kv> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>      inp_rs;

-    llm_graph_input_attn_kv_unified * get_attn() const { return inp_attn.get(); }
-    llm_graph_input_rs              * get_recr() const { return inp_rs.get(); }
+    llm_graph_input_attn_kv * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs      * get_recr() const { return inp_rs.get(); }

    const llama_memory_hybrid_context * mctx;
 };
@@ -680,14 +689,15 @@ struct llm_graph_context {
    //

    ggml_tensor * build_attn_mha(
-             ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
-             ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
-             ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
-             ggml_tensor * kq_b,
-             ggml_tensor * kq_mask,
-             ggml_tensor * sinks,
-             ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                   float   kq_scale) const;
+            ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
+            ggml_tensor * kq_b,
+            ggml_tensor * kq_mask,
+            ggml_tensor * sinks,   // [n_head_q]
+            ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;

    llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;

@@ -699,50 +709,39 @@ struct llm_graph_context {
            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
            ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                  float   kq_scale,
                    int   il) const;

-    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
+    llm_graph_input_attn_kv * build_attn_inp_kv() const;

    ggml_tensor * build_attn(
-            llm_graph_input_attn_kv_unified * inp,
+            llm_graph_input_attn_kv * inp,
            ggml_tensor * wo,
            ggml_tensor * wo_b,
            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
            ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                  float   kq_scale,
                    int   il) const;

-    llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
+    llm_graph_input_attn_kv_iswa * build_attn_inp_kv_iswa() const;

    // note: if k_cur or v_cur are not provided, they will not be stored in the memory
    ggml_tensor * build_attn(
-            llm_graph_input_attn_kv_unified_iswa * inp,
+            llm_graph_input_attn_kv_iswa * inp,
            ggml_tensor * wo,
            ggml_tensor * wo_b,
            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
            ggml_tensor * kq_b,
-            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                  float   kq_scale,
-                    int   il) const;
-
-    // TODO: temporary to keep the diff small. after the code is public will refactor to simplify this
-    ggml_tensor * build_attn_with_sinks(
-            llm_graph_input_attn_kv_unified_iswa * inp,
-            ggml_tensor * wo,
-            ggml_tensor * wo_b,
-            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
-            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
-            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
-            ggml_tensor * kq_b,
-            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
            ggml_tensor * sinks, // [n_head_q]
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                  float   kq_scale,
                    int   il) const;

@@ -756,6 +755,7 @@ struct llm_graph_context {
            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
            ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                  float   kq_scale,
                    int   il) const;
@@ -765,7 +765,7 @@ struct llm_graph_context {
    //

    // TODO: move this implementation to llama_memory_recurrent.
-    //       this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
+    //       this is analogous to llama_kv_cache::cpy_k / cpy_v
    //       when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
    //         implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
    //         `llama_memory_recurrent`
--- a/llama/llama.cpp/src/llama-hparams.cpp
+++ b/llama/llama.cpp/src/llama-hparams.cpp
@@ -1,6 +1,7 @@
 #include "llama-hparams.h"

 #include "ggml.h"
+#include <cassert>

 void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {
    if (dense_first) {
@@ -161,3 +162,64 @@ bool llama_hparams::is_swa(uint32_t il) const {

    GGML_ABORT("fatal error");
 }
+
+bool llama_hparams::has_kv(uint32_t il) const {
+    if (n_layer_kv_from_start >= 0) {
+        if (il < (uint32_t) n_layer_kv_from_start) {
+            return true;
+        }
+
+        return false;
+    }
+
+    // by default, all layers have kv
+    return true;
+}
+
+uint32_t llama_hparams::n_layer_kv() const {
+    uint32_t res = 0;
+
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (has_kv(il)) {
+            res++;
+        }
+    }
+
+    return res;
+}
+
+bool llama_hparams::is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
+    assert(p0 >= 0 && p1 >= 0);
+
+    switch (swa_type) {
+        case LLAMA_SWA_TYPE_NONE:
+            {
+            } break;
+        case LLAMA_SWA_TYPE_STANDARD:
+            {
+                if (p1 - p0 >= (int32_t) n_swa) {
+                    return true;
+                }
+            } break;
+        case LLAMA_SWA_TYPE_CHUNKED:
+            {
+                const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
+
+                if (p0 < pos_chunk_start) {
+                    return true;
+                }
+            } break;
+        case LLAMA_SWA_TYPE_SYMMETRIC:
+            {
+                const int32_t half_n_swa = (int32_t) n_swa / 2;
+                const int32_t pos_diff = p1 - p0;
+
+                // Mask if outside the symmetric window
+                if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
+                    return true;
+                }
+            } break;
+    }
+
+    return false;
+}
--- a/Show More
+++ b/Show More