Remove llama.cpp submodule and shift new build to top

---------

Co-authored-by: Bill Wang <bill.wang@bill.wang>

.dockerignore

@@ -3,9 +3,7 @@ ollama
 app
 macapp
 dist
-llm/llama.cpp
 .env
 .cache
 test_data
-llm/build
 llama/build

.gitattributes

@@ -1,4 +1,3 @@
-llm/ext_server/* linguist-vendored
 llama/**/*.cpp linguist-vendored
 llama/**/*.hpp linguist-vendored
 llama/**/*.h linguist-vendored

.github/workflows/release.yaml

@@ -48,8 +48,8 @@ jobs:
         with:
           name: dist-darwin
           path: |
-            dist/*arwin*
-            !dist/*-cov
+            dist/Ollama-darwin.zip
+            dist/ollama-darwin
 
   # Windows builds take a long time to both install the dependencies and build, so parallelize
   # CPU generation step
@@ -92,19 +92,19 @@ jobs:
       - run: go get ./...
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
-          go generate -x ./...
-        name: go generate
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          make -j $cores
+        name: make
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cpu
           path: |
             build/**/*
             build/**/*.a
-            llm/build/**/*.a
             dist/windows-amd64/**
 
   # ROCm generation step
@@ -158,31 +158,21 @@ jobs:
       - run: go get ./...
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          go generate -x ./...
-        name: go generate
-      - name: 'gather rocm dependencies'
-        run: |
-          $HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          md "dist\deps\bin\rocblas\library"
-          cp "${HIP_PATH}\bin\hipblas.dll" "dist\deps\bin\"
-          cp "${HIP_PATH}\bin\rocblas.dll" "dist\deps\bin\"
-          cp "${HIP_PATH}\bin\rocblas\library\*" "dist\deps\bin\rocblas\library\"
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          make -j $cores
+        name: make
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-rocm
           path: |
             build/**/*
             dist/windows-amd64/**
-      - uses: actions/upload-artifact@v4
-        with:
-          name: windows-rocm-deps
-          path: dist/deps/*
 
   # CUDA generation step
   generate-windows-cuda:
@@ -245,34 +235,23 @@ jobs:
       - name: 'Verify CUDA'
         run: nvcc -V
       - run: go get ./...
-      - name: go generate
+      - name: make
         run: |
           $gopath=(get-command go).source | split-path -parent
           $cudabin=(get-command nvcc).source | split-path
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$cudabin;$env:PATH"
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
-          go generate -x ./...
-      - name: 'gather cuda dependencies'
-        run: |
-          $NVIDIA_DIR=(resolve-path 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*\bin\')[0]
-          md "dist\deps"
-          cp "${NVIDIA_DIR}\cudart64_*.dll" "dist\deps\"
-          cp "${NVIDIA_DIR}\cublas64_*.dll" "dist\deps\"
-          cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          make -j $cores
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cuda-${{ matrix.cuda.version }}
           path: |
             build/**/*
             dist/windows-amd64/**
-      - uses: actions/upload-artifact@v4
-        with:
-          name: windows-cuda-deps-${{ matrix.cuda.version }}
-          path: dist/deps/*
-
 
   # windows arm64 generate, go build, and zip file (no installer)
   # Output of this build is aggregated into the final x86 build
@@ -292,6 +271,30 @@ jobs:
           choco install -y --no-progress git gzip
           echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "C:\ProgramData\chocolatey\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      # pacman is buggy on win arm64, so we avoid using it, but rely on the binary artifacts
+      # we download the sfx (7zip bundle) which isn't fully set up, but the binaries we need to build work
+      - name: Install msys2 x64
+        run: |
+          $url="https://github.com/msys2/msys2-installer/releases/download/2024-07-27/msys2-base-x86_64-20240727.sfx.exe"
+          write-host "Downloading MSYS2"
+          Invoke-WebRequest -Uri "$url" -outfile "${env:RUNNER_TEMP}\msys2.exe"
+          write-host "Installing msys2"
+          Start-Process "${env:RUNNER_TEMP}\msys2.exe" -ArgumentList @(
+              '-y', '-oC:\'
+              ) -NoNewWindow -Wait
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      # since pacman isn't reliable, we just download the tar file and extract directly
+      - name: Downloading and extracting msys2 make tar file
+        run: |
+          $url="https://mirror.msys2.org/msys/x86_64/make-4.4.1-2-x86_64.pkg.tar.zst"
+          write-host "Downloading make"
+          Invoke-WebRequest -Uri "$url" -outfile c:\msys64\make.tar.zst
+          cd c:\msys64; tar -xf make.tar.zst
+          rm c:\msys64\make.tar.zst
+      - name: Verify Make works properly
+        run: |
+          echo $env:PATH
+          make --version
       - name: Install Visual Studio 2022
         run: |
           $components = @(
@@ -385,10 +388,9 @@ jobs:
       - run: |
           $gopath=(get-command go).source | split-path -parent
           $gccpath=(get-command gcc).source | split-path -parent
-          & "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$gccpath;$env:PATH;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin"
+          import-module 'C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\Program Files\Microsoft Visual Studio\2022\Community' -skipautomaticlocation
+          $env:PATH="$gopath;$gccpath;$env:PATH"
           echo $env:PATH
           $env:ARCH="arm64"
           .\scripts\build_windows.ps1 buildOllama buildApp gatherDependencies distZip
@@ -455,15 +457,6 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-cuda-12
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-cuda-deps-11
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-cuda-deps-12
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-rocm-deps
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-rocm
@@ -474,11 +467,12 @@ jobs:
       - run: dir build
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_GENERATE="1"
+          $env:ARCH="amd64"
           & .\scripts\build_windows.ps1
       - uses: actions/upload-artifact@v4
         with:

.github/workflows/test.yaml

@@ -21,9 +21,6 @@ jobs:
   changes:
     runs-on: ubuntu-latest
     outputs:
-      GENERATE: ${{ steps.changes.outputs.GENERATE }}
-      GENERATE_CUDA: ${{ steps.changes.outputs.GENERATE_CUDA }}
-      GENERATE_ROCM: ${{ steps.changes.outputs.GENERATE_ROCM }}
       RUNNERS: ${{ steps.changes.outputs.RUNNERS }}
     steps:
       - uses: actions/checkout@v4
@@ -39,53 +36,12 @@ jobs:
           }
 
           {
-            echo GENERATE=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
-            echo GENERATE_CUDA=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
-            echo GENERATE_ROCM=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
             echo RUNNERS=$(changed 'llama/**')
           } >>$GITHUB_OUTPUT
 
-  generate:
+  runners-linux-cuda:
     needs: [changes]
-    if: ${{ needs.changes.outputs.GENERATE == 'True' }}
-    strategy:
-      matrix:
-        os: [ubuntu-latest, macos-latest, windows-2019]
-        arch: [amd64, arm64]
-        exclude:
-          - os: ubuntu-latest
-            arch: arm64
-          - os: windows-2019
-            arch: arm64
-    runs-on: ${{ matrix.os }}
-    env:
-      GOARCH: ${{ matrix.arch }}
-      CGO_ENABLED: '1'
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-go@v5
-        with:
-          go-version-file: go.mod
-          cache: true
-      - run: go get ./...
-      - run: |
-          $gopath=(get-command go).source | split-path -parent
-          $gccpath=(get-command gcc).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$gccpath;$env:PATH"
-          echo $env:PATH
-          go generate -x ./...
-        if: ${{ startsWith(matrix.os, 'windows-') }}
-        name: 'Windows Go Generate'
-      - run: go generate -x ./...
-        if: ${{ ! startsWith(matrix.os, 'windows-') }}
-        name: 'Unix Go Generate'
-      - run: go build .
-  generate-cuda:
-    needs: [changes]
-    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     strategy:
       matrix:
         cuda-version:
@@ -95,8 +51,6 @@ jobs:
     steps:
       - run: |
           apt-get update && apt-get install -y git build-essential curl
-          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.28.1/cmake-3.28.1-linux-x86_64.tar.gz \
-            | tar -zx -C /usr --strip-components 1
         env:
           DEBIAN_FRONTEND: noninteractive
       - uses: actions/checkout@v4
@@ -107,12 +61,11 @@ jobs:
       - run: go get ./...
       - run: |
           git config --global --add safe.directory /__w/ollama/ollama
-          go generate -x ./...
-        env:
-          OLLAMA_SKIP_CPU_GENERATE: '1'
-  generate-rocm:
+          cores=$(grep '^core id' /proc/cpuinfo |sort -u|wc -l)
+          make -j $cores cuda_v11
+  runners-linux-rocm:
     needs: [changes]
-    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     strategy:
       matrix:
         rocm-version:
@@ -122,8 +75,6 @@ jobs:
     steps:
       - run: |
           apt-get update && apt-get install -y git build-essential curl rocm-libs
-          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.28.1/cmake-3.28.1-linux-x86_64.tar.gz \
-            | tar -zx -C /usr --strip-components 1
         env:
           DEBIAN_FRONTEND: noninteractive
       - uses: actions/checkout@v4
@@ -134,14 +85,13 @@ jobs:
       - run: go get ./...
       - run: |
           git config --global --add safe.directory /__w/ollama/ollama
-          go generate -x ./...
-        env:
-          OLLAMA_SKIP_CPU_GENERATE: '1'
+          cores=$(grep '^core id' /proc/cpuinfo |sort -u|wc -l)
+          make -j $cores rocm
 
   # ROCm generation step
-  generate-windows-rocm:
+  runners-windows-rocm:
     needs: [changes]
-    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     runs-on: windows
     steps:
       - uses: actions/checkout@v4
@@ -163,21 +113,21 @@ jobs:
       - run: go get ./...
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          go generate -x ./...
-        name: go generate
-        env:
-          OLLAMA_SKIP_CPU_GENERATE: '1'
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          write-host $env:HIP_PATH
+          make -C llama print-HIP_PATH print-HIP_LIB_DIR
+          make -j $cores rocm
+        name: make
 
   # CUDA generation step
-  generate-windows-cuda:
+  runners-windows-cuda:
     needs: [changes]
-    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     runs-on: windows
     steps:
       - uses: actions/checkout@v4
@@ -202,20 +152,21 @@ jobs:
       - name: 'Verify CUDA'
         run: nvcc -V
       - run: go get ./...
-      - name: go generate
+      - name: make
         run: |
           $gopath=(get-command go).source | split-path -parent
           $cudabin=(get-command nvcc).source | split-path
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$cudabin;$env:PATH"
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
-          go generate -x ./...
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          make -j $cores cuda_v11
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
 
-  runners:
+  runners-cpu:
     needs: [changes]
     if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     strategy:
@@ -244,15 +195,15 @@ jobs:
         run: |
           $gopath=(get-command go).source | split-path -parent
           $gccpath=(get-command gcc).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$gccpath;$env:PATH"
           echo $env:PATH
-          make -C llama -j 4      
+          make -j 4      
       - name: 'Build Unix Go Runners'
         if: ${{ ! startsWith(matrix.os, 'windows-') }}
-        run: make -C llama -j 4
+        run: make -j 4
       - run: go build .
 
   lint:
@@ -302,9 +253,6 @@ jobs:
     env:
       GOARCH: ${{ matrix.arch }}
       CGO_ENABLED: '1'
-      OLLAMA_CPU_TARGET: 'static'
-      OLLAMA_SKIP_CPU_GENERATE: '1'
-      OLLAMA_SKIP_METAL_GENERATE: '1'
     steps:
       - uses: actions/checkout@v4
         with:
@@ -319,7 +267,6 @@ jobs:
             arm64) echo ARCH=arm64 ;;
           esac >>$GITHUB_ENV
         shell: bash
-      - run: go generate ./...
       - run: go build
       - run: go test -v ./...
 
@@ -333,4 +280,4 @@ jobs:
           submodules: recursive
       - name: Verify patches carry all the changes
         run: |
-          cd llama && ./sync.sh && git diff --compact-summary --exit-code .
+          make apply-patches sync && git diff --compact-summary --exit-code llama

.gitignore

@@ -14,4 +14,5 @@ llm/build
 build/*/*/*
 !build/**/placeholder
 llama/build
-__debug_bin*
+__debug_bin*
+llama/vendor

.gitmodules

@@ -1,4 +0,0 @@
-[submodule "llama.cpp"]
-	path = llm/llama.cpp
-	url = https://github.com/ggerganov/llama.cpp.git
-	shallow = true

Dockerfile

@@ -1,3 +1,4 @@
+# Note: once we have fully transitioned to the Go server, this will replace the old Dockerfile at the top of the tree
 ARG GOLANG_VERSION=1.22.5
 ARG CMAKE_VERSION=3.22.1
 ARG CUDA_VERSION_11=11.3.1
@@ -6,168 +7,134 @@ ARG CUDA_VERSION_12=12.4.0
 ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
 ARG ROCM_VERSION=6.1.2
 
-# Copy the minimal context we need to run the generate scripts
-FROM scratch AS llm-code
-COPY .git .git
-COPY .gitmodules .gitmodules
-COPY llm llm
-
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_11-devel-centos7 AS cuda-11-build-amd64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V11_ARCHITECTURES
-ENV GOARCH=amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
-    CUDA_VARIANT="_v11" \
-    bash gen_linux.sh
-
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_12-devel-centos7 AS cuda-12-build-amd64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V12_ARCHITECTURES
-ENV GOARCH=amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
-    CUDA_VARIANT="_v12" \
-    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
-    bash gen_linux.sh
-
-FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_11-devel-rockylinux8 AS cuda-11-build-runner-arm64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V11_ARCHITECTURES
-ENV GOARCH=arm64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
-    CUDA_VARIANT="_v11" \
-    bash gen_linux.sh
-
-FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_12-devel-rockylinux8 AS cuda-12-build-runner-arm64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V12_ARCHITECTURES
-ENV GOARCH=arm64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
-    CUDA_VARIANT="_v12" \
-    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
-    bash gen_linux.sh
-
-
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
-ENV LIBRARY_PATH=/opt/amdgpu/lib64
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG AMDGPU_TARGETS
-ENV GOARCH=amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
-RUN mkdir -p ../../dist/linux-amd64-rocm/lib/ollama && \
-    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64-rocm/lib/ollama && tar xf - )
-
-FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
+### To create a local image for building linux binaries on mac or windows with efficient incremental builds
+#
+# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile --target unified-builder-amd64 .
+# docker run --platform linux/amd64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-amd64
+#
+### Then incremental builds will be much faster in this container
+#
+# make -C llama -j 10 && go build -trimpath -o dist/linux-amd64/ollama .
+#
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
 COPY ./scripts/rh_linux_deps.sh /
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:/usr/local/cuda/bin:$PATH
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-ARG OLLAMA_CUSTOM_CPU_DEFS
-ARG CGO_CFLAGS
-ENV GOARCH=amd64
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+# TODO intel oneapi goes here...
+ENV GOARCH amd64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]
 
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" bash gen_linux.sh
-
-FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
+### To create a local image for building linux binaries on mac or linux/arm64 with efficient incremental builds
+# Note: this does not contain jetson variants
+#
+# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile --target unified-builder-arm64 .
+# docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
+#
+FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-ARG OLLAMA_CUSTOM_CPU_DEFS
-ARG CGO_CFLAGS
-ENV GOARCH=arm64
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
+    dnf config-manager --set-enabled appstream && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH:/usr/local/cuda/bin
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
+ENV GOARCH amd64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]
 
-FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
+FROM --platform=linux/amd64 unified-builder-amd64 AS runners-amd64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_CUDA_11_GENERATE
+ARG OLLAMA_SKIP_CUDA_12_GENERATE
+ARG OLLAMA_SKIP_ROCM_GENERATE
+ARG CUDA_V11_ARCHITECTURES
+ARG CUDA_V12_ARCHITECTURES
+ARG OLLAMA_FAST_BUILD
 RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
+    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
+        make -C llama -j $(expr $(nproc) / 2 ) ; \
+    else \
+        make -C llama -j 5 ; \
+    fi
+
+FROM --platform=linux/arm64 unified-builder-arm64 AS runners-arm64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_CUDA_11_GENERATE
+ARG OLLAMA_SKIP_CUDA_12_GENERATE
+ARG CUDA_V11_ARCHITECTURES
+ARG CUDA_V12_ARCHITECTURES
+ARG OLLAMA_FAST_BUILD
+RUN --mount=type=cache,target=/root/.ccache \
+    make -C llama -j 8
 
 
 # Intermediate stages used for ./scripts/build_linux.sh
-FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
-ENV CGO_ENABLED=1
+FROM --platform=linux/amd64 centos:7 AS builder-amd64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV CGO_ENABLED 1
+ENV GOARCH amd64
 WORKDIR /go/src/github.com/ollama/ollama
+
+FROM --platform=linux/amd64 builder-amd64 AS build-amd64
 COPY . .
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
+ARG OLLAMA_SKIP_ROCM_GENERATE
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-amd64/bin/ollama .
 RUN cd dist/linux-$GOARCH && \
     tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
-RUN cd dist/linux-$GOARCH-rocm && \
-    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz
+RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
+    cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
+    fi
 
-FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
-ENV CGO_ENABLED=1
+FROM --platform=linux/arm64 rockylinux:8 AS builder-arm64
+ARG CMAKE_VERSION
 ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+ENV CGO_ENABLED 1
+ENV GOARCH arm64
 WORKDIR /go/src/github.com/ollama/ollama
+
+FROM --platform=linux/arm64 builder-arm64 AS build-arm64
 COPY . .
-COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
@@ -179,11 +146,11 @@ FROM --platform=linux/amd64 scratch AS dist-amd64
 COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
 FROM --platform=linux/arm64 scratch AS dist-arm64
 COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
-FROM dist-$TARGETARCH as dist
+FROM dist-$TARGETARCH AS dist
 
 
 # Optimized container images do not cary nested payloads
-FROM --platform=linux/amd64 cpu-builder-amd64 AS container-build-amd64
+FROM --platform=linux/amd64 builder-amd64 AS container-build-amd64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 ARG GOFLAGS
@@ -191,7 +158,7 @@ ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-amd64/bin/ollama .
 
-FROM --platform=linux/arm64 cpu-builder-arm64 AS container-build-arm64
+FROM --platform=linux/arm64 builder-arm64 AS container-build-arm64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 ARG GOFLAGS
@@ -199,48 +166,52 @@ ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-arm64/bin/ollama .
 
+# For amd64 container images, filter out cuda/rocm to minimize size
+FROM runners-amd64 AS runners-cuda-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_hipblas.so \
+    ./dist/linux-amd64/lib/ollama/runners/rocm*
+
+FROM runners-amd64 AS runners-rocm-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_cuda*.so \
+    ./dist/linux-amd64/lib/ollama/libcu*.so* \
+    ./dist/linux-amd64/lib/ollama/runners/cuda*
+
 FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
-    apt-get clean && rm -rf /var/lib/apt/lists/*
+    rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-COPY --from=cpu-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
 
 FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
-    apt-get clean && rm -rf /var/lib/apt/lists/*
+    rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
-COPY --from=cpu-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
-COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
-COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
 
 # ROCm libraries larger so we keep it distinct from the CPU/CUDA image
 FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
 # Frontload the rocm libraries which are large, and rarely change to increase chance of a common layer
 # across releases
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
-    apt-get clean && rm -rf /var/lib/apt/lists/*
+    rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-COPY --from=cpu-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
 EXPOSE 11434
-ENV OLLAMA_HOST=0.0.0.0
+ENV OLLAMA_HOST 0.0.0.0
 
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]
 
 FROM runtime-$TARGETARCH
 EXPOSE 11434
-ENV OLLAMA_HOST=0.0.0.0
+ENV OLLAMA_HOST 0.0.0.0
 ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility

Makefile

@@ -0,0 +1,4 @@
+GOALS := $(or $(MAKECMDGOALS),all)
+.PHONY: $(GOALS)
+$(GOALS):
+	$(MAKE) -C llama $@

README.md

@@ -412,6 +412,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [High-level function abstraction in Go](https://gitlab.com/tozd/go/fun)
 - [Ollama PHP](https://github.com/ArdaGnsrn/ollama-php)
 - [Agents-Flex for Java](https://github.com/agents-flex/agents-flex) with [example](https://github.com/agents-flex/agents-flex/tree/main/agents-flex-llm/agents-flex-llm-ollama/src/test/java/com/agentsflex/llm/ollama)
+- [Ollama for Swift](https://github.com/mattt/ollama-swift)
 
 ### Mobile
 

cmd/cmd.go

@@ -21,7 +21,6 @@ import (
 	"path/filepath"
 	"regexp"
 	"runtime"
-	"slices"
 	"strconv"
 	"strings"
 	"sync/atomic"
@@ -47,28 +46,58 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
+var (
+	errModelNotFound     = errors.New("no Modelfile or safetensors files found")
+	errModelfileNotFound = errors.New("specified Modelfile wasn't found")
+)
+
+func getModelfileName(cmd *cobra.Command) (string, error) {
+	fn, _ := cmd.Flags().GetString("file")
+
+	filename := fn
+	if filename == "" {
+		filename = "Modelfile"
+	}
+
+	absName, err := filepath.Abs(filename)
+	if err != nil {
+		return "", err
+	}
+
+	_, err = os.Stat(absName)
+	if err != nil {
+		return fn, err
+	}
+
+	return absName, nil
+}
+
 func CreateHandler(cmd *cobra.Command, args []string) error {
-	filename, _ := cmd.Flags().GetString("file")
-	filename, err := filepath.Abs(filename)
-	if err != nil {
-		return err
-	}
-
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
-	f, err := os.Open(filename)
-	if err != nil {
-		return err
-	}
-	defer f.Close()
+	var reader io.Reader
 
-	modelfile, err := parser.ParseFile(f)
+	filename, err := getModelfileName(cmd)
+	if os.IsNotExist(err) {
+		if filename == "" {
+			reader = strings.NewReader("FROM .\n")
+		} else {
+			return errModelfileNotFound
+		}
+	} else if err != nil {
+		return err
+	} else {
+		f, err := os.Open(filename)
+		if err != nil {
+			return err
+		}
+
+		reader = f
+		defer f.Close()
+	}
+
+	modelfile, err := parser.ParseFile(reader)
 	if err != nil {
 		return err
 	}
@@ -83,6 +112,11 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p.Add(status, spinner)
 	defer p.Stop()
 
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
 	for i := range modelfile.Commands {
 		switch modelfile.Commands[i].Name {
 		case "model", "adapter":
@@ -221,7 +255,7 @@ func tempZipFiles(path string) (string, error) {
 		// covers consolidated.x.pth, consolidated.pth
 		files = append(files, pt...)
 	} else {
-		return "", errors.New("no safetensors or torch files found")
+		return "", errModelNotFound
 	}
 
 	// add configuration files, json files are detected as text/plain
@@ -453,7 +487,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	opts.MultiModal = slices.Contains(info.Details.Families, "clip")
+	opts.MultiModal = len(info.ProjectorInfo) != 0
 	opts.ParentModel = info.Details.ParentModel
 
 	if interactive {
@@ -1316,7 +1350,7 @@ func NewCLI() *cobra.Command {
 		RunE:    CreateHandler,
 	}
 
-	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile")
+	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\"")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_0)")
 
 	showCmd := &cobra.Command{

cmd/cmd_test.go

@@ -270,3 +270,102 @@ func TestDeleteHandler(t *testing.T) {
 		t.Fatalf("DeleteHandler failed: expected error about stopping non-existent model, got %v", err)
 	}
 }
+
+func TestGetModelfileName(t *testing.T) {
+	tests := []struct {
+		name          string
+		modelfileName string
+		fileExists    bool
+		expectedName  string
+		expectedErr   error
+	}{
+		{
+			name:          "no modelfile specified, no modelfile exists",
+			modelfileName: "",
+			fileExists:    false,
+			expectedName:  "",
+			expectedErr:   os.ErrNotExist,
+		},
+		{
+			name:          "no modelfile specified, modelfile exists",
+			modelfileName: "",
+			fileExists:    true,
+			expectedName:  "Modelfile",
+			expectedErr:   nil,
+		},
+		{
+			name:          "modelfile specified, no modelfile exists",
+			modelfileName: "crazyfile",
+			fileExists:    false,
+			expectedName:  "crazyfile",
+			expectedErr:   os.ErrNotExist,
+		},
+		{
+			name:          "modelfile specified, modelfile exists",
+			modelfileName: "anotherfile",
+			fileExists:    true,
+			expectedName:  "anotherfile",
+			expectedErr:   nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cmd := &cobra.Command{
+				Use: "fakecmd",
+			}
+			cmd.Flags().String("file", "", "path to modelfile")
+
+			var expectedFilename string
+
+			if tt.fileExists {
+				tempDir, err := os.MkdirTemp("", "modelfiledir")
+				defer os.RemoveAll(tempDir)
+				if err != nil {
+					t.Fatalf("temp modelfile dir creation failed: %v", err)
+				}
+				var fn string
+				if tt.modelfileName != "" {
+					fn = tt.modelfileName
+				} else {
+					fn = "Modelfile"
+				}
+
+				tempFile, err := os.CreateTemp(tempDir, fn)
+				if err != nil {
+					t.Fatalf("temp modelfile creation failed: %v", err)
+				}
+
+				expectedFilename = tempFile.Name()
+				err = cmd.Flags().Set("file", expectedFilename)
+				if err != nil {
+					t.Fatalf("couldn't set file flag: %v", err)
+				}
+			} else {
+				if tt.modelfileName != "" {
+					expectedFilename = tt.modelfileName
+					err := cmd.Flags().Set("file", tt.modelfileName)
+					if err != nil {
+						t.Fatalf("couldn't set file flag: %v", err)
+					}
+				}
+			}
+
+			actualFilename, actualErr := getModelfileName(cmd)
+
+			if actualFilename != expectedFilename {
+				t.Errorf("expected filename: '%s' actual filename: '%s'", expectedFilename, actualFilename)
+			}
+
+			if tt.expectedErr != os.ErrNotExist {
+				if actualErr != tt.expectedErr {
+					t.Errorf("expected err: %v actual err: %v", tt.expectedErr, actualErr)
+				}
+			} else {
+				if !os.IsNotExist(actualErr) {
+					t.Errorf("expected err: %v actual err: %v", tt.expectedErr, actualErr)
+				}
+			}
+		})
+	}
+}

cmd/interactive.go

@@ -494,28 +494,22 @@ func buildModelfile(opts runOptions) string {
 }
 
 func normalizeFilePath(fp string) string {
-	// Define a map of escaped characters and their replacements
-	replacements := map[string]string{
-		"\\ ":  " ",  // Escaped space
-		"\\(":  "(",  // Escaped left parenthesis
-		"\\)":  ")",  // Escaped right parenthesis
-		"\\[":  "[",  // Escaped left square bracket
-		"\\]":  "]",  // Escaped right square bracket
-		"\\{":  "{",  // Escaped left curly brace
-		"\\}":  "}",  // Escaped right curly brace
-		"\\$":  "$",  // Escaped dollar sign
-		"\\&":  "&",  // Escaped ampersand
-		"\\;":  ";",  // Escaped semicolon
-		"\\'":  "'",  // Escaped single quote
-		"\\\\": "\\", // Escaped backslash
-		"\\*":  "*",  // Escaped asterisk
-		"\\?":  "?",  // Escaped question mark
-	}
-
-	for escaped, actual := range replacements {
-		fp = strings.ReplaceAll(fp, escaped, actual)
-	}
-	return fp
+	return strings.NewReplacer(
+		"\\ ", " ", // Escaped space
+		"\\(", "(", // Escaped left parenthesis
+		"\\)", ")", // Escaped right parenthesis
+		"\\[", "[", // Escaped left square bracket
+		"\\]", "]", // Escaped right square bracket
+		"\\{", "{", // Escaped left curly brace
+		"\\}", "}", // Escaped right curly brace
+		"\\$", "$", // Escaped dollar sign
+		"\\&", "&", // Escaped ampersand
+		"\\;", ";", // Escaped semicolon
+		"\\'", "'", // Escaped single quote
+		"\\\\", "\\", // Escaped backslash
+		"\\*", "*", // Escaped asterisk
+		"\\?", "?", // Escaped question mark
+	).Replace(fp)
 }
 
 func extractFileNames(input string) []string {
@@ -535,10 +529,9 @@ func extractFileData(input string) (string, []api.ImageData, error) {
 	for _, fp := range filePaths {
 		nfp := normalizeFilePath(fp)
 		data, err := getImageData(nfp)
-		if err != nil {
-			if os.IsNotExist(err) {
-				continue
-			}
+		if errors.Is(err, os.ErrNotExist) {
+			continue
+		} else if err != nil {
 			fmt.Fprintf(os.Stderr, "Couldn't process image: %q\n", err)
 			return "", imgs, err
 		}
@@ -546,7 +539,7 @@ func extractFileData(input string) (string, []api.ImageData, error) {
 		input = strings.ReplaceAll(input, fp, "")
 		imgs = append(imgs, data)
 	}
-	return input, imgs, nil
+	return strings.TrimSpace(input), imgs, nil
 }
 
 func getImageData(filePath string) ([]byte, error) {

convert/convert.go

@@ -9,7 +9,7 @@ import (
 	"log/slog"
 	"strings"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type ModelParameters struct {
@@ -27,8 +27,8 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) llm.KV {
-	kv := llm.KV{
+func (ModelParameters) KV(t *Tokenizer) fileutils.KV {
+	kv := fileutils.KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -54,7 +54,7 @@ func (ModelParameters) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() llm.KV {
+func (p AdapterParameters) KV() fileutils.KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -62,7 +62,7 @@ func (p AdapterParameters) KV() llm.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := llm.KV{
+	kv := fileutils.KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -79,19 +79,19 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
-	return llm.WriteGGUF(ws, kv, ts)
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv fileutils.KV, ts []fileutils.Tensor) error {
+	return fileutils.WriteGGUF(ws, kv, ts)
 }
 
-func (AdapterParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
-	return llm.WriteGGUF(ws, kv, ts)
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv fileutils.KV, ts []fileutils.Tensor) error {
+	return fileutils.WriteGGUF(ws, kv, ts)
 }
 
 type ModelConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) llm.KV
+	KV(*Tokenizer) fileutils.KV
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
-	Tensors([]Tensor) []llm.Tensor
+	Tensors([]Tensor) []fileutils.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
@@ -99,7 +99,7 @@ type ModelConverter interface {
 	// specialTokenTypes returns any special token types the model uses
 	specialTokenTypes() []string
 	// writeFile writes the model to the provided io.WriteSeeker
-	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
+	writeFile(io.WriteSeeker, fileutils.KV, []fileutils.Tensor) error
 }
 
 type moreParser interface {
@@ -108,17 +108,17 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(llm.KV) llm.KV
+	KV(fileutils.KV) fileutils.KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
-	Tensors([]Tensor) []llm.Tensor
+	Tensors([]Tensor) []fileutils.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
 
-	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
+	writeFile(io.WriteSeeker, fileutils.KV, []fileutils.Tensor) error
 }
 
-func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV llm.KV) error {
+func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV fileutils.KV) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err

convert/convert_bert.go

@@ -8,7 +8,7 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type bertModel struct {
@@ -85,7 +85,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *bertModel) KV(t *Tokenizer) llm.KV {
+func (p *bertModel) KV(t *Tokenizer) fileutils.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false
@@ -132,8 +132,8 @@ func (p *bertModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *bertModel) Tensors(ts []Tensor) []fileutils.Tensor {
+	var out []fileutils.Tensor
 	for _, t := range ts {
 		if slices.Contains([]string{
 			"embeddings.position_ids",
@@ -143,7 +143,7 @@ func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
 			continue
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type gemmaModel struct {
@@ -23,7 +23,7 @@ type gemmaModel struct {
 
 var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
+func (p *gemmaModel) KV(t *Tokenizer) fileutils.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings
@@ -42,14 +42,14 @@ func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *gemmaModel) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *gemmaModel) Tensors(ts []Tensor) []fileutils.Tensor {
+	var out []fileutils.Tensor
 	for _, t := range ts {
 		if strings.HasSuffix(t.Name(), "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma2.go

@@ -1,7 +1,7 @@
 package convert
 
 import (
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type gemma2Model struct {
@@ -11,7 +11,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }
 
-func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
+func (p *gemma2Model) KV(t *Tokenizer) fileutils.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2_adapter.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type gemma2Adapter struct {
@@ -15,14 +15,14 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV llm.KV) llm.KV {
+func (p *gemma2Adapter) KV(baseKV fileutils.KV) fileutils.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv
 }
 
-func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *gemma2Adapter) Tensors(ts []Tensor) []fileutils.Tensor {
+	var out []fileutils.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -31,7 +31,7 @@ func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama.go

@@ -9,7 +9,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type llamaModel struct {
@@ -46,7 +46,7 @@ type llamaModel struct {
 
 var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llamaModel) KV(t *Tokenizer) llm.KV {
+func (p *llamaModel) KV(t *Tokenizer) fileutils.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize
@@ -120,11 +120,11 @@ func (p *llamaModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *llamaModel) Tensors(ts []Tensor) []fileutils.Tensor {
+	var out []fileutils.Tensor
 
 	if p.RopeScaling.factors != nil {
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     "rope_freqs.weight",
 			Kind:     0,
 			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
@@ -138,7 +138,7 @@ func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama_adapter.go

@@ -7,7 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type llamaAdapter struct {
@@ -18,7 +18,7 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
+func (p *llamaAdapter) KV(baseKV fileutils.KV) fileutils.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
 	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
@@ -29,8 +29,8 @@ func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
 	return kv
 }
 
-func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *llamaAdapter) Tensors(ts []Tensor) []fileutils.Tensor {
+	var out []fileutils.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -41,7 +41,7 @@ func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    shape,

convert/convert_mixtral.go

@@ -6,7 +6,7 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type mixtralModel struct {
@@ -15,7 +15,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
+func (p *mixtralModel) KV(t *Tokenizer) fileutils.KV {
 	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {
@@ -29,7 +29,7 @@ func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []fileutils.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -56,10 +56,10 @@ func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
 		return true
 	})
 
-	var out []llm.Tensor
+	var out []fileutils.Tensor
 	for n, e := range experts {
 		// TODO(mxyng): sanity check experts
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     n,
 			Kind:     e[0].Kind(),
 			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),

convert/convert_phi3.go

@@ -8,7 +8,7 @@ import (
 	"strings"
 	"sync"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type phi3Model struct {
@@ -37,7 +37,7 @@ type phi3Model struct {
 
 var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3Model) KV(t *Tokenizer) llm.KV {
+func (p *phi3Model) KV(t *Tokenizer) fileutils.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings
@@ -68,19 +68,19 @@ func (p *phi3Model) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
+func (p *phi3Model) Tensors(ts []Tensor) []fileutils.Tensor {
 	var addRopeFactors sync.Once
 
-	out := make([]llm.Tensor, 0, len(ts)+2)
+	out := make([]fileutils.Tensor, 0, len(ts)+2)
 	for _, t := range ts {
 		if strings.HasPrefix(t.Name(), "blk.0.") {
 			addRopeFactors.Do(func() {
-				out = append(out, llm.Tensor{
+				out = append(out, fileutils.Tensor{
 					Name:     "rope_factors_long.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
 					WriterTo: p.RopeScaling.LongFactor,
-				}, llm.Tensor{
+				}, fileutils.Tensor{
 					Name:     "rope_factors_short.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
@@ -89,7 +89,7 @@ func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
 			})
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_test.go

@@ -20,7 +20,7 @@ import (
 
 	"golang.org/x/exp/maps"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type tensorData struct {
@@ -29,7 +29,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }
 
-func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, fileutils.KV, *fileutils.Tensors) {
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -48,7 +48,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
 	}
 	t.Cleanup(func() { r.Close() })
 
-	m, _, err := llm.DecodeGGML(r, math.MaxInt)
+	m, _, err := fileutils.DecodeGGML(r, math.MaxInt)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -60,7 +60,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
 	return r, m.KV(), m.Tensors()
 }
 
-func generateResultsJSON(t *testing.T, f *os.File, kv llm.KV, tensors llm.Tensors) map[string]string {
+func generateResultsJSON(t *testing.T, f *os.File, kv fileutils.KV, tensors *fileutils.Tensors) map[string]string {
 	actual := make(map[string]string)
 	for k, v := range kv {
 		if s, ok := v.(json.Marshaler); !ok {
@@ -330,7 +330,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()
 
-			m, _, err := llm.DecodeGGML(r, math.MaxInt)
+			m, _, err := fileutils.DecodeGGML(r, math.MaxInt)
 			if err != nil {
 				t.Fatal(err)
 			}

discover/README.md

@@ -0,0 +1,3 @@
+# `discover`
+
+This package is responsible for discovering information about the system and the capabilities to run LLM.  This includes GPU and CPU discovery so the optimal runner can be chosen for a given model.  The ollama scheduler relies on up-to-date available memory information, so this package provides the ability to refresh free memory as efficiently as possible.

docs/development.md

@@ -1,183 +1,5 @@
 # Development
 
-> [!IMPORTANT]
-> The `llm` package that loads and runs models is being updated to use a new [Go runner](#transition-to-go-runner): this should only impact a small set of PRs however it does change how the project is built.
-
-Install required tools:
-
-- cmake version 3.24 or higher
-- go version 1.22 or higher
-- gcc version 11.4.0 or higher
-
-### MacOS
-
-```bash
-brew install go cmake gcc
-```
-
-Optionally enable debugging and more verbose logging:
-
-```bash
-# At build time
-export CGO_CFLAGS="-g"
-
-# At runtime
-export OLLAMA_DEBUG=1
-```
-
-Get the required libraries and build the native LLM code:
-
-```bash
-go generate ./...
-```
-
-Then build ollama:
-
-```bash
-go build .
-```
-
-Now you can run `ollama`:
-
-```bash
-./ollama
-```
-
-### Linux
-
-#### Linux CUDA (NVIDIA)
-
-_Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
-
-Install `cmake` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
-development and runtime packages.
-
-Typically the build scripts will auto-detect CUDA, however, if your Linux distro
-or installation approach uses unusual paths, you can specify the location by
-specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
-libraries, and `CUDACXX` to the location of the nvcc compiler. You can customize
-a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
-
-Then generate dependencies:
-
-```
-go generate ./...
-```
-
-Then build the binary:
-
-```
-go build .
-```
-
-#### Linux ROCm (AMD)
-
-_Your operating system distribution may already have packages for AMD ROCm and CLBlast. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
-
-Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/) development packages first, as well as `cmake` and `golang`.
-
-Typically the build scripts will auto-detect ROCm, however, if your Linux distro
-or installation approach uses unusual paths, you can specify the location by
-specifying an environment variable `ROCM_PATH` to the location of the ROCm
-install (typically `/opt/rocm`), and `CLBlast_DIR` to the location of the
-CLBlast install (typically `/usr/lib/cmake/CLBlast`). You can also customize
-the AMD GPU targets by setting AMDGPU_TARGETS (e.g. `AMDGPU_TARGETS="gfx1101;gfx1102"`)
-
-```
-go generate ./...
-```
-
-Then build the binary:
-
-```
-go build .
-```
-
-ROCm requires elevated privileges to access the GPU at runtime. On most distros you can add your user account to the `render` group, or run as root.
-
-#### Advanced CPU Settings
-
-By default, running `go generate ./...` will compile a few different variations
-of the LLM library based on common CPU families and vector math capabilities,
-including a lowest-common-denominator which should run on almost any 64 bit CPU
-somewhat slowly. At runtime, Ollama will auto-detect the optimal variation to
-load. If you would like to build a CPU-based build customized for your
-processor, you can set `OLLAMA_CUSTOM_CPU_DEFS` to the llama.cpp flags you would
-like to use. For example, to compile an optimized binary for an Intel i9-9880H,
-you might use:
-
-```
-OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
-go build .
-```
-
-#### Containerized Linux Build
-
-If you have Docker available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
-
-### Windows
-
-Note: The Windows build for Ollama is still under development.
-
-First, install required tools:
-
-- MSVC toolchain - C/C++ and cmake as minimal requirements
-- Go version 1.22 or higher
-- MinGW (pick one variant) with GCC.
-  - [MinGW-w64](https://www.mingw-w64.org/)
-  - [MSYS2](https://www.msys2.org/)
-- The `ThreadJob` Powershell module: `Install-Module -Name ThreadJob -Scope CurrentUser`
-
-Then, build the `ollama` binary:
-
-```powershell
-$env:CGO_ENABLED="1"
-go generate ./...
-go build .
-```
-
-#### Windows CUDA (NVIDIA)
-
-In addition to the common Windows development tools described above, install CUDA after installing MSVC.
-
-- [NVIDIA CUDA](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html)
-
-
-#### Windows ROCm (AMD Radeon)
-
-In addition to the common Windows development tools described above, install AMDs HIP package after installing MSVC.
-
-- [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
-- [Strawberry Perl](https://strawberryperl.com/)
-
-Lastly, add `ninja.exe` included with MSVC to the system path (e.g. `C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja`).
-
-#### Windows arm64
-
-The default `Developer PowerShell for VS 2022` may default to x86 which is not what you want.  To ensure you get an arm64 development environment, start a plain PowerShell terminal and run:
-
-```powershell
-import-module 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\Common7\\Tools\\Microsoft.VisualStudio.DevShell.dll'
-Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community' -skipautomaticlocation
-```
-
-You can confirm with `write-host $env:VSCMD_ARG_TGT_ARCH`
-
-Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64 msys2 environment.  Ollama requires gcc and mingw32-make to compile, which is not currently available on Windows arm64, but a gcc compatibility adapter is available via `mingw-w64-clang-aarch64-gcc-compat`. At a minimum you will need to install the following:
-
-```
-pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
-```
-
-You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)
-
-
-## Transition to Go runner
-
-The Ollama team is working on moving to a new Go based runner that loads and runs models in a subprocess to replace the previous code under `ext_server`. During this transition period, this new Go runner is "opt in" at build time, and requires using a different approach to build.
-
-After the transition to use the Go server exclusively, both `make` and `go generate` will build the Go runner.
-
 Install required tools:
 
 - go version 1.22 or higher
@@ -201,7 +23,7 @@ export OLLAMA_DEBUG=1
 Get the required libraries and build the native LLM code:  (Adjust the job count based on your number of processors for a faster build)
 
 ```bash
-make -C llama -j 5
+make -j 5
 ```
 
 Then build ollama:
@@ -238,7 +60,7 @@ a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "
 Then generate dependencies:  (Adjust the job count based on your number of processors for a faster build)
 
 ```
-make -C llama -j 5
+make -j 5
 ```
 
 Then build the binary:
@@ -263,7 +85,7 @@ the AMD GPU targets by setting AMDGPU_TARGETS (e.g. `AMDGPU_TARGETS="gfx1101;gfx
 Then generate dependencies:  (Adjust the job count based on your number of processors for a faster build)
 
 ```
-make -C llama -j 5
+make -j 5
 ```
 
 Then build the binary:
@@ -305,7 +127,7 @@ Then, build the `ollama` binary:
 
 ```powershell
 $env:CGO_ENABLED="1"
-make -C llama -j 8
+make -j 8
 go build .
 ```
 

docs/openai.md

@@ -37,7 +37,7 @@ response = client.chat.completions.create(
                 {"type": "text", "text": "What's in this image?"},
                 {
                     "type": "image_url",
-                    "image_url": "iVBORw0KGgoAAAANSUhEUgAAAG0AAABmCAYAAADBPx+VAAAACXBIWXMAAAsTAAALEwEAmpwYAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAA3VSURBVHgB7Z27r0zdG8fX743i1bi1ikMoFMQloXRpKFFIqI7LH4BEQ+NWIkjQuSWCRIEoULk0gsK1kCBI0IhrQVT7tz/7zZo888yz1r7MnDl7z5xvsjkzs2fP3uu71nNfa7lkAsm7d++Sffv2JbNmzUqcc8m0adOSzZs3Z+/XES4ZckAWJEGWPiCxjsQNLWmQsWjRIpMseaxcuTKpG/7HP27I8P79e7dq1ars/yL4/v27S0ejqwv+cUOGEGGpKHR37tzJCEpHV9tnT58+dXXCJDdECBE2Ojrqjh071hpNECjx4cMHVycM1Uhbv359B2F79+51586daxN/+pyRkRFXKyRDAqxEp4yMlDDzXG1NPnnyJKkThoK0VFd1ELZu3TrzXKxKfW7dMBQ6bcuWLW2v0VlHjx41z717927ba22U9APcw7Nnz1oGEPeL3m3p2mTAYYnFmMOMXybPPXv2bNIPpFZr1NHn4HMw0KRBjg9NuRw95s8PEcz/6DZELQd/09C9QGq5RsmSRybqkwHGjh07OsJSsYYm3ijPpyHzoiacg35MLdDSIS/O1yM778jOTwYUkKNHWUzUWaOsylE00MyI0fcnOwIdjvtNdW/HZwNLGg+sR1kMepSNJXmIwxBZiG8tDTpEZzKg0GItNsosY8USkxDhD0Rinuiko2gfL/RbiD2LZAjU9zKQJj8RDR0vJBR1/Phx9+PHj9Z7REF4nTZkxzX4LCXHrV271qXkBAPGfP/atWvu/PnzHe4C97F48eIsRLZ9+3a3f/9+87dwP1JxaF7/3r17ba+5l4EcaVo0lj3SBq5kGTJSQmLWMjgYNei2GPT1MuMqGTDEFHzeQSP2wi/jGnkmPJ/nhccs44jvDAxpVcxnq0F6eT8h4ni/iIWpR5lPyA6ETkNXoSukvpJAD3AsXLiwpZs49+fPn5ke4j10TqYvegSfn0OnafC+Tv9ooA/JPkgQysqQNBzagXY55nO/oa1F7qvIPWkRL12WRpMWUvpVDYmxAPehxWSe8ZEXL20sadYIozfmNch4QJPAfeJgW3rNsnzphBKNJM2KKODo1rVOMRYik5ETy3ix4qWNI81qAAirizgMIc+yhTytx0JWZuNI03qsrgWlGtwjoS9XwgUhWGyhUaRZZQNNIEwCiXD16tXcAHUs79co0vSD8rrJCIW98pzvxpAWyyo3HYwqS0+H0BjStClcZJT5coMm6D2LOF8TolGJtK9fvyZpyiC5ePFi9nc/oJU4eiEP0jVoAnHa9wyJycITMP78+eMeP37sXrx44d6+fdt6f82aNdkx1pg9e3Zb5W+RSRE+n+VjksQWifvVaTKFhn5O8my63K8Qabdv33b379/PiAP//vuvW7BggZszZ072/+TJk91YgkafPn166zXB1rQHFvouAWHq9z3SEevSUerqCn2/dDCeta2jxYbr69evk4MHDyY7d+7MjhMnTiTPnz9Pfv/+nfQT2ggpO2dMF8cghuoM7Ygj5iWCqRlGFml0QC/ftGmTmzt3rmsaKDsgBSPh0/8yPeLLBihLkOKJc0jp8H8vUzcxIA1k6QJ/c78tWEyj5P3o4u9+jywNPdJi5rAH9x0KHcl4Hg570eQp3+vHXGyrmEeigzQsQsjavXt38ujRo44LQuDDhw+TW7duRS1HGgMxhNXHgflaNTOsHyKvHK5Ijo2jbFjJBQK9YwFd6RVMzfgRBmEfP37suBBm/p49e1qjEP2mwTViNRo0VJWH1deMXcNK08uUjVUu7s/zRaL+oLNxz1bpANco4npUgX4G2eFbpDFyQoQxojBCpEGSytmOH8qrH5Q9vuzD6ofQylkCUmh8DBAr+q8JCyVNtWQIidKQE9wNtLSQnS4jDSsxNHogzFuQBw4cyM61UKVsjfr3ooBkPSqqQHesUPWVtzi9/vQi1T+rJj7WiTz4Pt/l3LxUkr5P2VYZaZ4URpsE+st/dujQoaBBYokbrz/8TJNQYLSonrPS9kUaSkPeZyj1AWSj+d+VBoy1pIWVNed8P0Ll/ee5HdGRhrHhR5GGN0r4LGZBaj8oFDJitBTJzIZgFcmU0Y8ytWMZMzJOaXUSrUs5RxKnrxmbb5YXO9VGUhtpXldhEUogFr3IzIsvlpmdosVcGVGXFWp2oU9kLFL3dEkSz6NHEY1sjSRdIuDFWEhd8KxFqsRi1uM/nz9/zpxnwlESONdg6dKlbsaMGS4EHFHtjFIDHwKOo46l4TxSuxgDzi+rE2jg+BaFruOX4HXa0Nnf1lwAPufZeF8/r6zD97WK2qFnGjBxTw5qNGPxT+5T/r7/7RawFC3j4vTp09koCxkeHjqbHJqArmH5UrFKKksnxrK7FuRIs8STfBZv+luugXZ2pR/pP9Ois4z+TiMzUUkUjD0iEi1fzX8GmXyuxUBRcaUfykV0YZnlJGKQpOiGB76x5GeWkWWJc3mOrK6S7xdND+W5N6XyaRgtWJFe13GkaZnKOsYqGdOVVVbGupsyA/l7emTLHi7vwTdirNEt0qxnzAvBFcnQF16xh/TMpUuXHDowhlA9vQVraQhkudRdzOnK+04ZSP3DUhVSP61YsaLtd/ks7ZgtPcXqPqEafHkdqa84X6aCeL7YWlv6edGFHb+ZFICPlljHhg0bKuk0CSvVznWsotRu433alNdFrqG45ejoaPCaUkWERpLXjzFL2Rpllp7PJU2a/v7Ab8N05/9t27Z16KUqoFGsxnI9EosS2niSYg9SpU6B4JgTrvVW1flt1sT+0ADIJU2maXzcUTraGCRaL1Wp9rUMk16PMom8QhruxzvZIegJjFU7LLCePfS8uaQdPny4jTTL0dbee5mYokQsXTIWNY46kuMbnt8Kmec+LGWtOVIl9cT1rCB0V8WqkjAsRwta93TbwNYoGKsUSChN44lgBNCoHLHzquYKrU6qZ8lolCIN0Rh6cP0Q3U6I6IXILYOQI513hJaSKAorFpuHXJNfVlpRtmYBk1Su1obZr5dnKAO+L10Hrj3WZW+E3qh6IszE37F6EB+68mGpvKm4eb9bFrlzrok7fvr0Kfv727dvWRmdVTJHw0qiiCUSZ6wCK+7XL/AcsgNyL74DQQ730sv78Su7+t/A36MdY0sW5o40ahslXr58aZ5HtZB8GH64m9EmMZ7FpYw4T6QnrZfgenrhFxaSiSGXtPnz57e9TkNZLvTjeqhr734CNtrK41L40sUQckmj1lGKQ0rC37x544r8eNXRpnVE3ZZY7zXo8NomiO0ZUCj2uHz58rbXoZ6gc0uA+F6ZeKS/jhRDUq8MKrTho9fEkihMmhxtBI1DxKFY9XLpVcSkfoi8JGnToZO5sU5aiDQIW716ddt7ZLYtMQlhECdBGXZZMWldY5BHm5xgAroWj4C0hbYkSc/jBmggIrXJWlZM6pSETsEPGqZOndr2uuuR5rF169a2HoHPdurUKZM4CO1WTPqaDaAd+GFGKdIQkxAn9RuEWcTRyN2KSUgiSgF5aWzPTeA/lN5rZubMmR2bE4SIC4nJoltgAV/dVefZm72AtctUCJU2CMJ327hxY9t7EHbkyJFseq+EJSY16RPo3Dkq1kkr7+q0bNmyDuLQcZBEPYmHVdOBiJyIlrRDq41YPWfXOxUysi5fvtyaj+2BpcnsUV/oSoEMOk2CQGlr4ckhBwaetBhjCwH0ZHtJROPJkyc7UjcYLDjmrH7ADTEBXFfOYmB0k9oYBOjJ8b4aOYSe7QkKcYhFlq3QYLQhSidNmtS2RATwy8YOM3EQJsUjKiaWZ+vZToUQgzhkHXudb/PW5YMHD9yZM2faPsMwoc7RciYJXbGuBqJ1UIGKKLv915jsvgtJxCZDubdXr165mzdvtr1Hz5LONA8jrUwKPqsmVesKa49S3Q4WxmRPUEYdTjgiUcfUwLx589ySJUva3oMkP6IYddq6HMS4o55xBJBUeRjzfa4Zdeg56QZ43LhxoyPo7Lf1kNt7oO8wWAbNwaYjIv5lhyS7kRf96dvm5Jah8vfvX3flyhX35cuX6HfzFHOToS1H4BenCaHvO8pr8iDuwoUL7tevX+b5ZdbBair0xkFIlFDlW4ZknEClsp/TzXyAKVOmmHWFVSbDNw1l1+4f90U6IY/q4V27dpnE9bJ+v87QEydjqx/UamVVPRG+mwkNTYN+9tjkwzEx+atCm/X9WvWtDtAb68Wy9LXa1UmvCDDIpPkyOQ5ZwSzJ4jMrvFcr0rSjOUh+GcT4LSg5ugkW1Io0/SCDQBojh0hPlaJdah+tkVYrnTZowP8iq1F1TgMBBauufyB33x1v+NWFYmT5KmppgHC+NkAgbmRkpD3yn9QIseXymoTQFGQmIOKTxiZIWpvAatenVqRVXf2nTrAWMsPnKrMZHz6bJq5jvce6QK8J1cQNgKxlJapMPdZSR64/UivS9NztpkVEdKcrs5alhhWP9NeqlfWopzhZScI6QxseegZRGeg5a8C3Re1Mfl1ScP36ddcUaMuv24iOJtz7sbUjTS4qBvKmstYJoUauiuD3k5qhyr7QdUHMeCgLa1Ear9NquemdXgmum4fvJ6w1lqsuDhNrg1qSpleJK7K3TF0Q2jSd94uSZ60kK1e3qyVpQK6PVWXp2/FC3mp6jBhKKOiY2h3gtUV64TWM6wDETRPLDfSakXmH3w8g9Jlug8ZtTt4kVF0kLUYYmCCtD/DrQ5YhMGbA9L3ucdjh0y8kOHW5gU/VEEmJTcL4Pz/f7mgoAbYkAAAAAElFTkSuQmCC",
+                    "image_url": "data:image/png;base64,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",
                 },
             ],
         }
@@ -86,7 +86,7 @@ const response = await openai.chat.completions.create({
             { type: "text", text: "What's in this image?" },
             {
             type: "image_url",
-            image_url: "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",
+            image_url: "data:image/png;base64,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",
             },
         ],
         },
@@ -142,7 +142,7 @@ curl http://localhost:11434/v1/chat/completions \
           {
             "type": "image_url",
             "image_url": {
-               "url": "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"
+               "url": "data:image/png;base64,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"
             }
           }
         ]

envconfig/config.go

@@ -72,6 +72,7 @@ func Origins() (origins []string) {
 		"app://*",
 		"file://*",
 		"tauri://*",
+		"vscode-webview://*",
 	)
 
 	return origins

envconfig/config_test.go

@@ -68,6 +68,7 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
+			"vscode-webview://*",
 		}},
 		{"http://10.0.0.1", []string{
 			"http://10.0.0.1",
@@ -86,6 +87,7 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
+			"vscode-webview://*",
 		}},
 		{"http://172.16.0.1,https://192.168.0.1", []string{
 			"http://172.16.0.1",
@@ -105,6 +107,7 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
+			"vscode-webview://*",
 		}},
 		{"http://totally.safe,http://definitely.legit", []string{
 			"http://totally.safe",
@@ -124,6 +127,7 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
+			"vscode-webview://*",
 		}},
 	}
 	for _, tt := range cases {

fileutils/README.md

@@ -0,0 +1,3 @@
+# `modelfile`
+
+This package provides utilities for loading and inspecting model files

fileutils/filetype.go

@@ -1,9 +1,11 @@
-package llm
+package fileutils
 
 import "fmt"
 
 type fileType uint32
 
+// TODO this should map over to the GGML CGO enum type
+
 const (
 	fileTypeF32 fileType = iota
 	fileTypeF16

fileutils/ggla.go

@@ -1,4 +1,4 @@
-package llm
+package fileutils
 
 import (
 	"encoding/binary"
@@ -51,8 +51,8 @@ func (llm *ggla) KV() KV {
 	return llm.kv
 }
 
-func (llm *ggla) Tensors() Tensors {
-	return Tensors{
+func (llm *ggla) Tensors() *Tensors {
+	return &Tensors{
 		Items:  llm.tensors,
 		Offset: llm.tensorOffset,
 	}

fileutils/ggml.go

@@ -1,11 +1,14 @@
-package llm
+package fileutils
 
 import (
 	"encoding/binary"
 	"errors"
 	"fmt"
 	"io"
+	"os"
+	"slices"
 	"strings"
+	"sync"
 
 	"github.com/ollama/ollama/util/bufioutil"
 )
@@ -17,7 +20,7 @@ type GGML struct {
 
 type model interface {
 	KV() KV
-	Tensors() Tensors
+	Tensors() *Tensors
 }
 
 type KV map[string]any
@@ -123,25 +126,34 @@ func (kv KV) ChatTemplate() string {
 type Tensors struct {
 	Items  []*Tensor
 	Offset uint64
+
+	layers     map[string]Layer
+	layersOnce sync.Once
 }
 
-func (ts Tensors) Layers() map[string]Layer {
-	layers := make(map[string]Layer)
-	for _, t := range ts.Items {
-		parts := strings.Split(t.Name, ".")
-		if parts[0] == "blk" {
-			// join first and second part, e.g. blk.%d
-			parts = append([]string{fmt.Sprintf("%s.%s", parts[0], parts[1])}, parts[2:]...)
+func (ts *Tensors) Layers() map[string]Layer {
+	ts.layersOnce.Do(func() {
+		ts.layers = make(map[string]Layer)
+		for _, t := range ts.Items {
+			parts := strings.Split(t.Name, ".")
+			if index := slices.IndexFunc(parts, func(s string) bool { return s == "blk" || s == "mm" }); index != -1 {
+				if len(parts) > index+2 {
+					// blk and mm should have a number after them, join it
+					parts = append(
+						[]string{strings.Join(parts[:index+2], ".")},
+						parts[index+2:]...)
+				}
+			}
+
+			if _, ok := ts.layers[parts[0]]; !ok {
+				ts.layers[parts[0]] = make(Layer)
+			}
+
+			ts.layers[parts[0]][strings.Join(parts[1:], ".")] = t
 		}
+	})
 
-		if _, ok := layers[parts[0]]; !ok {
-			layers[parts[0]] = make(Layer)
-		}
-
-		layers[parts[0]][strings.Join(parts[1:], ".")] = t
-	}
-
-	return layers
+	return ts.layers
 }
 
 type Layer map[string]*Tensor
@@ -477,3 +489,23 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 
 	return
 }
+
+// LoadModel will load a model from disk. The model must be in the GGML format.
+//
+// It collects array values for arrays with a size less than or equal to
+// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
+// the maxArraySize is negative, all arrays are collected.
+func LoadModel(model string, maxArraySize int) (*GGML, error) {
+	if _, err := os.Stat(model); err != nil {
+		return nil, err
+	}
+
+	f, err := os.Open(model)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	ggml, _, err := DecodeGGML(f, maxArraySize)
+	return ggml, err
+}

fileutils/ggml_test.go

@@ -0,0 +1 @@
+package fileutils

fileutils/gguf.go

@@ -1,4 +1,4 @@
-package llm
+package fileutils
 
 import (
 	"bytes"
@@ -110,8 +110,8 @@ func (llm *gguf) KV() KV {
 	return llm.kv
 }
 
-func (llm *gguf) Tensors() Tensors {
-	return Tensors{
+func (llm *gguf) Tensors() *Tensors {
+	return &Tensors{
 		Items:  llm.tensors,
 		Offset: llm.tensorOffset,
 	}

fileutils/memory.go

@@ -1,8 +1,9 @@
-package llm
+package fileutils
 
 import (
 	"fmt"
 	"log/slog"
+	"os"
 	"strconv"
 	"strings"
 
@@ -63,6 +64,8 @@ type MemoryEstimate struct {
 	memoryLayerOutput   uint64
 	graphFullOffload    uint64
 	graphPartialOffload uint64
+
+	projectorWeights, projectorGraph uint64
 }
 
 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
@@ -78,7 +81,8 @@ func EstimateGPULayers(gpus []discover.GpuInfo, ggml *GGML, projectors []string,
 	var graphOffload uint64
 
 	// Projectors loaded into GPU0 only
-	var projectorSize uint64
+	var projectorWeights uint64
+	var projectorGraph uint64
 
 	// Conditional output size on GPU 0
 	var memoryLayerOutput uint64
@@ -103,7 +107,9 @@ func EstimateGPULayers(gpus []discover.GpuInfo, ggml *GGML, projectors []string,
 	slog.Debug("evaluating", "library", gpus[0].Library, "gpu_count", len(gpus), "available", availableList)
 
 	for _, projector := range projectors {
-		projectorSize += projectorMemoryRequirements(projector)
+		weight, graph := projectorMemoryRequirements(projector)
+		projectorWeights += weight
+		projectorGraph += graph
 
 		// multimodal models require at least 2048 context
 		opts.NumCtx = max(opts.NumCtx, 2048)
@@ -149,7 +155,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, ggml *GGML, projectors []string,
 	}
 
 	// Output layer handled at the end if we have space
-	gpuZeroOverhead := projectorSize
+	gpuZeroOverhead := projectorWeights + projectorGraph
 
 	// Reduce set of GPUs to only those that have sufficient space to fit overhead and at least one layer
 	var layerCount int
@@ -303,6 +309,8 @@ func EstimateGPULayers(gpus []discover.GpuInfo, ggml *GGML, projectors []string,
 		memoryLayerOutput:   memoryLayerOutput,
 		graphFullOffload:    graphFullOffload,
 		graphPartialOffload: graphPartialOffload,
+		projectorWeights:    projectorWeights,
+		projectorGraph:      projectorGraph,
 	}
 
 	if gpus[0].Library == "cpu" {
@@ -321,9 +329,21 @@ func EstimateGPULayers(gpus []discover.GpuInfo, ggml *GGML, projectors []string,
 	return estimate
 }
 
-func (m MemoryEstimate) log() {
+func (m MemoryEstimate) Log() {
 	overhead := envconfig.GpuOverhead()
-	slog.Info(
+
+	log := slog.With()
+	if m.projectorWeights > 0 {
+		log = log.With(
+			slog.Group(
+				"projector",
+				"weights", format.HumanBytes2(m.projectorWeights),
+				"graph", format.HumanBytes2(m.projectorGraph),
+			),
+		)
+	}
+
+	log.Info(
 		"offload to "+m.inferenceLibrary,
 		slog.Group(
 			"layers",
@@ -371,3 +391,52 @@ func (m MemoryEstimate) log() {
 		),
 	)
 }
+
+func projectorMemoryRequirements(filename string) (weights, graphSize uint64) {
+	file, err := os.Open(filename)
+	if err != nil {
+		return 0, 0
+	}
+	defer file.Close()
+
+	ggml, _, err := DecodeGGML(file, 0)
+	if err != nil {
+		return 0, 0
+	}
+
+	for _, layer := range ggml.Tensors().Layers() {
+		weights += layer.size()
+	}
+
+	switch arch := ggml.KV().Architecture(); arch {
+	case "mllama":
+		kv := func(n string) uint64 {
+			if v, ok := ggml.KV()[arch+".vision."+n].(uint32); ok {
+				return uint64(v)
+			}
+
+			return 0
+		}
+
+		imageSize := kv("image_size")
+
+		maxNumTiles := kv("max_num_tiles")
+		embeddingLength := kv("embedding_length")
+		headCount := kv("attention.head_count")
+
+		numPatches := (imageSize / kv("patch_size")) * (imageSize / kv("patch_size"))
+		if _, ok := ggml.Tensors().Layers()["v"]["class_embd"]; ok {
+			numPatches++
+		}
+
+		numPaddedPatches := numPatches + 8 - (numPatches%8)%8
+
+		graphSize = 4 * (8 +
+			imageSize*imageSize*kv("num_channels")*maxNumTiles +
+			embeddingLength*numPatches*maxNumTiles +
+			9*embeddingLength*numPaddedPatches*maxNumTiles +
+			numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
+	}
+
+	return weights, graphSize
+}

fileutils/memory_test.go

@@ -1,4 +1,4 @@
-package llm
+package fileutils
 
 import (
 	"bytes"

go.mod

@@ -22,6 +22,7 @@ require (
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
+	golang.org/x/image v0.14.0
 )
 
 require (

go.sum

@@ -230,6 +230,8 @@ golang.org/x/image v0.0.0-20200430140353-33d19683fad8/go.mod h1:FeLwcggjj3mMvU+o
 golang.org/x/image v0.0.0-20200618115811-c13761719519/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
 golang.org/x/image v0.0.0-20201208152932-35266b937fa6/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
 golang.org/x/image v0.0.0-20210216034530-4410531fe030/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/image v0.14.0 h1:tNgSxAFe3jC4uYqvZdTr84SZoM1KfwdC9SKIFrLjFn4=
+golang.org/x/image v0.14.0/go.mod h1:HUYqC05R2ZcZ3ejNQsIHQDQiwWM4JBqmm6MKANTp4LE=
 golang.org/x/lint v0.0.0-20181026193005-c67002cb31c3/go.mod h1:UVdnD1Gm6xHRNCYTkRU2/jEulfH38KcIWyp/GAMgvoE=
 golang.org/x/lint v0.0.0-20190227174305-5b3e6a55c961/go.mod h1:wehouNa3lNwaWXcvxsM5YxQ5yQlVC4a0KAMCusXpPoU=
 golang.org/x/lint v0.0.0-20190313153728-d0100b6bd8b3/go.mod h1:6SW0HCj/g11FgYtHlgUYUwCkIfeOF89ocIRzGO/8vkc=

integration/basic_test.go

@@ -30,6 +30,22 @@ func TestOrcaMiniBlueSky(t *testing.T) {
 	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
 }
 
+func TestUnicodeOutput(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",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	GenerateTestHelper(ctx, t, req, []string{"😀", "😊", "😁", "😂", "😄", "😃"})
+}
+
 func TestUnicodeModelDir(t *testing.T) {
 	// This is only useful for Windows with utf-16 characters, so skip this test for other platforms
 	if runtime.GOOS != "windows" {

llama/Dockerfile

@@ -1,221 +0,0 @@
-# Note: once we have fully transitioned to the Go server, this will replace the old Dockerfile at the top of the tree
-ARG GOLANG_VERSION=1.22.5
-ARG CMAKE_VERSION=3.22.1
-ARG CUDA_VERSION_11=11.3.1
-ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
-ARG CUDA_VERSION_12=12.4.0
-ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
-ARG ROCM_VERSION=6.1.2
-
-### To create a local image for building linux binaries on mac or windows with efficient incremental builds
-#
-# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile.new --target unified-builder-amd64 .
-# docker run --platform linux/amd64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-amd64
-#
-### Then incremental builds will be much faster in this container
-#
-# make -C llama -j 10 && go build -trimpath -o dist/linux-amd64/ollama .
-#
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
-ARG CMAKE_VERSION
-ARG GOLANG_VERSION
-ARG CUDA_VERSION_11
-ARG CUDA_VERSION_12
-COPY ./scripts/rh_linux_deps.sh /
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:/usr/local/cuda/bin:$PATH
-ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
-ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
-    dnf clean all && \
-    dnf install -y \
-    zsh \
-    cuda-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
-    cuda-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
-# TODO intel oneapi goes here...
-ENV GOARCH amd64
-ENV CGO_ENABLED 1
-WORKDIR /go/src/github.com/ollama/ollama/
-ENTRYPOINT [ "zsh" ]
-
-### To create a local image for building linux binaries on mac or linux/arm64 with efficient incremental builds
-# Note: this does not contain jetson variants
-#
-# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile.new --target unified-builder-arm64 .
-# docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
-#
-FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
-ARG CMAKE_VERSION
-ARG GOLANG_VERSION
-ARG CUDA_VERSION_11
-ARG CUDA_VERSION_12
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
-    dnf config-manager --set-enabled appstream && \
-    dnf clean all && \
-    dnf install -y \
-    zsh \
-    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
-    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH:/usr/local/cuda/bin
-ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
-ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
-ENV GOARCH amd64
-ENV CGO_ENABLED 1
-WORKDIR /go/src/github.com/ollama/ollama/
-ENTRYPOINT [ "zsh" ]
-
-FROM --platform=linux/amd64 unified-builder-amd64 AS runners-amd64
-COPY . .
-ARG OLLAMA_SKIP_CUDA_GENERATE
-ARG OLLAMA_SKIP_CUDA_11_GENERATE
-ARG OLLAMA_SKIP_CUDA_12_GENERATE
-ARG OLLAMA_SKIP_ROCM_GENERATE
-ARG CUDA_V11_ARCHITECTURES
-ARG CUDA_V12_ARCHITECTURES
-ARG OLLAMA_FAST_BUILD
-RUN --mount=type=cache,target=/root/.ccache \
-    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
-        make -C llama -j $(expr $(nproc) / 2 ) ; \
-    else \
-        make -C llama -j 5 ; \
-    fi
-
-FROM --platform=linux/arm64 unified-builder-arm64 AS runners-arm64
-COPY . .
-ARG OLLAMA_SKIP_CUDA_GENERATE
-ARG OLLAMA_SKIP_CUDA_11_GENERATE
-ARG OLLAMA_SKIP_CUDA_12_GENERATE
-ARG CUDA_V11_ARCHITECTURES
-ARG CUDA_V12_ARCHITECTURES
-ARG OLLAMA_FAST_BUILD
-RUN --mount=type=cache,target=/root/.ccache \
-    make -C llama -j 8
-
-
-# Intermediate stages used for ./scripts/build_linux.sh
-FROM --platform=linux/amd64 centos:7 AS builder-amd64
-ARG CMAKE_VERSION
-ARG GOLANG_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-ENV CGO_ENABLED 1
-ENV GOARCH amd64
-WORKDIR /go/src/github.com/ollama/ollama
-
-FROM --platform=linux/amd64 builder-amd64 AS build-amd64
-COPY . .
-COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/build/ build/
-ARG GOFLAGS
-ARG CGO_CFLAGS
-ARG OLLAMA_SKIP_ROCM_GENERATE
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-amd64/bin/ollama .
-RUN cd dist/linux-$GOARCH && \
-    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
-RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
-    cd dist/linux-$GOARCH-rocm && \
-    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
-    fi
-
-FROM --platform=linux/arm64 rockylinux:8 AS builder-arm64
-ARG CMAKE_VERSION
-ARG GOLANG_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-ENV CGO_ENABLED 1
-ENV GOARCH arm64
-WORKDIR /go/src/github.com/ollama/ollama
-
-FROM --platform=linux/arm64 builder-arm64 AS build-arm64
-COPY . .
-COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/build/ build/
-ARG GOFLAGS
-ARG CGO_CFLAGS
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-arm64/bin/ollama .
-RUN cd dist/linux-$GOARCH && \
-    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
-
-FROM --platform=linux/amd64 scratch AS dist-amd64
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
-FROM --platform=linux/arm64 scratch AS dist-arm64
-COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
-FROM dist-$TARGETARCH AS dist
-
-
-# Optimized container images do not cary nested payloads
-FROM --platform=linux/amd64 builder-amd64 AS container-build-amd64
-WORKDIR /go/src/github.com/ollama/ollama
-COPY . .
-ARG GOFLAGS
-ARG CGO_CFLAGS
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-amd64/bin/ollama .
-
-FROM --platform=linux/arm64 builder-arm64 AS container-build-arm64
-WORKDIR /go/src/github.com/ollama/ollama
-COPY . .
-ARG GOFLAGS
-ARG CGO_CFLAGS
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-arm64/bin/ollama .
-
-# For amd64 container images, filter out cuda/rocm to minimize size
-FROM runners-amd64 AS runners-cuda-amd64
-RUN rm -rf \
-    ./dist/linux-amd64/lib/ollama/libggml_hipblas.so \
-    ./dist/linux-amd64/lib/ollama/runners/rocm*
-
-FROM runners-amd64 AS runners-rocm-amd64
-RUN rm -rf \
-    ./dist/linux-amd64/lib/ollama/libggml_cuda*.so \
-    ./dist/linux-amd64/lib/ollama/libcu*.so* \
-    ./dist/linux-amd64/lib/ollama/runners/cuda*
-
-FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
-RUN apt-get update && \
-    apt-get install -y ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
-COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-
-FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
-RUN apt-get update && \
-    apt-get install -y ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
-COPY --from=container-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
-COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
-
-# ROCm libraries larger so we keep it distinct from the CPU/CUDA image
-FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
-# Frontload the rocm libraries which are large, and rarely change to increase chance of a common layer
-# across releases
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
-RUN apt-get update && \
-    apt-get install -y ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
-COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-
-EXPOSE 11434
-ENV OLLAMA_HOST 0.0.0.0
-
-ENTRYPOINT ["/bin/ollama"]
-CMD ["serve"]
-
-FROM runtime-$TARGETARCH
-EXPOSE 11434
-ENV OLLAMA_HOST 0.0.0.0
-ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
-ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
-ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
-ENV NVIDIA_VISIBLE_DEVICES=all
-
-ENTRYPOINT ["/bin/ollama"]
-CMD ["serve"]

llama/Makefile

@@ -9,8 +9,7 @@ ifeq ($(OS),windows)
 	CUDA_BASE_DIR := $(dir $(shell cygpath -m -s "$(CUDA_PATH)\\.." 2>/dev/null))
 	CUDA_11:=$(shell ls -d $(CUDA_BASE_DIR)/v11.? 2>/dev/null)
 	CUDA_12:=$(shell ls -d $(CUDA_BASE_DIR)/v12.? 2>/dev/null)
-	HIP_PATH_83 := $(shell cygpath -m -s "$(subst \,/,$(HIP_PATH))" 2>/dev/null)
-	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH_83)/lib 2>/dev/null)
+	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH)/lib 2>/dev/null)
 else ifeq ($(OS),linux)
 	HIP_PATH?=/opt/rocm
 	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH)/lib 2>/dev/null)
@@ -41,8 +40,12 @@ runners: $(RUNNER_TARGETS)
 $(RUNNER_TARGETS):
 	$(MAKE) -f make/Makefile.$@
 
+help-sync apply-patches create-patches sync:
+	$(MAKE) -f make/Makefile.sync $@
+
 clean:
 	rm -rf $(BUILD_DIR) $(DIST_RUNNERS) $(PAYLOAD_RUNNERS)
+	go clean -cache
 
 clean-payload:
 	rm -rf $(addprefix $(RUNNERS_PAYLOAD_DIR)/, $(RUNNER_TARGETS) metal cpu cpu_avx cpu_avx2)

llama/README.md

@@ -91,10 +91,70 @@ go build -tags avx,rocm .
 make -j
 ```
 
-## Syncing with llama.cpp
+## Vendoring
 
-To update this package to the latest llama.cpp code, use the `sync.sh` script:
+Ollama currently vendors [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [ggml](https://github.com/ggerganov/ggml) through a vendoring model.  While we generally strive to contribute changes back upstream to avoid drift, we cary a small set of patches which are applied to the tracking commit.  A set of make targets are available to aid developers in updating to a newer tracking commit, or to work on changes.
+
+If you update the vendoring code, start by running the following command to establish the tracking llama.cpp repo in the `./vendor/` directory.
 
 ```
-./sync.sh ../../llama.cpp
+make apply-patches
 ```
+
+### Updating Base Commit
+
+**Pin to new base commit**
+
+To update to a newer base commit, select the upstream git tag or commit and update `llama/vendoring.env`
+
+#### Applying patches
+
+When updating to a newer base commit, the existing patches may not apply cleanly and require manual merge resolution.
+
+Start by applying the patches.  If any of the patches have conflicts, the `git am` will stop at the first failure.
+
+```
+make apply-patches
+```
+
+If you see an error message about a conflict, go into the `./vendor/` directory, and perform merge resolution using your preferred tool to the patch commit which failed.  Save the file(s) and continue the patch series with `git am --continue` .  If any additional patches fail, follow the same pattern until the full patch series is applied.  Once finished, run a final `create-patches` and `sync` target to ensure everything is updated.
+
+```
+make create-patches sync
+```
+
+Build and test Ollama, and make any necessary changes to the Go code based on the new base commit.  Submit your PR to the Ollama repo.
+
+### Generating Patches
+
+When working on new fixes or features that impact vendored code, use the following model.  First get a clean tracking repo with all current patches applied:
+
+```
+make apply-patches
+```
+
+Now edit the upstream native code in the `./vendor/` directory.  You do not need to commit every change in order to build, a dirty working tree in the tracking repo is OK while developing.  Simply save in your editor, and run the following to refresh the vendored code with your changes, build the backend(s) and build ollama:
+
+```
+make sync
+make -j 8
+go build .
+```
+
+> [!IMPORTANT]
+> Do **NOT** run `apply-patches` while you're iterating as that will reset the tracking repo.  It will detect a dirty tree and abort, but if your tree is clean and you accidentally ran this target, use `git reflog` to recover your commit(s).
+
+Iterate until you're ready to submit PRs.  Once your code is ready, commit a change in the `./vendor/` directory, then generate the patches for ollama with
+
+```
+make create-patches
+```
+
+> [!IMPORTANT]
+> Once you have completed this step, it is safe to run `apply-patches` since your change is preserved in the patches.
+
+In your `./vendor/` directory, create a branch, and cherry-pick the new commit to that branch, then submit a PR upstream to llama.cpp.
+
+Commit the changes in the ollama repo and submit a PR to Ollama, which will include the vendored code update with your change, along with the patches.
+
+After your PR upstream is merged, follow the **Updating Base Commit** instructions above, however first remove your patch before running `apply-patches` since the new base commit contains your change already.

llama/build-info.cpp

@@ -1,30 +1,4 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "";
+char const *LLAMA_COMMIT = "3f1ae2e32cde00c39b96be6d01c2997c29bae555";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/ggml-cuda.cu

@@ -2296,6 +2296,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_PAD:
             ggml_cuda_op_pad(ctx, dst);
             break;
+        case GGML_OP_UNPAD:
+            ggml_cuda_op_unpad(ctx, dst);
+            break;
         case GGML_OP_ARANGE:
             ggml_cuda_op_arange(ctx, dst);
             break;
@@ -3018,6 +3021,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_GROUP_NORM:
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:

llama/ggml-cuda/pad.cu

@@ -73,3 +73,49 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
 }
+
+static __global__ void unpad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
+    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
+    // blockIdx.y: idx of ne1
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+        dst[offset_dst] = x[offset_src];
+    }
+}
+
+static void unpad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02, const int ne03,
+    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2*ne3);
+    unpad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+}
+
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    unpad_f32_cuda(src0_d, dst_d,
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}

llama/ggml-cuda/pad.cuh

@@ -29,3 +29,4 @@
 #define CUDA_PAD_BLOCK_SIZE 256
 
 void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-metal.metal

@@ -2055,6 +2055,51 @@ kernel void kernel_pad_f32(
     }
 }
 
+kernel void kernel_unpad_f32(
+    device  const char * src0,
+    device        char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+
+    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
+        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+            if (i0 < ne00) {
+                dst_ptr[i0] = src0_ptr[i0];
+            }
+        }
+
+        return;
+    }
+}
+
 kernel void kernel_arange_f32(
     device        char * dst,
     constant   int64_t & ne0,

llama/ggml-metal_darwin_arm64.m

@@ -219,6 +219,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_IM2COL_F32,
     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
     GGML_METAL_KERNEL_TYPE_PAD_F32,
+    GGML_METAL_KERNEL_TYPE_UNPAD_F32,
     GGML_METAL_KERNEL_TYPE_ARANGE_F32,
     GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
@@ -715,6 +716,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                       pad_f32,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UNPAD_F32,                     unpad_f32,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,        timestep_embedding_f32,         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                    arange_f32,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,           argsort_f32_i32_asc,            true);
@@ -872,6 +874,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
             return false;
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
@@ -2681,6 +2684,36 @@ static void ggml_metal_encode_node(
 
                 const int nth = MIN(1024, ne0);
 
+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_UNPAD:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UNPAD_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+
+                const int nth = MIN(1024, ne0);
+
                 [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_ARANGE:

llama/ggml.c

@@ -3023,6 +3023,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "POOL_2D_BACK",
     "UPSCALE",
     "PAD",
+    "UNPAD",
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
@@ -3056,7 +3057,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "OPT_STEP_ADAMW",
 };
 
-static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80");
+static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -3117,6 +3118,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "pool_2d_back(x)",
     "upscale(x)",
     "pad(x)",
+    "unpad(x)",
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
@@ -3150,7 +3152,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "adamw(x)",
 };
 
-static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80");
+static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -6981,6 +6983,32 @@ struct ggml_tensor * ggml_pad(
     return result;
 }
 
+// ggml_unpad
+
+struct ggml_tensor * ggml_unpad(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    int p0, int p1, int p2, int p3) {
+    bool is_node = false;
+
+    if (a->grad) {
+        GGML_ABORT("fatal error"); // TODO: implement backward
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
+            a->ne[0] - p0,
+            a->ne[1] - p1,
+            a->ne[2] - p2,
+            a->ne[3] - p3);
+
+    result->op = GGML_OP_UNPAD;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_arange
 
 struct ggml_tensor * ggml_arange(
@@ -15338,6 +15366,58 @@ static void ggml_compute_forward_pad(
     }
 }
 
+static void ggml_compute_forward_unpad_f32(
+    const struct ggml_compute_params *params,
+    struct ggml_tensor *dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT( dst->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    float * dst_ptr = (float *) dst->data;
+
+    // TODO: optimize
+
+    for (int64_t i2 = 0; i2 < ne2; ++i2) {
+        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                for (int64_t i3 = 0; i3 < ne3; ++i3) {
+                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+
+                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        dst_ptr[dst_idx] = *src_ptr;
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_unpad(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_unpad_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
 
 // ggml_compute_forward_arange
 
@@ -17320,6 +17400,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_pad(params, tensor);
             } break;
+        case GGML_OP_UNPAD:
+            {
+                ggml_compute_forward_unpad(params, tensor);
+            } break;
         case GGML_OP_ARANGE:
             {
                 ggml_compute_forward_arange(params, tensor);
@@ -18395,6 +18479,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ABORT("fatal error"); // TODO: not implemented
             }
+        case GGML_OP_UNPAD:
+            {
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_ARANGE:
             {
                 GGML_ABORT("fatal error"); // TODO: not implemented
@@ -19191,6 +19279,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:

llama/ggml.h

@@ -532,6 +532,7 @@ extern "C" {
         GGML_OP_POOL_2D_BACK,
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
+        GGML_OP_UNPAD,
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
@@ -1790,6 +1791,15 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    // unpad each dimension: [x, ..., x, y, ..., y] -> [x, ..., x]
+    GGML_API struct ggml_tensor * ggml_unpad(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                  p0,
+            int                  p1,
+            int                  p2,
+            int                  p3);
+
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
     // return: [N, dim]

llama/llama.cpp

@@ -195,6 +195,7 @@ static std::string format(const char * fmt, ...) {
 
 enum llm_arch {
     LLM_ARCH_LLAMA,
+    LLM_ARCH_MLLAMA,
     LLM_ARCH_FALCON,
     LLM_ARCH_BAICHUAN,
     LLM_ARCH_GROK,
@@ -249,6 +250,7 @@ enum llm_arch {
 
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_LLAMA,           "llama"        },
+    { LLM_ARCH_MLLAMA,          "mllama"       },
     { LLM_ARCH_FALCON,          "falcon"       },
     { LLM_ARCH_GROK,            "grok"         },
     { LLM_ARCH_GPT2,            "gpt2"         },
@@ -356,6 +358,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
+    LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_FREQ_BASE,
@@ -465,6 +468,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"           },
     { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                    },
     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection.%d" },
+    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers"   },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
@@ -639,6 +643,14 @@ enum llm_tensor {
     LLM_TENSOR_CLS,
     LLM_TENSOR_CLS_OUT,
     LLM_TENSOR_BSKCN_TV,
+    LLM_TENSOR_CROSS_ATTN_K_NORM,
+    LLM_TENSOR_CROSS_ATTN_K_PROJ,
+    LLM_TENSOR_CROSS_ATTN_O_PROJ,
+    LLM_TENSOR_CROSS_ATTN_Q_NORM,
+    LLM_TENSOR_CROSS_ATTN_Q_PROJ,
+    LLM_TENSOR_CROSS_ATTN_V_PROJ,
+    LLM_TENSOR_CROSS_ATTN_ATTN_GATE,
+    LLM_TENSOR_CROSS_ATTN_MLP_GATE,
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@@ -668,6 +680,40 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_MLLAMA,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_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_CROSS_ATTN_K_NORM,    "blk.%d.cross_attn_k_norm" },
+            { LLM_TENSOR_CROSS_ATTN_K_PROJ,    "blk.%d.cross_attn_k_proj" },
+            { LLM_TENSOR_CROSS_ATTN_O_PROJ,    "blk.%d.cross_attn_o_proj" },
+            { LLM_TENSOR_CROSS_ATTN_Q_NORM,    "blk.%d.cross_attn_q_norm" },
+            { LLM_TENSOR_CROSS_ATTN_Q_PROJ,    "blk.%d.cross_attn_q_proj" },
+            { LLM_TENSOR_CROSS_ATTN_V_PROJ,    "blk.%d.cross_attn_v_proj" },
+            { LLM_TENSOR_CROSS_ATTN_ATTN_GATE, "blk.%d.cross_attn_attn_gate" },
+            { LLM_TENSOR_CROSS_ATTN_MLP_GATE,  "blk.%d.cross_attn_mlp_gate" },
+        },
+    },
     {
         LLM_ARCH_BAICHUAN,
         {
@@ -2416,6 +2462,7 @@ enum e_model {
     MODEL_40B,
     MODEL_65B,
     MODEL_70B,
+    MODEL_90B,
     MODEL_236B,
     MODEL_314B,
     MODEL_SMALL,
@@ -2460,6 +2507,7 @@ struct llama_hparams {
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
 
     std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> cross_attn_layers;
 
     uint32_t n_layer_dense_lead = 0;
     uint32_t n_lora_q = 0;
@@ -2528,10 +2576,11 @@ struct llama_hparams {
         if (this->n_expert      != other.n_expert)      return true;
         if (this->n_expert_used != other.n_expert_used) return true;
 
-        if (this->n_head_arr    != other.n_head_arr)    return true;
-        if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
-        if (this->n_ff_arr      != other.n_ff_arr)      return true;
-        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
+        if (this->n_head_arr        != other.n_head_arr)        return true;
+        if (this->n_head_kv_arr     != other.n_head_kv_arr)     return true;
+        if (this->n_ff_arr          != other.n_ff_arr)          return true;
+        if (this->n_bskcn_arr       != other.n_bskcn_arr)       return true;
+        if (this->cross_attn_layers != other.cross_attn_layers) return true;
 
         if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
@@ -2649,6 +2698,10 @@ struct llama_hparams {
 
         GGML_ABORT("fatal error");
     }
+
+    bool cross_attention_layer(uint32_t il) const {
+        return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
+    }
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
@@ -2832,6 +2885,16 @@ struct llama_layer {
     struct ggml_tensor * ffn_down_scale;
 
     struct ggml_tensor * bskcn_tv;
+
+    // cross attention
+    struct ggml_tensor * cross_attn_k_norm;
+    struct ggml_tensor * cross_attn_k_proj;
+    struct ggml_tensor * cross_attn_o_proj;
+    struct ggml_tensor * cross_attn_q_norm;
+    struct ggml_tensor * cross_attn_q_proj;
+    struct ggml_tensor * cross_attn_v_proj;
+    struct ggml_tensor * cross_attn_attn_gate;
+    struct ggml_tensor * cross_attn_mlp_gate;
 };
 
 // very similar to llama_batch,
@@ -3478,6 +3541,12 @@ struct llama_context {
     struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
     struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
     struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+
+    // TODO (jmorganca): this should most likely be passed in as part of a batch
+    // and not set on the context for all batches.
+    float * cross_attn_state = nullptr;
+    bool cross_attn_state_first_pass = true;
+    struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061]
 };
 
 struct llama_lora_weight {
@@ -3712,6 +3781,18 @@ static bool llama_kv_cache_init(
     cache.v_l.reserve(n_layer);
 
     for (int i = 0; i < (int) n_layer; i++) {
+        // for cross attention layers
+        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layer(i)) {
+            struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front();
+            ggml_tensor * k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
+            ggml_tensor * v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
+            ggml_format_name(k, "cache_k_l%d", i);
+            ggml_format_name(v, "cache_v_l%d", i);
+            cache.k_l.push_back(k);
+            cache.v_l.push_back(v);
+            continue;
+        }
+
         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
 
@@ -5486,12 +5567,14 @@ static void llm_load_hparams(
     }
 
     // zero-out the per-layer hparams
-    std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
-    std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
-    std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
+    std::fill(hparams.n_head_arr.begin(),             hparams.n_head_arr.end(),        0);
+    std::fill(hparams.n_head_kv_arr.begin(),          hparams.n_head_kv_arr.end(),     0);
+    std::fill(hparams.n_ff_arr.begin(),               hparams.n_ff_arr.end(),          0);
+    std::fill(hparams.cross_attn_layers.begin(),      hparams.cross_attn_layers.end(), -1);
 
-    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff_arr,   hparams.n_layer);
-    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,       hparams.n_ff_arr,          hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT,      hparams.n_head_arr,        hparams.n_layer);
+    ml.get_arr(LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, hparams.cross_attn_layers, false);
 
     // n_head_kv is optional, default to n_head
     hparams.n_head_kv_arr = hparams.n_head_arr;
@@ -5540,7 +5623,7 @@ static void llm_load_hparams(
 
         ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
 
-        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
+        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_MLLAMA || model.arch == LLM_ARCH_FALCON) {
             if (hparams.n_rot != hparams.n_embd_head_k) {
                 throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
             }
@@ -5580,6 +5663,16 @@ static void llm_load_hparams(
                     }
                 }
             } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 40: model.type = e_model::MODEL_11B; break;
+                    case 100: model.type = e_model::MODEL_90B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_MINICPM:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -7275,6 +7368,55 @@ static bool llm_load_tensors(
                         layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
                     }
                 } break;
+            case LLM_ARCH_MLLAMA:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        if (hparams.cross_attention_layer(i)) {
+                            layer.cross_attn_k_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_NORM,   "weight", i), {128});
+                            layer.cross_attn_k_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_PROJ,   "weight", i), {n_embd, 1024});
+                            layer.cross_attn_o_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_O_PROJ,   "weight", i), {n_embd, n_embd});
+                            layer.cross_attn_q_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128});
+                            layer.cross_attn_q_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd});
+                            layer.cross_attn_v_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024});
+                            layer.cross_attn_attn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1});
+                            layer.cross_attn_mlp_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1});
+                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        } else {
+                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
+                            layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                            layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                            layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                            layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        }
+                    }
+                } break;
             case LLM_ARCH_GROK:
                 {
                     if (n_expert == 0) {
@@ -9119,7 +9261,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
 
         if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
             model.hparams.n_vocab != model.vocab.id_to_token.size()) {
-            throw std::runtime_error("vocab size mismatch");
+            LLAMA_LOG_WARN("%s: vocab mismatch %u !- %zu ...\n", __func__, model.hparams.n_vocab, model.vocab.id_to_token.size());
         }
 
         if (params.vocab_only) {
@@ -9204,7 +9346,7 @@ static struct ggml_tensor * llm_build_inp_embd(
 
         inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
     } else {
-       lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
+        lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
         inpL = lctx.inp_embd;
         ggml_set_input(lctx.inp_embd);
     }
@@ -9219,6 +9361,22 @@ static struct ggml_tensor * llm_build_inp_embd(
     return inpL;
 }
 
+static struct ggml_tensor * llm_build_inp_cross_attn_state(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+        const llama_hparams & hparams,
+         const llm_build_cb & cb) {
+    const int64_t n_embd = hparams.n_embd;
+
+    struct ggml_tensor * inpCAS;
+    lctx.inp_cross_attn_state = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, 1601, 4);
+    cb(lctx.inp_cross_attn_state, "inp_cross_attn_state", -1);
+    ggml_set_input(lctx.inp_cross_attn_state);
+    inpCAS = lctx.inp_cross_attn_state;
+
+    return inpCAS;
+}
+
 static void llm_build_kv_store(
         struct ggml_context * ctx,
         const llama_hparams & hparams,
@@ -10193,6 +10351,7 @@ struct llm_build_context {
         lctx.inp_pos_bucket    = nullptr;
         lctx.inp_embd_enc      = nullptr;
         lctx.inp_KQ_mask_cross = nullptr;
+        lctx.inp_cross_attn_state = nullptr;
     }
 
     void free() {
@@ -10780,6 +10939,253 @@ struct llm_build_context {
                 LLM_NORM_RMS, cb, -1);
         cb(cur, "result_norm", -1);
 
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_mllama() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+        struct ggml_tensor * inpCAS;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpCAS = llm_build_inp_cross_attn_state(ctx0, lctx, hparams, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            if (hparams.cross_attention_layer(il)) {
+                if (!lctx.cross_attn_state) {
+                    continue;
+                }
+
+                // cross attention layer
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_q_proj, cur);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+                cb(Qcur, "Qcur", il);
+
+                // TODO: is this required?
+                Qcur = ggml_cont(ctx0, Qcur);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur;
+                if (lctx.cross_attn_state_first_pass) {
+                    Kcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_k_proj, inpCAS);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 6404);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = ggml_permute(ctx0, Kcur, 0, 2, 1, 3);
+                    cb(Kcur, "Kcur", il);
+
+                    // TODO: is this required?
+                    Kcur = ggml_cont(ctx0, Kcur);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, cb, il);
+                    cb(Kcur, "Kcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self.k_l[il]));
+                } else {
+                    Kcur = ggml_view_tensor(ctx0, kv_self.k_l[il]);
+                    cb(Kcur, "Kcur (view)", il);
+                }
+
+                struct ggml_tensor * Vcur;
+                if (lctx.cross_attn_state_first_pass) {
+                    Vcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_v_proj, inpCAS);
+                    cb(Vcur, "Vcur", il);
+
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, 6404);
+                    cb(Vcur, "Vcur", il);
+
+                    Vcur = ggml_permute(ctx0, Vcur, 0, 2, 1, 3);
+                    cb(Vcur, "Vcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, kv_self.v_l[il]));
+                } else {
+                    Vcur = ggml_view_tensor(ctx0, kv_self.v_l[il]);
+                    cb(Vcur, "Vcur (view)", il);
+                }
+
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, Kcur, Qcur);
+                cb(kq, "kq", il);
+
+                kq = ggml_scale_inplace(ctx0, kq, 1.0f/sqrtf(float(n_embd_head)));
+                cb(kq, "kq_scaled", il);
+
+                // TODO: apply causal masks
+                struct ggml_tensor * kq_soft_max = ggml_soft_max_inplace(ctx0, kq);
+                cb(kq_soft_max, "kq_soft_max", il);
+
+                Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, Vcur));
+                cb(Vcur, "Vcur", il);
+
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, Vcur, kq_soft_max);
+                cb(kqv, "kqv", il);
+
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
+
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_o_proj, cur);
+                cb(cur, "cur", il);
+
+                // TODO: do this in place once?
+                cur = ggml_mul(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_attn_gate));
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+
+                // TODO: do this inplace once?
+                cur = ggml_add_inplace(ctx0, ggml_mul_inplace(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_mlp_gate)), ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            } else {
+                // self attention layer
+
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+
+
+                if (il == n_layer - 1) {
+                    // skip computing output for unused tokens
+                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                    n_tokens = n_outputs;
+                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                }
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            }
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
         // lm_head
         cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
         cb(cur, "result_output", -1);
@@ -16527,6 +16933,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_llama();
             } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                result = llm.build_mllama();
+            } break;
         case LLM_ARCH_BAICHUAN:
             {
                 result = llm.build_baichuan();
@@ -16799,6 +17209,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
     }
 
+    // TODO (jmorganca): this might copy a lot of data on every request of a
+    // single generation even though it doesn't change, so we should
+    // find a way to not set this more than one time per image
+    if (lctx.inp_cross_attn_state &&
+        lctx.inp_cross_attn_state->buffer) {
+        ggml_backend_tensor_set(lctx.inp_cross_attn_state, lctx.cross_attn_state, 0, hparams.n_embd * 1601 * 4 * ggml_element_size(lctx.inp_cross_attn_state));
+    }
+
     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
         GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs");
         const int64_t n_tokens = batch.n_tokens;
@@ -17481,6 +17899,10 @@ static int llama_decode_internal(
 
         llama_set_inputs(lctx, ubatch);
 
+        // TODO: replace with something better to find out if its
+        // our first actual pass
+        lctx.cross_attn_state_first_pass = false;
+
         llama_graph_compute(lctx, gf, n_threads, threadpool);
 
         // update the kv ring buffer
@@ -18674,7 +19096,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         if (llama_model_has_encoder(&model)) {
             n_attn_layer *= 3;
         }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+        if (qs.n_attention_wv != n_attn_layer) {
+            LLAMA_LOG_WARN("%s: n_attention_wv is unexpected, expected: %d, found: %d\n", __func__, n_attn_layer, qs.n_attention_wv);
+        }
     }
 
     size_t total_size_org = 0;
@@ -19770,6 +20194,11 @@ struct llama_context * llama_new_context_with_model(
     return ctx;
 }
 
+void llama_set_cross_attn_state(struct llama_context * ctx, float * cross_attn_state) {
+    ctx->cross_attn_state_first_pass = true;
+    ctx->cross_attn_state = cross_attn_state;
+}
+
 void llama_free(struct llama_context * ctx) {
     delete ctx;
 }
@@ -19840,6 +20269,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_MLLAMA:
         case LLM_ARCH_BAICHUAN:
         case LLM_ARCH_STARCODER:
         case LLM_ARCH_PLAMO:

llama/llama.go

@@ -1,45 +1,70 @@
 package llama
 
+//go:generate make -j 8
+
 /*
 #cgo CFLAGS: -O2 -std=c11 -DGGML_BUILD=1 -DNDEBUG -DLOG_DISABLE_LOGS -DGGML_USE_LLAMAFILE
 #cgo CXXFLAGS: -O2 -std=c++11 -DGGML_BUILD=1 -DNDEBUG -DLOG_DISABLE_LOGS -DGGML_USE_LLAMAFILE
-#cgo darwin,arm64 CFLAGS: -DGGML_USE_METAL -DGGML_USE_ACCELERATE -DGGML_METAL_EMBED_LIBRARY -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_BLAS
-#cgo darwin,arm64 CXXFLAGS: -DGGML_USE_METAL -DGGML_USE_ACCELERATE -DGGML_METAL_EMBED_LIBRARY -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_BLAS
-#cgo darwin,arm64 LDFLAGS: -framework Foundation -framework Metal -framework MetalKit -framework Accelerate
+#cgo amd64,avx CFLAGS: -mavx
+#cgo amd64,avx CXXFLAGS: -mavx
+#cgo amd64,avx2 CFLAGS: -mavx2 -mfma
+#cgo amd64,avx2 CXXFLAGS: -mavx2 -mfma
+#cgo amd64,f16c CFLAGS: -mf16c
+#cgo amd64,f16c CXXFLAGS: -mf16c
+#cgo amd64,fma CFLAGS: -mfma
+#cgo amd64,fma CXXFLAGS: -mfma
+#cgo avx CFLAGS: -mavx
+#cgo avx CXXFLAGS: -mavx
+#cgo avx2 CFLAGS: -mavx2 -mfma -mf16c
+#cgo avx2 CXXFLAGS: -mavx2 -mfma -mf16c
+#cgo cuda CFLAGS: -fPIE -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+#cgo cuda CFLAGS: -fPIE -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+#cgo cuda CXXFLAGS: -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+#cgo cuda CXXFLAGS: -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+#cgo cuda_v11 LDFLAGS: -lggml_cuda_v11 -L/usr/local/cuda-11/lib64
+#cgo cuda_v12 LDFLAGS: -lggml_cuda_v12 -L/usr/local/cuda-12/lib64
 #cgo darwin,amd64 CFLAGS: -Wno-incompatible-pointer-types-discards-qualifiers
 #cgo darwin,amd64 CXXFLAGS: -Wno-incompatible-pointer-types-discards-qualifiers
 #cgo darwin,amd64 LDFLAGS: -framework Foundation
 #cgo darwin,amd64,avx2 CFLAGS: -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
 #cgo darwin,amd64,avx2 CXXFLAGS: -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
 #cgo darwin,amd64,avx2 LDFLAGS: -framework Accelerate
+#cgo darwin,arm64 CFLAGS: -DGGML_USE_METAL -DGGML_USE_ACCELERATE -DGGML_METAL_EMBED_LIBRARY -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_BLAS
+#cgo darwin,arm64 CXXFLAGS: -DGGML_USE_METAL -DGGML_USE_ACCELERATE -DGGML_METAL_EMBED_LIBRARY -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_BLAS
+#cgo darwin,arm64 LDFLAGS: -framework Foundation -framework Metal -framework MetalKit -framework Accelerate
 #cgo linux CFLAGS: -D_GNU_SOURCE
 #cgo linux CXXFLAGS: -D_GNU_SOURCE
-#cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/Linux/arm64
 #cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/Linux/amd64
-#cgo windows CFLAGS: -Wno-discarded-qualifiers
-#cgo windows LDFLAGS: -lmsvcrt -static-libstdc++ -static-libgcc -static
-#cgo windows,arm64 LDFLAGS: -L${SRCDIR}/build/Windows/arm64
-#cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/Windows/amd64
-#cgo avx CFLAGS: -mavx
-#cgo avx CXXFLAGS: -mavx
-#cgo avx2 CFLAGS: -mavx2 -mfma -mf16c
-#cgo avx2 CXXFLAGS: -mavx2 -mfma -mf16c
-#cgo cuda CFLAGS: -fPIE -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1
-#cgo cuda CXXFLAGS: -fPIE -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1
-#cgo rocm CFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -D__HIP_PLATFORM_AMD__=1 -D__HIP_ROCclr__=1
-#cgo rocm CXXFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -D__HIP_PLATFORM_AMD__=1 -D__HIP_ROCclr__=1
-#cgo rocm LDFLAGS: -L${SRCDIR} -lggml_rocm -lhipblas -lamdhip64 -lrocblas
-#cgo cuda_v11 LDFLAGS: -lggml_cuda_v11 -L/usr/local/cuda-11/lib64
-#cgo cuda_v12 LDFLAGS: -lggml_cuda_v12 -L/usr/local/cuda-12/lib64
-#cgo windows,cuda LDFLAGS: -lcuda -lcudart -lcublas -lcublasLt
-#cgo windows,rocm LDFLAGS: -lggml_rocm -lhipblas -lamdhip64 -lrocblas
+#cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/Linux/amd64
+#cgo linux,arm64 CFLAGS: -D__aarch64__ -D__ARM_NEON -D__ARM_FEATURE_FMA -D__ARM_FEATURE_MATMUL_INT8
+#cgo linux,arm64 CXXFLAGS: -D__aarch64__ -D__ARM_NEON -D__ARM_FEATURE_FMA -D__ARM_FEATURE_MATMUL_INT8
+#cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/Linux/arm64
+#cgo linux,arm64,sve CFLAGS: -march=armv8.6-a+sve
+#cgo linux,arm64,sve CXXFLAGS: -march=armv8.6-a+sve
 #cgo linux,cuda LDFLAGS: -lcuda -lcudart -lcublas -lcublasLt -lpthread -ldl -lrt -lresolv
 #cgo linux,rocm LDFLAGS: -L/opt/rocm/lib -lpthread -ldl -lrt -lresolv
+#cgo rocm CFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+#cgo rocm CXXFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+#cgo rocm LDFLAGS: -L${SRCDIR} -lggml_rocm -lhipblas -lamdhip64 -lrocblas
+#cgo windows CFLAGS: -Wno-discarded-qualifiers -D_WIN32_WINNT=0x602
+#cgo windows CXXFLAGS: -D_WIN32_WINNT=0x602
+#cgo windows LDFLAGS: -lmsvcrt
+#cgo windows LDFLAGS: -lmsvcrt -static-libstdc++ -static-libgcc -static
+#cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/Windows/amd64
+#cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/Windows/amd64
+#cgo windows,arm64 CFLAGS: -D__aarch64__ -D__ARM_NEON -D__ARM_FEATURE_FMA
+#cgo windows,arm64 CXXFLAGS: -D__aarch64__ -D__ARM_NEON -D__ARM_FEATURE_FMA
+#cgo windows,arm64 LDFLAGS: -L${SRCDIR}/build/Windows/arm64
+#cgo windows,arm64 LDFLAGS: -L${SRCDIR}/build/Windows/arm64
+#cgo windows,cuda LDFLAGS: -lcuda -lcudart -lcublas -lcublasLt
+#cgo windows,rocm LDFLAGS: -lggml_rocm -lhipblas -lamdhip64 -lrocblas
 
 #include <stdlib.h>
 #include "llama.h"
 #include "clip.h"
+#include "ggml.h"
 #include "llava.h"
+#include "mllama.h"
 #include "sampling_ext.h"
 
 bool llamaProgressCallback(float progress, void *user_data);
@@ -389,18 +414,60 @@ func Quantize(infile, outfile string, ftype uint32) error {
 
 // llava
 type ClipContext struct {
-	c *C.struct_clip_ctx
+	c        *C.struct_clip_ctx
+	m        *C.struct_mllama_ctx
+	IsMllama bool
+	embedPin runtime.Pinner
+	pinned   bool
 }
 
-func NewClipContext(modelPath string) *ClipContext {
+func getVisionArch(mp *C.char) (string, error) {
+	gguf_ctx := C.gguf_init_from_file(mp, C.struct_gguf_init_params{no_alloc: true, ctx: (**C.struct_ggml_context)(C.NULL)})
+	if gguf_ctx == nil {
+		return "", errors.New("unable to load vision projector")
+	}
+	defer C.gguf_free(gguf_ctx)
+
+	arch_index := C.gguf_find_key(gguf_ctx, C.CString("general.architecture"))
+	if int(arch_index) < 0 {
+		return "", errors.New("unknown vision model architecture")
+	}
+
+	arch := C.gguf_get_val_str(gguf_ctx, arch_index)
+
+	return C.GoString(arch), nil
+}
+
+func NewClipContext(modelPath string) (*ClipContext, error) {
 	mp := C.CString(modelPath)
 	defer C.free(unsafe.Pointer(mp))
-	cc := C.clip_model_load(mp, 1)
-	return &ClipContext{c: cc}
+
+	arch, err := getVisionArch(mp)
+	if err != nil {
+		return nil, err
+	}
+
+	var cc ClipContext
+	if arch == "clip" {
+		cc.c = C.clip_model_load(mp, 1)
+	} else if arch == "mllama" {
+		cc.m = C.mllama_model_load(mp, 1)
+		cc.IsMllama = true
+	} else {
+		return nil, fmt.Errorf("unknown vision model architecture: %s", arch)
+	}
+
+	// XXX: check embedding size?
+	return &cc, nil
 }
 
 func (c *ClipContext) Free() {
-	C.clip_free(c.c)
+	if c.c != nil {
+		C.clip_free(c.c)
+	}
+	if c.m != nil {
+		C.mllama_free(c.m)
+	}
 }
 
 func NewLlavaImageEmbed(llamaContext *Context, clipContext *ClipContext, data []byte) [][]float32 {
@@ -424,6 +491,48 @@ func NewLlavaImageEmbed(llamaContext *Context, clipContext *ClipContext, data []
 	return embed
 }
 
+func NewMllamaImageEmbed(llamaContext *Context, clipContext *ClipContext, data []byte, aspectRatioId int) [][]float32 {
+	img := C.mllama_image_init()
+	defer C.mllama_image_free(img)
+
+	C.mllama_image_load_from_data(unsafe.Pointer(&data[0]), C.int(len(data)), 560, 560, 3, 4, C.int(aspectRatioId), img)
+
+	numTokens := int(C.mllama_n_positions(clipContext.m) * C.mllama_n_tiles(clipContext.m))
+	numEmbed := llamaContext.Model().NEmbd()
+
+	rows := make([]float32, numEmbed*numTokens)
+	C.mllama_image_encode(clipContext.m, C.int(llamaContext.numThreads), img, (*C.float)(unsafe.Pointer(&rows[0])))
+
+	embed := make([][]float32, numTokens)
+	for i := range embed {
+		embed[i] = rows[i*numEmbed : (i+1)*numEmbed]
+	}
+
+	return embed
+}
+
+// This really needs to be set on a batch instead
+func MllamaSetCrossAttn(llamaContext *Context, clipContext *ClipContext, embed [][]float32) {
+	if embed != nil {
+		if clipContext.pinned {
+			panic("Cross attention state already pinned")
+		}
+
+		embedData := &embed[0][0]
+		clipContext.embedPin.Pin(embedData)
+		clipContext.pinned = true
+
+		C.llama_set_cross_attn_state(llamaContext.c, (*C.float)(unsafe.Pointer(embedData)))
+	} else {
+		C.llama_set_cross_attn_state(llamaContext.c, (*C.float)(C.NULL))
+
+		if clipContext.pinned {
+			clipContext.embedPin.Unpin()
+			clipContext.pinned = false
+		}
+	}
+}
+
 // sampling
 // TODO: this is a temporary wrapper to allow calling C++ code from CGo
 type SamplingContext struct {

llama/llama.h

@@ -449,6 +449,10 @@ extern "C" {
                      struct llama_model * model,
             struct llama_context_params   params);
 
+    // TODO (jmorganca): this should most likely be passed in as part of a batch
+    // and not set on the context for all batches.
+    LLAMA_API void llama_set_cross_attn_state(struct llama_context * ctx, float * cross_attn_state);
+
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 

llama/make/Makefile.default

@@ -8,8 +8,10 @@ CPU_GOFLAGS="-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$(VERS
 DEFAULT_RUNNER := $(if $(and $(filter darwin,$(OS)),$(filter arm64,$(ARCH))),metal,cpu)
 RUNNERS := $(DEFAULT_RUNNER)
 ifeq ($(ARCH),amd64)
+ifeq ($(CUSTOM_CPU_FLAGS),)
 	RUNNERS += cpu_avx cpu_avx2
 endif
+endif
 
 DIST_RUNNERS = $(addprefix $(RUNNERS_DIST_DIR)/,$(addsuffix /ollama_llama_server$(EXE_EXT),$(RUNNERS)))
 ifneq ($(OS),windows)
@@ -19,20 +21,20 @@ BUILD_RUNNERS = $(addprefix $(RUNNERS_BUILD_DIR)/,$(addsuffix /ollama_llama_serv
 
 all: $(BUILD_RUNNERS) $(DIST_RUNNERS) $(PAYLOAD_RUNNERS)
 
-$(RUNNERS_BUILD_DIR)/$(DEFAULT_RUNNER)/ollama_llama_server$(EXE_EXT): TARGET_CPU_FLAGS=
+$(RUNNERS_BUILD_DIR)/$(DEFAULT_RUNNER)/ollama_llama_server$(EXE_EXT): TARGET_CPU_FLAGS=$(CUSTOM_CPU_FLAGS)
 $(RUNNERS_BUILD_DIR)/$(DEFAULT_RUNNER)/ollama_llama_server$(EXE_EXT): *.go ./runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
 	@-mkdir -p $(dir $@)
-	GOARCH=$(ARCH) go build $(CPU_GOFLAGS) -o $@ ./runner
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath $(if $(CUSTOM_CPU_FLAGS),-tags $(subst $(space),$(comma),$(CUSTOM_CPU_FLAGS)))  -o $@ ./runner
 
 $(RUNNERS_BUILD_DIR)/cpu_avx/ollama_llama_server$(EXE_EXT): TARGET_CPU_FLAGS="avx"
 $(RUNNERS_BUILD_DIR)/cpu_avx/ollama_llama_server$(EXE_EXT): *.go ./runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
 	@-mkdir -p $(dir $@)
-	GOARCH=$(ARCH) go build $(CPU_GOFLAGS) -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./runner
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./runner
 
 $(RUNNERS_BUILD_DIR)/cpu_avx2/ollama_llama_server$(EXE_EXT): TARGET_CPU_FLAGS="avx avx2"
 $(RUNNERS_BUILD_DIR)/cpu_avx2/ollama_llama_server$(EXE_EXT): *.go ./runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
 	@-mkdir -p $(dir $@)
-	GOARCH=$(ARCH) go build $(CPU_GOFLAGS) -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./runner
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./runner
 
 $(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%
 	@-mkdir -p $(dir $@)
@@ -47,3 +49,6 @@ clean:
 
 .PHONY: clean all
 
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'

llama/make/Makefile.rocm

@@ -9,19 +9,16 @@ HIP_ARCHS_COMMON := gfx900 gfx940 gfx941 gfx942 gfx1010 gfx1012 gfx1030 gfx1100
 HIP_ARCHS_LINUX := gfx906:xnack- gfx908:xnack- gfx90a:xnack+ gfx90a:xnack-
 
 ifeq ($(OS),windows)
-	GPU_LIB_DIR_WIN := $(shell cygpath -m -s "$(HIP_PATH)\bin")
-	# If HIP_PATH has spaces, hipcc trips over them when subprocessing
-	HIP_PATH := $(shell cygpath -m -s "$(HIP_PATH)\")
-	CGO_EXTRA_LDFLAGS_WIN := -L$(shell cygpath -m -s "$(HIP_PATH)\lib")
-	export HIP_PATH
-	GPU_COMPILER_WIN := $(HIP_PATH)bin/hipcc.bin.exe
+	GPU_LIB_DIR_WIN := $(shell cygpath -m -s "$(HIP_PATH)/bin")
+	CGO_EXTRA_LDFLAGS_WIN := -L$(shell cygpath -m -s "$(HIP_PATH)/lib")
+	GPU_COMPILER_WIN := $(HIP_PATH)/bin/hipcc.bin.exe
 	GPU_COMPILER:=$(GPU_COMPILER_WIN)
 else ifeq ($(OS),linux)
-	HIP_PATH?=/opt/rocm
 	GPU_LIB_DIR_LINUX := $(HIP_PATH)/lib
 	GPU_COMPILER_LINUX := $(shell X=$$(which hipcc 2>/dev/null) && echo $$X)
 	GPU_COMPILER:=$(GPU_COMPILER_LINUX)
-	ROCM_TRANSITIVE_LIBS = $(shell ldd $(ROCM_LIBS) | grep "=>" | cut -f2 -d= | cut -f2 -d' '  | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf | sort -u )
+	ROCM_TRANSITIVE_LIBS_INITIAL = $(sort $(shell ldd $(GPU_LIBS) | grep "=>" | cut -f2 -d= | cut -f2 -d' '  | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf))
+	GPU_TRANSITIVE_LIBS = $(sort $(shell readlink -f $(ROCM_TRANSITIVE_LIBS_INITIAL)) $(ROCM_TRANSITIVE_LIBS_INITIAL))
 endif
 
 # TODO future multi-variant support for ROCm
@@ -36,14 +33,18 @@ GPU_RUNNER_DRIVER_LIB_LINK := -lamdhip64
 GPU_RUNNER_LIBS_SHORT := hipblas rocblas
 GPU_PATH_ROOT_WIN=$(dir $(GPU_LIB_DIR_WIN))
 GPU_PATH_ROOT_LINUX=$(dir $(GPU_LIB_DIR_LINUX))
-GPU_COMPILER_CFLAGS_WIN = $(CFLAGS)
+GPU_COMPILER_CFLAGS_WIN = $(CFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CFLAGS_LINUX = $(CFLAGS) -fPIC -D_GNU_SOURCE
-GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS)
+GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CXXFLAGS_LINUX = $(CXXFLAGS) -fPIC -D_GNU_SOURCE
 
-ROCM_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
-ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH)-rocm)/lib/ollama
-ROCM_DIST_DEPS_LIBS = $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(ROCM_LIBS)) $(notdir $(ROCM_TRANSITIVE_LIBS)))
+GPU_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
+ifeq ($(OS),windows)
+	ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH))/lib/ollama
+else ifeq ($(OS),linux)
+	ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH)-rocm)/lib/ollama
+endif
+GPU_DIST_DEPS_LIBS= $(sort $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(GPU_LIBS)) $(notdir $(GPU_TRANSITIVE_LIBS))))
 ROCBLAS_DIST_DEP_MANIFEST = $(ROCM_DIST_DEPS_DIR)/rocblas/library/TensileManifest.txt
 
 ifeq ($(OS),linux)
@@ -84,18 +85,14 @@ GPU_COMPILER_CUFLAGS = \
 	-Wno-pass-failed \
 	-Wno-deprecated-declarations \
 	-Wno-unused-result \
-	-I. \
-	$(foreach arch, $(HIP_ARCHS_COMMON), --offload-arch=$(arch))
+	-I.
 
 include make/gpu.make
 
 # Adjust the rules from gpu.make to handle the ROCm dependencies properly
-$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST) $(ROCM_DIST_DEPS_LIBS)
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST)
 $(ROCBLAS_DIST_DEP_MANIFEST):
 	@-mkdir -p $(dir $@)
 	@echo "Copying rocblas library..."
 	cd $(GPU_LIB_DIR)/rocblas/library/ && tar cf - . | (cd $(dir $@) && tar xf - )
 	@echo "rocblas library copy complete"
-$(ROCM_DIST_DEPS_LIBS): 
-	@-mkdir -p $(dir $@)
-	$(CP) $(dir $(filter %$(notdir $@),$(ROCM_LIBS) $(ROCM_TRANSITIVE_LIBS)))/$(notdir $@)* $(dir $@)

llama/make/Makefile.sync

@@ -0,0 +1,191 @@
+# Helpers for managing our vendored llama.cpp repo and patch set
+
+REPO_ROOT:=$(dir $(patsubst %/,%,$(dir $(patsubst %/,%,$(dir $(abspath $(lastword $(MAKEFILE_LIST))))))))
+DST_DIR:=$(dir $(patsubst %/,%,$(dir $(abspath $(lastword $(MAKEFILE_LIST))))))
+
+include $(REPO_ROOT)llama/vendoring.env
+
+LLAMACPP_REPO := $(REPO_ROOT)llama/vendor/
+
+LLAMACPP_PATCH_DIR := $(DST_DIR)patches/
+
+
+help-sync:
+	@echo "The following make targets will help you update llama.cpp to a new base commit, or work on new features/fixes"
+	@echo ""
+	@echo "\tmake apply-patches   # Establish the tracking repo if not already present, reset to the base commit, and apply our patch set"
+	@echo "\tmake sync            # Vendor llama.cpp and ggml from the tracking repo working tree"
+	@echo "\tmake create-patches  # Generate the patch set based on the current commits in the tracking repo since the base commit"
+	@echo ""
+	@echo "For more details on the workflow, see the Vendoring section in ../docs/development.md"
+
+apply-patches: $(LLAMACPP_REPO)
+	@if ! git -C $(LLAMACPP_REPO) --no-pager diff --exit-code ; then \
+  		echo "ERROR: Your llama.cpp repo is dirty.  The apply-patches target requires a clean working tree"; \
+		echo "To clobber: git -C $(LLAMACPP_REPO) reset --hard HEAD" ; \
+  		exit 1; \
+	fi
+	@echo "Checking out $(LLAMACPP_BASE_COMMIT)"
+	@git -C $(LLAMACPP_REPO) checkout -q $(LLAMACPP_BASE_COMMIT) || \
+		git -C $(LLAMACPP_REPO) fetch --all && git -C $(LLAMACPP_REPO) checkout -q $(LLAMACPP_BASE_COMMIT)
+	@echo "Applying ollama patches..."
+	@git -c 'user.name=nobody' -c 'user.email=<>' -C $(LLAMACPP_REPO) am -3 $(LLAMACPP_PATCH_DIR)/*.patch || \
+		echo "Please resolve the conflicts in $(LLAMACPP_REPO), and run 'git am --continue' to continue applying subsequent patches"
+	@echo ""
+	@echo "The tracking repo $(LLAMACPP_REPO) is now in a detached state with all patches applied."
+	@echo "Don't forget to commit any changes you make and run 'make create-patches' "
+	
+$(LLAMACPP_REPO):
+	@echo "Cloning llama.cpp to $(LLAMACPP_REPO)"
+	git clone https://github.com/ggerganov/llama.cpp.git $@
+
+create-patches: $(LLAMACPP_REPO)
+	@if ! git -C $(LLAMACPP_REPO) --no-pager diff --exit-code ; then \
+  		echo "ERROR: Your llama.cpp repo is dirty.  You must commit any pending changes for format-patch to generate patches"; \
+  		exit 1; \
+	fi
+	git -C $(LLAMACPP_REPO) format-patch --no-signature --no-numbered --zero-commit -o $(LLAMACPP_PATCH_DIR) $(LLAMACPP_BASE_COMMIT)
+
+# Vendoring template logic
+EXCLUDED_FILES=sgemm.cpp sgemm.h sampling_ext.cpp sampling_ext.h stb_image.h json.hpp llama_darwin.c base64.hpp
+OLLAMA_NATIVE_FILES=mllama.cpp mllama.h llama_darwin.c sampling_ext.cpp sampling_ext.h
+define vendor_file
+$(strip $(addprefix $(2),$(notdir $1))) : $(addprefix $(LLAMACPP_REPO),$(1))
+ifneq ($$(filter-out $(EXCLUDED_FILES),$(notdir $1)),)
+	@echo "vendoring $1"; \
+		mkdir -p $$(dir $$@) && \
+		echo "/**" > $$@ && \
+		echo " * llama.cpp - commit $$(LLAMACPP_BASE_COMMIT) - do not edit this file" >> $$@ && \
+		echo " *" >> $$@ && \
+		sed 's/^/ * /' <$(LLAMACPP_REPO)/LICENSE | sed 's/ *$$$$//' >> $$@ && \
+		echo " */" >> $$@ && \
+		echo "" >> $$@ && \
+		cat $$< >> $$@
+else
+	@echo "vendoring $1"; \
+		mkdir -p $$(dir $$@) && \
+		cat $$< > $$@
+endif
+VENDORED_FILES += $(strip $(addprefix $(2),$(notdir $1)))
+endef
+
+# llama.cpp files -> llama/
+LLAMACPP_FILES=\
+	src/unicode.cpp \
+	src/unicode.h \
+	src/unicode-data.cpp \
+	src/unicode-data.h \
+	src/llama.cpp \
+	src/llama-impl.h \
+	src/llama-vocab.cpp \
+	src/llama-vocab.h \
+	src/llama-grammar.cpp \
+	src/llama-grammar.h \
+	src/llama-sampling.cpp \
+	src/llama-sampling.h \
+	include/llama.h \
+	ggml/src/llamafile/sgemm.cpp \
+	ggml/src/llamafile/sgemm.h
+$(foreach name,$(LLAMACPP_FILES),$(eval $(call vendor_file,$(name),$(DST_DIR))))
+
+# llama.cpp files -> llama/llamafile
+LLAMAFILE_FILES= \
+	ggml/src/llamafile/sgemm.h
+$(foreach name,$(LLAMAFILE_FILES),$(eval $(call vendor_file,$(name),$(DST_DIR)llamafile/)))
+
+# ggml files -> llama/
+GGML_FILES= \
+	ggml/src/ggml.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-quants.c \
+	ggml/src/ggml-quants.h \
+	ggml/src/ggml-metal.metal \
+	ggml/include/ggml-metal.h \
+	ggml/src/ggml-impl.h \
+	ggml/include/ggml-cuda.h \
+	ggml/src/ggml-cuda.cu \
+	ggml/src/ggml-common.h \
+	ggml/include/ggml-backend.h \
+	ggml/src/ggml-backend.c \
+	ggml/src/ggml-backend-impl.h \
+	ggml/include/ggml-alloc.h \
+	ggml/src/ggml-alloc.c \
+	ggml/src/ggml-aarch64.h \
+	ggml/src/ggml-aarch64.c \
+	ggml/src/ggml-cpu-impl.h \
+	ggml/include/ggml-blas.h \
+	ggml/src/ggml-blas.cpp
+$(foreach name,$(GGML_FILES),$(eval $(call vendor_file,$(name),$(DST_DIR))))
+
+# TODO generalize renaming pattern if we have more of these
+$(DST_DIR)ggml-metal_darwin_arm64.m : $(LLAMACPP_REPO)ggml/src/ggml-metal.m
+	@echo "vendoring $(subst $(LLAMACPP_REPO),,$<)"; \
+		mkdir -p $(dir $@) && \
+		echo "/**" > $@ && \
+		echo " * llama.cpp - commit $(LLAMACPP_BASE_COMMIT) - do not edit this file" >> $@ && \
+		echo " *" >> $@ && \
+		sed 's/^/ * /' <$(LLAMACPP_REPO)/LICENSE | sed 's/ *$$//' >> $@ && \
+		echo " */" >> $@ && \
+		echo "" >> $@ && \
+		cat $< >> $@
+VENDORED_FILES += $(DST_DIR)ggml-metal_darwin_arm64.m
+
+# ggml-cuda -> llama/ggml-cuda/
+GGML_CUDA_FILES= ggml/src/ggml-cuda/*.cu ggml/src/ggml-cuda/*.cuh
+GGML_CUDA_FILES_EXPANDED = $(addprefix ggml/src/ggml-cuda/,$(notdir $(wildcard $(addprefix $(LLAMACPP_REPO),$(GGML_CUDA_FILES)))))
+$(foreach name,$(GGML_CUDA_FILES_EXPANDED),$(eval $(call vendor_file,$(name),$(DST_DIR)ggml-cuda/)))
+
+GGML_TEMPLATE_FILES= ggml/src/ggml-cuda/template-instances/*.cu
+GGML_TEMPLATE_FILES_EXPANDED = 	$(addprefix ggml/src/ggml-cuda/template-instances/,$(notdir $(wildcard $(addprefix $(LLAMACPP_REPO),$(GGML_TEMPLATE_FILES)))))
+$(foreach name,$(GGML_TEMPLATE_FILES_EXPANDED),$(eval $(call vendor_file,$(name),$(DST_DIR)ggml-cuda/template-instances/)))
+
+GGML_VENDOR_FILES= ggml/src/ggml-cuda/vendors/*.h
+GGML_VENDOR_FILES_EXPANDED=$(addprefix ggml/src/ggml-cuda/vendors/,$(notdir $(wildcard $(addprefix $(LLAMACPP_REPO),$(GGML_VENDOR_FILES)))))
+$(foreach name,$(GGML_VENDOR_FILES_EXPANDED),$(eval $(call vendor_file,$(name),$(DST_DIR)ggml-cuda/vendors/)))
+
+# llava -> llama/
+LAVA_FILES= \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	common/log.h \
+	common/log.cpp \
+	common/stb_image.h
+# These files are mostly used by the llava code
+# and shouldn't be necessary once we use clip.cpp directly
+LAVA_FILES+= \
+	common/common.cpp \
+	common/common.h \
+	common/sampling.cpp \
+	common/sampling.h \
+	common/json.hpp \
+	common/json-schema-to-grammar.cpp \
+	common/json-schema-to-grammar.h \
+	common/base64.hpp
+$(foreach name,$(LAVA_FILES),$(eval $(call vendor_file,$(name),$(DST_DIR))))
+
+$(DST_DIR)build-info.cpp:
+	@echo "Generating $@"
+	@echo "int LLAMA_BUILD_NUMBER = 0;" > $@
+	@echo "char const *LLAMA_COMMIT = \"$(LLAMACPP_BASE_COMMIT)\";" >> $@
+	@echo "char const *LLAMA_COMPILER = \"\";" >> $@
+	@echo "char const *LLAMA_BUILD_TARGET = \"\";" >> $@
+VENDORED_FILES += $(DST_DIR)build-info.cpp
+
+
+sync: $(LLAMACPP_REPO) .WAIT $(VENDORED_FILES) .WAIT remove-stale-files
+
+PATS=*.c *.h *.cpp *.m *.metal *.cu *.cuh
+NATIVE_DIRS=$(DST_DIR) $(DST_DIR)llamafile/ $(DST_DIR)ggml-cuda/ $(DST_DIR)ggml-cuda/template-instances/ $(DST_DIR)ggml-cuda/vendors/
+ALL_NATIVE_FILES=$(foreach dir,$(NATIVE_DIRS),$(wildcard $(addprefix $(dir),$(PATS))))
+EXTRA_NATIVE_FILES=$(filter-out $(VENDORED_FILES) $(addprefix $(DST_DIR),$(OLLAMA_NATIVE_FILES)), $(ALL_NATIVE_FILES))
+remove-stale-files:
+	@rm -f $(EXTRA_NATIVE_FILES)
+
+.PHONY: help-sync apply-patches sync create-patches remove-stale-fails .WAIT 
+
+
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'

llama/make/common-defs.make

@@ -46,8 +46,6 @@ endif
 
 # Override in environment space separated to tune GPU runner CPU vector flags
 ifeq ($(ARCH),amd64)
-# TODO may need a bit more work - setting 'GPU_RUNNER_CPU_FLAGS="avx avx2 avx512f avx512bw"' doesn't yield
-# a system_info showing 'AVX512 = 1' so there may be additional macros that are needed in GGML
 	GPU_RUNNER_CPU_FLAGS ?= avx
 endif
 
@@ -59,12 +57,18 @@ ifeq ($(OS),windows)
 	EXE_EXT := .exe
 	SHARED_PREFIX := 
 	CPU_FLAG_PREFIX := /arch:
+ifneq ($(HIP_PATH),)
+	# If HIP_PATH has spaces, hipcc trips over them when subprocessing
+	HIP_PATH := $(shell cygpath -m -s "$(patsubst %\,%,$(HIP_PATH))")
+	export HIP_PATH
+endif
 else ifeq ($(OS),linux)
 	CP := cp -af
 	OBJ_EXT := o
 	SHARED_EXT := so
 	SHARED_PREFIX := lib
 	CPU_FLAG_PREFIX := -m
+	HIP_PATH?=/opt/rocm
 else
 	OBJ_EXT := o
 	SHARED_EXT := so

llama/make/cuda.make

@@ -19,6 +19,9 @@ GPU_COMPILER_CFLAGS_WIN = $(CFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CFLAGS_LINUX = $(CFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
 GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CXXFLAGS_LINUX = $(CXXFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
+GPU_LIBS = $(sort $(wildcard $(addsuffix *.$(SHARED_EXT)*,$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT)))))
+GPU_DIST_DEPS_LIBS= $(sort $(addprefix $(DIST_LIB_DIR)/,$(notdir $(GPU_LIBS))))
+
 ifeq ($(OS),linux)
 	CUDA_PATH?=/usr/local/cuda
 	GPU_COMPILER_FPIC = -fPIC -Wno-unused-function -std=c++11

llama/make/gpu.make

@@ -79,7 +79,7 @@ $(GPU_RUNNER_NAME): $(BUILD_RUNNERS) $(DIST_RUNNERS) $(PAYLOAD_RUNNERS)
 # Build targets
 $(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.cu
 	@-mkdir -p $(dir $@)
-	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CUFLAGS) $(GPU_RUNNER_ARCH_FLAGS) -o $@ $<
+	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CFLAGS) $(GPU_COMPILER_CUFLAGS) $(GPU_RUNNER_ARCH_FLAGS) -o $@ $<
 $(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.c
 	@-mkdir -p $(dir $@)
 	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CFLAGS) -o $@ $<
@@ -89,7 +89,7 @@ $(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.cpp
 $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): TARGET_CGO_LDFLAGS = -L"$(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/" $(CGO_EXTRA_LDFLAGS)
 $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT) *.go ./runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
 	@-mkdir -p $(dir $@)
-	GOARCH=$(ARCH) CGO_LDFLAGS="$(TARGET_CGO_LDFLAGS)" go build $(GPU_GOFLAGS) -tags $(subst $(space),$(comma),$(GPU_RUNNER_CPU_FLAGS) $(GPU_RUNNER_GO_TAGS)) -o $@ ./runner
+	GOARCH=$(ARCH) CGO_LDFLAGS="$(TARGET_CGO_LDFLAGS)" go build -buildmode=pie  $(GPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(GPU_RUNNER_CPU_FLAGS) $(GPU_RUNNER_GO_TAGS)) -o $@ ./runner
 $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(GPU_RUNNER_OBJS) $(DIST_GPU_RUNNER_LIB_DEPS) $(COMMON_HDRS) $(GPU_RUNNER_HDRS)
 	@-mkdir -p $(dir $@)
 	$(CCACHE) $(GPU_COMPILER) --shared $(GPU_RUNNER_DRIVER_LIB_LINK) -L${DIST_GPU_RUNNER_DEPS_DIR} $(foreach lib, $(GPU_RUNNER_LIBS_SHORT), -l$(lib)) $(GPU_RUNNER_OBJS) -o $@
@@ -97,14 +97,17 @@ $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).
 # Distribution targets
 $(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%
 	@-mkdir -p $(dir $@)
-	cp $< $@
-$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
+	$(CP) $< $@
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT) $(GPU_DIST_DEPS_LIBS)
 $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
 	@-mkdir -p $(dir $@)
-	cp $< $@
+	$(CP) $< $@
 $(DIST_GPU_RUNNER_LIB_DEPS): 
 	@-mkdir -p $(dir $@)
-	$(CP) $(GPU_LIB_DIR)/$(notdir $@)* $(dir $@)
+	$(CP) $(GPU_LIB_DIR)/$(notdir $@) $(dir $@)
+$(GPU_DIST_DEPS_LIBS): 
+	@-mkdir -p $(dir $@)
+	$(CP) $(dir $(filter %$(notdir $@),$(GPU_LIBS) $(GPU_TRANSITIVE_LIBS)))/$(notdir $@) $(dir $@)
 
 # Payload targets
 $(RUNNERS_PAYLOAD_DIR)/%/ollama_llama_server.gz: $(RUNNERS_BUILD_DIR)/%/ollama_llama_server 

llama/mllama.cpp

@@ -0,0 +1,900 @@
+// NOTE: This is modified from clip.cpp for Mllama only
+#include "mllama.h"
+
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "ggml.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#include <algorithm>
+#include <cmath>
+#include <cstdarg>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <stdexcept>
+#include <vector>
+
+#define REQUIRE(x)                                           \
+    do {                                                     \
+        if (!(x)) {                                          \
+            throw std::runtime_error("REQUIRE failed: " #x); \
+        }                                                    \
+    } while (0)
+
+#define LOG(fmt, ...) fprintf(stderr, "%s: " fmt "\n", __func__, ##__VA_ARGS__)
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#include <windows.h>
+#if __GLIBCXX__
+#include <cstdio>
+#include <ext/stdio_filebuf.h>
+#include <fcntl.h>
+#endif
+#endif
+
+struct mllama_image {
+    int width;
+    int height;
+
+    int num_channels = 3;
+    int num_tiles = 4;
+
+    int aspect_ratio_id;
+
+    std::vector<float> data;
+};
+
+static std::string format(const char *fmt, ...) {
+    va_list args;
+    va_start(args, fmt);
+    std::vector<char> b(128);
+    int n = vsnprintf(b.data(), b.size(), fmt, args);
+    REQUIRE(n >= 0 && n < b.size());
+    va_end(args);
+    return std::string(b.data(), b.size());
+}
+
+//
+// utilities to get data from a gguf file
+//
+
+static int get_key_index(const gguf_context *ctx, const char *key) {
+    int key_index = gguf_find_key(ctx, key);
+    REQUIRE(key_index != -1);
+    return key_index;
+}
+
+static std::vector<uint32_t> get_u32_array(const gguf_context *ctx, const std::string &key) {
+    const int i = get_key_index(ctx, key.c_str());
+    const int n = gguf_get_arr_n(ctx, i);
+    const uint32_t *data = (uint32_t *)gguf_get_arr_data(ctx, i);
+
+    std::vector<uint32_t> s(n);
+    for (size_t j = 0; j < s.size(); j++) {
+        s[j] = data[j];
+    }
+
+    return s;
+}
+
+static uint32_t get_u32(const gguf_context *ctx, const std::string &key) {
+    return gguf_get_val_u32(ctx, get_key_index(ctx, key.c_str()));
+}
+
+static float get_f32(const gguf_context *ctx, const std::string &key) {
+    return gguf_get_val_f32(ctx, get_key_index(ctx, key.c_str()));
+}
+
+static std::string get_ftype(int ftype) {
+    return ggml_type_name(static_cast<ggml_type>(ftype));
+}
+
+//
+// mllama layers
+//
+
+struct mllama_hparams {
+    uint32_t image_size;
+    uint32_t patch_size;
+    uint32_t hidden_size;
+    uint32_t n_intermediate;
+    uint32_t projection_dim;
+    uint32_t n_head;
+    uint32_t n_layer;
+    uint32_t n_global_layer;
+    uint32_t n_tiles;
+
+    float eps;
+
+    std::vector<bool> intermediate_layers;
+};
+
+struct mllama_layer {
+    // attention
+    struct ggml_tensor *k_w;
+    struct ggml_tensor *k_b;
+    struct ggml_tensor *q_w;
+    struct ggml_tensor *q_b;
+    struct ggml_tensor *v_w;
+    struct ggml_tensor *v_b;
+
+    struct ggml_tensor *o_w;
+    struct ggml_tensor *o_b;
+
+    struct ggml_tensor *attn_gate;
+
+    // layernorm 1
+    struct ggml_tensor *ln_1_w;
+    struct ggml_tensor *ln_1_b;
+
+    // ff
+    struct ggml_tensor *ff_i_w;
+    struct ggml_tensor *ff_i_b;
+
+    struct ggml_tensor *ff_o_w;
+    struct ggml_tensor *ff_o_b;
+
+    struct ggml_tensor *ff_gate;
+
+    // layernorm 2
+    struct ggml_tensor *ln_2_w;
+    struct ggml_tensor *ln_2_b;
+};
+
+struct mllama_vision_model {
+    struct mllama_hparams hparams;
+
+    // embeddings
+    struct ggml_tensor *class_embedding;
+    struct ggml_tensor *patch_embeddings;
+    struct ggml_tensor *position_embeddings;
+    struct ggml_tensor *position_embeddings_gate;
+    struct ggml_tensor *tile_position_embeddings;
+    struct ggml_tensor *tile_position_embeddings_gate;
+    struct ggml_tensor *pre_tile_position_embeddings;
+    struct ggml_tensor *pre_tile_position_embeddings_gate;
+    struct ggml_tensor *post_tile_position_embeddings;
+    struct ggml_tensor *post_tile_position_embeddings_gate;
+
+    struct ggml_tensor *pre_ln_w;
+    struct ggml_tensor *pre_ln_b;
+
+    std::vector<mllama_layer> layers;
+    std::vector<mllama_layer> global_layers;
+
+    struct ggml_tensor *post_ln_w;
+    struct ggml_tensor *post_ln_b;
+
+    struct ggml_tensor *mm_0_w;
+    struct ggml_tensor *mm_0_b;
+};
+
+struct mllama_ctx {
+    struct mllama_vision_model vision_model;
+
+    uint32_t ftype = 1;
+
+    struct gguf_context *ctx_gguf;
+    struct ggml_context *ctx_data;
+
+    std::vector<uint8_t> buf_compute_meta;
+
+    // memory buffers to evaluate the model
+    ggml_backend_buffer_t params_buffer = nullptr;
+
+    ggml_backend_t backend = nullptr;
+    ggml_gallocr_t compute_alloc = nullptr;
+};
+
+static ggml_tensor *mllama_image_build_encoder_layer(
+    struct ggml_context *ctx0, const size_t il, const struct mllama_layer &layer, struct ggml_tensor *embeddings,
+    const float eps, const int hidden_size, const int batch_size, const int n_head, const int d_head) {
+    struct ggml_tensor *cur = embeddings;
+
+    {
+        // layernorm1
+        cur = ggml_norm(ctx0, cur, eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_1_w), layer.ln_1_b);
+        ggml_set_name(cur, format("%d pre layernorm", il).c_str());
+    }
+
+    {
+        // self-attention
+        struct ggml_tensor *Q = ggml_mul_mat(ctx0, layer.q_w, cur);
+        if (layer.q_b != nullptr) {
+            Q = ggml_add(ctx0, Q, layer.q_b);
+        }
+
+        Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, Q->ne[1], batch_size);
+        Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+        ggml_set_name(Q, format("%d query", il).c_str());
+
+        struct ggml_tensor *K = ggml_mul_mat(ctx0, layer.k_w, cur);
+        if (layer.k_b != nullptr) {
+            K = ggml_add(ctx0, K, layer.k_b);
+        }
+
+        K = ggml_reshape_4d(ctx0, K, d_head, n_head, K->ne[1], batch_size);
+        K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+        ggml_set_name(K, format("%d key", il).c_str());
+
+        struct ggml_tensor *V = ggml_mul_mat(ctx0, layer.v_w, cur);
+        if (layer.v_b != nullptr) {
+            V = ggml_add(ctx0, V, layer.v_b);
+        }
+
+        V = ggml_reshape_4d(ctx0, V, d_head, n_head, V->ne[1], batch_size);
+        V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+        ggml_set_name(V, format("%d value", il).c_str());
+
+        struct ggml_tensor *KQ = ggml_mul_mat(ctx0, K, Q);
+        KQ = ggml_scale_inplace(ctx0, KQ, 1.0f / sqrtf((float)d_head));
+        KQ = ggml_soft_max_inplace(ctx0, KQ);
+        ggml_set_name(KQ, format("%d KQ", il).c_str());
+
+        struct ggml_tensor *KQV = ggml_mul_mat(ctx0, V, KQ);
+        KQV = ggml_reshape_4d(ctx0, KQV, d_head, KQV->ne[1], n_head, batch_size);
+        KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+        KQV = ggml_cont_3d(ctx0, KQV, hidden_size, KQV->ne[2], batch_size);
+        ggml_set_name(KQV, format("%d KQV", il).c_str());
+
+        cur = ggml_mul_mat(ctx0, layer.o_w, KQV);
+        if (layer.o_b != nullptr) {
+            cur = ggml_add(ctx0, cur, layer.o_b);
+        }
+        ggml_set_name(cur, format("%d self attention", il).c_str());
+
+        if (layer.attn_gate != nullptr) {
+            cur = ggml_mul_inplace(ctx0, cur, layer.attn_gate);
+            ggml_set_name(cur, format("%d self attention gate", il).c_str());
+        }
+    }
+
+    cur = ggml_add(ctx0, cur, embeddings);
+    ggml_set_name(cur, format("%d residual", il).c_str());
+
+    embeddings = cur;
+
+    {
+        // layernorm2
+        cur = ggml_norm(ctx0, cur, eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_2_w), layer.ln_2_b);
+        ggml_set_name(cur, format("%d post layernorm", il).c_str());
+    }
+
+    {
+        // feed forward
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, layer.ff_i_w, cur), layer.ff_i_b);
+        cur = ggml_gelu_inplace(ctx0, cur);
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, layer.ff_o_w, cur), layer.ff_o_b);
+        ggml_set_name(cur, format("%d feed forward", il).c_str());
+
+        if (layer.ff_gate != nullptr) {
+            cur = ggml_mul_inplace(ctx0, cur, layer.ff_gate);
+            ggml_set_name(cur, format("%d feed forward gate", il).c_str());
+        }
+    }
+
+    // residual 2
+    cur = ggml_add(ctx0, cur, embeddings);
+    ggml_set_name(cur, format("%d residual", il).c_str());
+
+    embeddings = cur;
+
+    return embeddings;
+}
+
+static ggml_cgraph *mllama_image_build_graph(mllama_ctx *ctx, const mllama_image_batch *imgs) {
+    const auto &model = ctx->vision_model;
+    const auto &hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    const int image_size_width = image_size;
+    const int image_size_height = image_size;
+
+    const int patch_size = hparams.patch_size;
+    const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_positions = num_patches + (model.class_embedding == nullptr ? 0 : 1);
+    const int hidden_size = hparams.hidden_size;
+    const int n_head = hparams.n_head;
+    const int d_head = hidden_size / n_head;
+
+    const int batch_size = imgs->size;
+    REQUIRE(batch_size == 1);
+
+    int num_tiles = 4;
+    int num_channels = 3;
+    if (imgs->data != nullptr) {
+        num_tiles = imgs->data[0].num_tiles > 0 ? imgs->data[0].num_tiles : num_tiles;
+        num_channels = imgs->data[0].num_channels > 0 ? imgs->data[0].num_channels : num_channels;
+    }
+
+    struct ggml_init_params params = {
+        ctx->buf_compute_meta.size(), // mem_size
+        ctx->buf_compute_meta.data(), // mem_buffer
+        true,                         // no_alloc
+    };
+
+    struct ggml_context *ctx0 = ggml_init(params);
+    struct ggml_cgraph *gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor *inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, num_channels, num_tiles);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor *inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, num_tiles);
+    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+
+    struct ggml_tensor *aspect_ratios = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, imgs->size);
+    ggml_set_name(aspect_ratios, "aspect_ratios");
+    ggml_set_input(aspect_ratios);
+
+    if (model.pre_tile_position_embeddings != nullptr) {
+        struct ggml_tensor *pre_tile_position_embeddings = ggml_get_rows(ctx0, model.pre_tile_position_embeddings, aspect_ratios);
+        ggml_set_name(pre_tile_position_embeddings, "pre_tile_position_embeddings");
+
+        pre_tile_position_embeddings = ggml_reshape_3d(ctx0, pre_tile_position_embeddings, hidden_size, 1, num_tiles);
+        if (model.pre_tile_position_embeddings_gate != nullptr) {
+            pre_tile_position_embeddings = ggml_mul_inplace(ctx0, pre_tile_position_embeddings, model.pre_tile_position_embeddings_gate);
+        }
+
+        inp = ggml_add(ctx0, inp, pre_tile_position_embeddings);
+    }
+
+    struct ggml_tensor *embeddings = inp;
+
+    if (model.class_embedding != nullptr) {
+        // concat class_embeddings and patch_embeddings
+        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, num_tiles);
+        ggml_set_name(embeddings, "embeddings");
+        ggml_set_input(embeddings);
+        for (int i = 0; i < num_tiles; ++i) {
+            // repeat class embeddings for each tile
+            embeddings = ggml_acc_inplace(ctx0, embeddings, model.class_embedding, embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], i * embeddings->nb[2]);
+        }
+
+        embeddings = ggml_acc_inplace(ctx0, embeddings, inp, embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    }
+
+    struct ggml_tensor *positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    struct ggml_tensor *position_embd = ggml_get_rows(ctx0, model.position_embeddings, positions);
+    if (model.position_embeddings_gate != nullptr) {
+        position_embd = ggml_mul_inplace(ctx0, position_embd, model.position_embeddings_gate);
+    }
+
+    embeddings = ggml_add(ctx0, embeddings, position_embd);
+
+    if (model.tile_position_embeddings != nullptr) {
+        struct ggml_tensor *tile_position_embeddings = ggml_get_rows(ctx0, model.tile_position_embeddings, aspect_ratios);
+        ggml_set_name(tile_position_embeddings, "tile_position_embeddings");
+
+        tile_position_embeddings = ggml_reshape_3d(ctx0, tile_position_embeddings, hidden_size, num_positions, num_tiles);
+        if (model.tile_position_embeddings_gate != nullptr) {
+            tile_position_embeddings = ggml_mul_inplace(ctx0, tile_position_embeddings, model.tile_position_embeddings_gate);
+        }
+
+        embeddings = ggml_add(ctx0, embeddings, tile_position_embeddings);
+    }
+
+    // pre-layernorm
+    if (model.pre_ln_w != nullptr) {
+        embeddings = ggml_mul(ctx0, ggml_norm(ctx0, embeddings, hparams.eps), model.pre_ln_w);
+        if (model.pre_ln_b != nullptr) {
+            embeddings = ggml_add(ctx0, embeddings, model.pre_ln_b);
+        }
+
+        ggml_set_name(embeddings, "pre layernorm");
+    }
+
+    const int num_padding_patches = 8 - (embeddings->ne[1] % 8) % 8;
+
+    embeddings = ggml_pad(ctx0, embeddings, 0, num_padding_patches, 0, 0);
+    embeddings = ggml_view_3d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1] * embeddings->ne[2], batch_size, embeddings->nb[1], embeddings->nb[2] * embeddings->ne[3], 0);
+
+    std::vector<struct ggml_tensor *> intermediate_embeddings;
+
+    // encoder
+    for (size_t il = 0; il < model.layers.size(); il++) {
+        if (hparams.intermediate_layers[il]) {
+            intermediate_embeddings.push_back(embeddings);
+        }
+
+        embeddings = mllama_image_build_encoder_layer(
+            ctx0, il, model.layers[il], embeddings,
+            hparams.eps, hidden_size, batch_size, n_head, d_head);
+    }
+
+    // post-layernorm
+    if (model.post_ln_w != nullptr) {
+        embeddings = ggml_mul(ctx0, ggml_norm(ctx0, embeddings, hparams.eps), model.post_ln_w);
+        if (model.post_ln_b != nullptr) {
+            embeddings = ggml_add(ctx0, embeddings, model.post_ln_b);
+        }
+
+        ggml_set_name(embeddings, "post layernorm");
+    }
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_positions + num_padding_patches, num_tiles);
+
+    if (model.post_tile_position_embeddings != nullptr) {
+        struct ggml_tensor *post_tile_position_embeddings = ggml_get_rows(ctx0, model.post_tile_position_embeddings, aspect_ratios);
+        ggml_set_name(post_tile_position_embeddings, "post_tile_position_embeddings");
+
+        post_tile_position_embeddings = ggml_reshape_3d(ctx0, post_tile_position_embeddings, hidden_size, 1, num_tiles);
+        if (model.post_tile_position_embeddings_gate != nullptr) {
+            post_tile_position_embeddings = ggml_mul(ctx0, post_tile_position_embeddings, model.post_tile_position_embeddings_gate);
+        }
+
+        embeddings = ggml_add(ctx0, embeddings, post_tile_position_embeddings);
+    }
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_tiles * (num_positions + num_padding_patches), 1);
+
+    // global encoder
+    for (size_t il = 0; il < model.global_layers.size(); il++) {
+        embeddings = mllama_image_build_encoder_layer(
+            ctx0, il, model.global_layers[il], embeddings,
+            hparams.eps, hidden_size, batch_size, n_head, d_head);
+    }
+
+    struct ggml_tensor *stacked_embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 0, hidden_size, (num_positions + num_padding_patches) * num_tiles);
+    for (size_t i = 0; i < intermediate_embeddings.size(); ++i) {
+        stacked_embeddings = ggml_concat(ctx0, stacked_embeddings, ggml_reshape_3d(ctx0, intermediate_embeddings[i], 1, intermediate_embeddings[i]->ne[0], intermediate_embeddings[i]->ne[1]), 0);
+    }
+
+    stacked_embeddings = ggml_reshape_4d(ctx0, stacked_embeddings, intermediate_embeddings.size() * hidden_size, num_positions + num_padding_patches, num_tiles, batch_size);
+    stacked_embeddings = ggml_unpad(ctx0, stacked_embeddings, 0, num_padding_patches, 0, 0);
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_positions + num_padding_patches, num_tiles);
+    embeddings = ggml_unpad(ctx0, embeddings, 0, num_padding_patches, 0, 0);
+    embeddings = ggml_concat(ctx0, embeddings, stacked_embeddings, 0);
+
+    // mllama projector
+    embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_0_w, embeddings), model.mm_0_b);
+    ggml_set_name(embeddings, "multi modal projector");
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static struct ggml_tensor *mllama_tensor_load(struct ggml_context *ctx, const char *name, const bool optional) {
+    struct ggml_tensor *cur = ggml_get_tensor(ctx, name);
+    REQUIRE(cur != nullptr || optional);
+    return cur;
+}
+
+static std::vector<struct mllama_layer> mllama_layers_load(struct ggml_context *ctx, const char *prefix, const int n) {
+    std::vector<struct mllama_layer> layers(n);
+    for (size_t i = 0; i < layers.size(); i++) {
+        auto &layer = layers[i];
+        layer.ln_1_w = mllama_tensor_load(ctx, format("%s.blk.%d.ln1.weight", prefix, i).c_str(), false);
+        layer.ln_1_b = mllama_tensor_load(ctx, format("%s.blk.%d.ln1.bias", prefix, i).c_str(), false);
+        layer.ln_2_w = mllama_tensor_load(ctx, format("%s.blk.%d.ln2.weight", prefix, i).c_str(), false);
+        layer.ln_2_b = mllama_tensor_load(ctx, format("%s.blk.%d.ln2.bias", prefix, i).c_str(), false);
+
+        layer.k_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_k.weight", prefix, i).c_str(), false);
+        layer.k_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_k.bias", prefix, i).c_str(), true);
+        layer.q_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_q.weight", prefix, i).c_str(), false);
+        layer.q_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_q.bias", prefix, i).c_str(), true);
+        layer.v_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_v.weight", prefix, i).c_str(), false);
+        layer.v_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_v.bias", prefix, i).c_str(), true);
+        layer.o_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_out.weight", prefix, i).c_str(), false);
+        layer.o_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_out.bias", prefix, i).c_str(), true);
+
+        layer.ff_i_w = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_down.weight", prefix, i).c_str(), false);
+        layer.ff_i_b = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_down.bias", prefix, i).c_str(), false);
+        layer.ff_o_w = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_up.weight", prefix, i).c_str(), false);
+        layer.ff_o_b = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_up.bias", prefix, i).c_str(), false);
+
+        layer.attn_gate = mllama_tensor_load(ctx, format("%s.blk.%d.attn_gate", prefix, i).c_str(), true);
+        layer.ff_gate = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_gate", prefix, i).c_str(), true);
+    }
+
+    return layers;
+}
+
+// read and create ggml_context containing the tensors and their data
+struct mllama_ctx *mllama_model_load(const char *fname, const int verbosity = 1) {
+    struct ggml_context *meta = nullptr;
+
+    struct gguf_init_params params = {
+        true,  // no_alloc
+        &meta, // ctx
+    };
+
+    struct gguf_context *ctx = gguf_init_from_file(fname, params);
+    REQUIRE(ctx != nullptr);
+
+    if (verbosity >= 1) {
+        const int n_tensors = gguf_get_n_tensors(ctx);
+        const int n_kv = gguf_get_n_kv(ctx);
+        const std::string ftype = get_ftype(get_u32(ctx, "general.file_type"));
+        const int idx_desc = get_key_index(ctx, "general.description");
+        const std::string description = gguf_get_val_str(ctx, idx_desc);
+        const int idx_name = gguf_find_key(ctx, "general.name");
+        if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
+            const std::string name = gguf_get_val_str(ctx, idx_name);
+            LOG("model name:   %s", name.c_str());
+        }
+        LOG("description:  %s", description.c_str());
+        LOG("GGUF version: %d", gguf_get_version(ctx));
+        LOG("alignment:    %zu", gguf_get_alignment(ctx));
+        LOG("n_tensors:    %d", n_tensors);
+        LOG("n_kv:         %d", n_kv);
+        LOG("ftype:        %s", ftype.c_str());
+        LOG("");
+    }
+    const int n_tensors = gguf_get_n_tensors(ctx);
+
+    mllama_ctx *new_mllama = new mllama_ctx{};
+
+#ifdef GGML_USE_CUDA
+    new_mllama->backend = ggml_backend_cuda_init(0);
+    LOG("vision using CUDA backend");
+#endif
+
+#ifdef GGML_USE_METAL
+    new_mllama->backend = ggml_backend_metal_init();
+    LOG("vision using Metal backend");
+#endif
+
+#ifdef GGML_USE_CANN
+    new_mllama->backend = ggml_backend_cann_init(0);
+    LOG("vision using CANN backend");
+#endif
+
+#ifdef GGML_USE_VULKAN
+    new_mllama->backend = ggml_backend_vk_init(0);
+    LOG("vision using Vulkan backend");
+#endif
+
+    if (!new_mllama->backend) {
+        new_mllama->backend = ggml_backend_cpu_init();
+        LOG("vision using CPU backend");
+    }
+
+    // load tensors
+    {
+        std::vector<uint8_t> read_buf;
+        struct ggml_init_params params = {
+            (n_tensors + 1) * ggml_tensor_overhead(), // mem_size
+            nullptr,                                  // mem_buffer
+            true,                                     // no_alloc
+        };
+
+        new_mllama->ctx_data = ggml_init(params);
+        if (!new_mllama->ctx_data) {
+            LOG("ggml_init() failed");
+            mllama_free(new_mllama);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+#ifdef _WIN32
+        int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
+        if (!wlen) {
+            return NULL;
+        }
+        wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
+        wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wbuf, wlen);
+        if (!wlen) {
+            free(wbuf);
+            return NULL;
+        }
+#if __GLIBCXX__
+        int fd = _wopen(wbuf, _O_RDONLY | _O_BINARY);
+        __gnu_cxx::stdio_filebuf<char> buffer(fd, std::ios_base::in);
+        std::istream fin(&buffer);
+#else // MSVC
+        // unused in our current build
+        auto fin = std::ifstream(wbuf, std::ios::binary);
+#endif
+        free(wbuf);
+#else
+        auto fin = std::ifstream(fname, std::ios::binary);
+#endif
+        if (!fin) {
+            LOG("cannot open model file for loading tensors\n");
+            mllama_free(new_mllama);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+        // add tensors to context
+        for (int i = 0; i < n_tensors; ++i) {
+            const char *name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor *t = ggml_get_tensor(meta, name);
+            struct ggml_tensor *cur = ggml_dup_tensor(new_mllama->ctx_data, t);
+            ggml_set_name(cur, name);
+        }
+
+        // alloc memory and offload data
+        new_mllama->params_buffer = ggml_backend_alloc_ctx_tensors(new_mllama->ctx_data, new_mllama->backend);
+        for (int i = 0; i < n_tensors; ++i) {
+            const char *name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor *cur = ggml_get_tensor(new_mllama->ctx_data, name);
+            const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
+            fin.seekg(offset, std::ios::beg);
+            if (!fin) {
+                LOG("failed to seek for tensor %s\n", name);
+                mllama_free(new_mllama);
+                gguf_free(ctx);
+                return nullptr;
+            }
+            int num_bytes = ggml_nbytes(cur);
+            if (ggml_backend_buffer_is_host(new_mllama->params_buffer)) {
+                // for the CPU and Metal backend, we can read directly into the tensor
+                fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+            } else {
+                // read into a temporary buffer first, then copy to device memory
+                read_buf.resize(num_bytes);
+                fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+            }
+        }
+
+#if defined(_WIN32) && defined(__GLIBCXX__)
+        close(fd);
+#else
+        fin.close();
+#endif
+    }
+
+    // vision model
+    // load vision model
+    auto &vision_model = new_mllama->vision_model;
+    auto &hparams = vision_model.hparams;
+    hparams.hidden_size = get_u32(ctx, "mllama.vision.embedding_length");
+    hparams.n_head = get_u32(ctx, "mllama.vision.attention.head_count");
+    hparams.n_intermediate = get_u32(ctx, "mllama.vision.feed_forward_length");
+    hparams.n_layer = get_u32(ctx, "mllama.vision.block_count");
+    hparams.n_global_layer = get_u32(ctx, "mllama.vision.global.block_count");
+    hparams.n_tiles = get_u32(ctx, "mllama.vision.max_num_tiles");
+    hparams.image_size = get_u32(ctx, "mllama.vision.image_size");
+    hparams.patch_size = get_u32(ctx, "mllama.vision.patch_size");
+    hparams.projection_dim = get_u32(ctx, "mllama.vision.projection_dim");
+    hparams.eps = get_f32(ctx, "mllama.vision.attention.layer_norm_epsilon");
+
+    std::vector<uint32_t> intermediate_layers_indices = get_u32_array(ctx, "mllama.vision.intermediate_layers_indices");
+    hparams.intermediate_layers.resize(hparams.n_layer);
+    for (size_t i = 0; i < intermediate_layers_indices.size(); i++) {
+        hparams.intermediate_layers[intermediate_layers_indices[i]] = true;
+    }
+
+    if (verbosity >= 2) {
+        LOG("");
+        LOG("vision model hparams");
+        LOG("image_size         %d", hparams.image_size);
+        LOG("patch_size         %d", hparams.patch_size);
+        LOG("v_hidden_size      %d", hparams.hidden_size);
+        LOG("v_n_intermediate   %d", hparams.n_intermediate);
+        LOG("v_projection_dim   %d", hparams.projection_dim);
+        LOG("v_n_head           %d", hparams.n_head);
+        LOG("v_n_layer          %d", hparams.n_layer);
+        LOG("v_n_global_layer   %d", hparams.n_global_layer);
+        LOG("v_eps              %f", hparams.eps);
+    }
+
+    vision_model.class_embedding = mllama_tensor_load(new_mllama->ctx_data, "v.class_embd", true);
+    vision_model.patch_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.patch_embd.weight", true);
+
+    vision_model.position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.position_embd.weight", true);
+    vision_model.position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.position_embd.gate", true);
+
+    vision_model.pre_ln_w = mllama_tensor_load(new_mllama->ctx_data, "v.pre_ln.weight", true);
+    vision_model.pre_ln_b = mllama_tensor_load(new_mllama->ctx_data, "v.pre_ln.bias", true);
+    vision_model.post_ln_w = mllama_tensor_load(new_mllama->ctx_data, "v.post_ln.weight", true);
+    vision_model.post_ln_b = mllama_tensor_load(new_mllama->ctx_data, "v.post_ln.bias", true);
+
+    vision_model.tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.tile_position_embd.weight", true);
+    vision_model.tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.tile_position_embd.gate", true);
+
+    vision_model.pre_tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.pre_tile_position_embd.weight", true);
+    vision_model.pre_tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.pre_tile_position_embd.gate", true);
+
+    vision_model.post_tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.post_tile_position_embd.weight", true);
+    vision_model.post_tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.post_tile_position_embd.gate", true);
+
+    vision_model.mm_0_w = mllama_tensor_load(new_mllama->ctx_data, "mm.0.weight", false);
+    vision_model.mm_0_b = mllama_tensor_load(new_mllama->ctx_data, "mm.0.bias", false);
+
+    vision_model.layers = mllama_layers_load(new_mllama->ctx_data, "v", hparams.n_layer);
+    vision_model.global_layers = mllama_layers_load(new_mllama->ctx_data, "v.global", hparams.n_global_layer);
+
+    ggml_free(meta);
+
+    new_mllama->ctx_gguf = ctx;
+
+    {
+        // measure mem requirement and allocate
+        new_mllama->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
+        new_mllama->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_mllama->backend));
+        struct mllama_image_batch batch;
+        batch.size = 1;
+        ggml_cgraph *gf = mllama_image_build_graph(new_mllama, &batch);
+        ggml_gallocr_reserve(new_mllama->compute_alloc, gf);
+        size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_mllama->compute_alloc, 0);
+        LOG("compute allocated memory: %.2f MB", compute_memory_buffer_size / 1024.0 / 1024.0);
+    }
+
+    return new_mllama;
+}
+
+struct mllama_image *mllama_image_init() {
+    return new mllama_image();
+}
+
+void mllama_image_free(struct mllama_image *img) { delete img; }
+void mllama_image_batch_free(struct mllama_image_batch *batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
+    }
+}
+
+bool mllama_image_load_from_data(const void *data, const int n, const int width, const int height, const int num_channels, const int num_tiles, const int aspect_ratio_id, struct mllama_image *img) {
+    img->width = width;
+    img->height = height;
+    img->num_channels = num_channels;
+    img->num_tiles = num_tiles;
+    img->aspect_ratio_id = aspect_ratio_id;
+    img->data.resize(n);
+
+    memcpy(img->data.data(), data, n);
+    return true;
+}
+
+inline int mllama(int x, int lower, int upper) {
+    return std::max(lower, std::min(x, upper));
+}
+
+void mllama_free(mllama_ctx *ctx) {
+    ggml_free(ctx->ctx_data);
+    gguf_free(ctx->ctx_gguf);
+
+    ggml_backend_buffer_free(ctx->params_buffer);
+    ggml_backend_free(ctx->backend);
+    ggml_gallocr_free(ctx->compute_alloc);
+    delete ctx;
+}
+
+bool mllama_image_encode(struct mllama_ctx *ctx, const int n_threads, mllama_image *img, float *vec) {
+    mllama_image_batch imgs{};
+    imgs.size = 1;
+    imgs.data = img;
+    return mllama_image_batch_encode(ctx, n_threads, &imgs, vec);
+}
+
+bool mllama_image_batch_encode(mllama_ctx *ctx, const int n_threads, const mllama_image_batch *imgs, float *vec) {
+    int batch_size = imgs->size;
+    REQUIRE(batch_size == 1);
+
+    // build the inference graph
+    ggml_cgraph *gf = mllama_image_build_graph(ctx, imgs);
+    ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
+
+    // set inputs
+    const auto &model = ctx->vision_model;
+    const auto &hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    int image_size_width = image_size;
+    int image_size_height = image_size;
+
+    const int patch_size = hparams.patch_size;
+    const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_positions = num_patches + (model.class_embedding == nullptr ? 0 : 1);
+
+    {
+        struct ggml_tensor *inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
+        ggml_backend_tensor_set(inp_raw, imgs->data[0].data.data(), 0, ggml_nbytes(inp_raw));
+    }
+
+    {
+        struct ggml_tensor *embeddings = ggml_graph_get_tensor(gf, "embeddings");
+        if (embeddings != nullptr) {
+            void *zeros = malloc(ggml_nbytes(embeddings));
+            memset(zeros, 0, ggml_nbytes(embeddings));
+            ggml_backend_tensor_set(embeddings, zeros, 0, ggml_nbytes(embeddings));
+            free(zeros);
+        }
+    }
+
+    {
+        struct ggml_tensor *positions = ggml_graph_get_tensor(gf, "positions");
+        if (positions != nullptr) {
+            int *positions_data = (int *)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = i;
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+    }
+
+    {
+        struct ggml_tensor *aspect_ratios = ggml_graph_get_tensor(gf, "aspect_ratios");
+        if (aspect_ratios != nullptr) {
+            int *aspect_ratios_data = (int *)malloc(ggml_nbytes(aspect_ratios));
+            aspect_ratios_data[0] = imgs->data[0].aspect_ratio_id;
+            ggml_backend_tensor_set(aspect_ratios, aspect_ratios_data, 0, ggml_nbytes(aspect_ratios));
+            free(aspect_ratios_data);
+        }
+    }
+
+    if (ggml_backend_is_cpu(ctx->backend)) {
+        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
+    }
+
+    ggml_backend_graph_compute(ctx->backend, gf);
+
+    // the last node is the embedding tensor
+    struct ggml_tensor *embeddings = ggml_graph_node(gf, ggml_graph_n_nodes(gf) - 1);
+
+    // copy the embeddings to the location passed by the user
+    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
+
+    return true;
+}
+
+int32_t mllama_image_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.image_size;
+}
+
+int32_t mllama_patch_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.patch_size;
+}
+
+int32_t mllama_hidden_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.hidden_size;
+}
+
+int mllama_n_patches(const struct mllama_ctx *ctx) {
+    const auto &hparams = ctx->vision_model.hparams;
+    return (hparams.image_size / hparams.patch_size) * (hparams.image_size / hparams.patch_size);
+}
+
+int mllama_n_positions(const struct mllama_ctx *ctx) {
+    return mllama_n_patches(ctx) + (ctx->vision_model.class_embedding == nullptr ? 0 : 1);
+}
+
+int mllama_n_tiles(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.n_tiles;
+}
+
+int mllama_n_embd(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.projection_dim;
+}
+
+size_t mllama_n_embd_bytes(const struct mllama_ctx *ctx) {
+    return mllama_n_positions(ctx) * mllama_n_embd(ctx) * mllama_n_tiles(ctx) * sizeof(float);
+}

llama/mllama.h

@@ -0,0 +1,61 @@
+#ifndef MLLAMA_H
+#define MLLAMA_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#if defined(_WIN32) && !defined(__MINGW32__)
+#ifdef LLAMA_BUILD
+#define MLLAMA_API __declspec(dllexport)
+#else
+#define MLLAMA_API __declspec(dllimport)
+#endif
+#else
+#define MLLAMA_API __attribute__((visibility("default")))
+#endif
+#else
+#define MLLAMA_API
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct mllama_ctx;
+
+struct mllama_image_batch {
+    struct mllama_image *data;
+    size_t size;
+};
+
+MLLAMA_API struct mllama_ctx *mllama_model_load(const char *fname, int verbosity);
+MLLAMA_API struct mllama_ctx *mllama_model_load_cpu(const char *fname, int verbosity);
+
+MLLAMA_API void mllama_free(struct mllama_ctx *ctx);
+
+MLLAMA_API int32_t mllama_image_size(const struct mllama_ctx *ctx);
+MLLAMA_API int32_t mllama_patch_size(const struct mllama_ctx *ctx);
+MLLAMA_API int32_t mllama_hidden_size(const struct mllama_ctx *ctx);
+
+MLLAMA_API int mllama_n_patches(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_positions(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_tiles(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_embd(const struct mllama_ctx *ctx);
+MLLAMA_API size_t mllama_n_embd_bytes(const struct mllama_ctx *ctx);
+
+MLLAMA_API struct mllama_image *mllama_image_init();
+
+MLLAMA_API void mllama_image_free(struct mllama_image *img);
+MLLAMA_API void mllama_image_batch_free(struct mllama_image_batch *batch);
+
+MLLAMA_API bool mllama_image_load_from_data(const void *data, const int n, const int nx, const int ny, const int nc, const int nt, const int aspect_ratio_id, struct mllama_image *img);
+
+MLLAMA_API bool mllama_image_encode(struct mllama_ctx *ctx, int n_threads, struct mllama_image *img, float *vec);
+MLLAMA_API bool mllama_image_batch_encode(struct mllama_ctx *ctx, int n_threads, const struct mllama_image_batch *imgs, float *vec);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // MLLAMA_H

llama/patches/0001-cuda.patch

@@ -1,3 +1,14 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Thu, 6 Jun 2024 23:55:47 -0700
+Subject: [PATCH] cuda
+
+---
+ ggml/include/ggml-cuda.h | 2 ++
+ ggml/src/ggml-backend.c  | 5 +++++
+ ggml/src/ggml-cuda.cu    | 6 ++++--
+ 3 files changed, 11 insertions(+), 2 deletions(-)
+
 diff --git a/ggml/include/ggml-cuda.h b/ggml/include/ggml-cuda.h
 index 71bb6dcf..08be0895 100644
 --- a/ggml/include/ggml-cuda.h

llama/patches/0002-pretokenizer.patch

@@ -1,7 +1,7 @@
-From 0e531d69786c4a96a3a2bcf7b2d576bd6f7edf25 Mon Sep 17 00:00:00 2001
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
 From: Michael Yang <mxyng@pm.me>
 Date: Mon, 16 Sep 2024 15:53:13 -0700
-Subject: [PATCH] 05-default-pretokenizer.diff
+Subject: [PATCH] pretokenizer
 
 ---
  src/llama.cpp | 14 +++-----------
@@ -39,6 +39,3 @@ index 4c0a1bb6..800dfb95 100644
              }
          } else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
              vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
--- 
-2.39.3 (Apple Git-146)
-

llama/patches/0003-metal.patch

@@ -1,6 +1,6 @@
-From 91d3f886f1645b38d9658c0e125603e8d5338146 Mon Sep 17 00:00:00 2001
-From: nobody <>
-Date: Tue, 1 Oct 2024 13:55:01 -0600
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:12 -0700
 Subject: [PATCH] metal
 
 ---
@@ -52,6 +52,3 @@ index 9da08fe2..3a433703 100644
  
                          // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                          // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
--- 
-2.39.3 (Apple Git-146)
-

llama/patches/0004-ggml-metal.patch

@@ -1,3 +1,12 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Wed, 12 Jun 2024 12:18:40 -0700
+Subject: [PATCH] ggml-metal
+
+---
+ ggml/src/ggml-metal.m | 4 ++--
+ 1 file changed, 2 insertions(+), 2 deletions(-)
+
 diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
 index 3a433703..829c5e39 100644
 --- a/ggml/src/ggml-metal.m

llama/patches/0005-embeddings.patch

@@ -1,6 +1,6 @@
-From 235b6d876a74cb09abe26985fa89ebe5bfc9f562 Mon Sep 17 00:00:00 2001
-From: Gabe Goodhart <ghart@us.ibm.com>
-Date: Thu, 19 Sep 2024 17:06:17 -0600
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:14 -0700
 Subject: [PATCH] embeddings
 
 ---
@@ -8,10 +8,10 @@ Subject: [PATCH] embeddings
  1 file changed, 9 insertions(+), 6 deletions(-)
 
 diff --git a/src/llama.cpp b/src/llama.cpp
-index 1a8e0c51..e55ec3f8 100644
+index 800dfb95..a639522d 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -16516,7 +16516,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
+@@ -16920,7 +16920,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
      const auto n_embd  = hparams.n_embd;
  
      // TODO: use a per-batch flag for logits presence instead
@@ -20,7 +20,7 @@ index 1a8e0c51..e55ec3f8 100644
      const bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
  
      const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
-@@ -16794,20 +16794,23 @@ static int llama_decode_internal(
+@@ -17192,20 +17192,23 @@ static int llama_decode_internal(
              // no output
              res  = nullptr;
              embd = nullptr;
@@ -49,6 +49,3 @@ index 1a8e0c51..e55ec3f8 100644
          // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
  
          ggml_backend_sched_alloc_graph(lctx.sched, gf);
--- 
-2.39.3 (Apple Git-146)
-

llama/patches/0006-clip-unicode.patch

@@ -1,3 +1,12 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:15 -0700
+Subject: [PATCH] clip-unicode
+
+---
+ examples/llava/clip.cpp | 40 +++++++++++++++++++++++++++++++++++++++-
+ 1 file changed, 39 insertions(+), 1 deletion(-)
+
 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
 index 14e02c8d..6e849d8e 100644
 --- a/examples/llava/clip.cpp

llama/patches/0007-solar-pro.patch

@@ -1,5 +1,21 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:16 -0700
+Subject: [PATCH] solar-pro
+
+solar-pro introduces block skip connections where blocks are connected
+to other, non-sequential blocks with a scale multiple
+
+this change adds 4 new keys to store the skip connections and one new
+tensor to store the scalar. the scalar is implemented a 1-dimensional
+tensor with 2 elements dervied from the model's bskcn_tv configuration.
+in general, the values are (bskcn_tv, 1 - bskcn_tv)
+---
+ src/llama.cpp | 269 +++++++++++++++++++++++++++++++++++++++++++++++---
+ 1 file changed, 255 insertions(+), 14 deletions(-)
+
 diff --git a/src/llama.cpp b/src/llama.cpp
-index bdad28b3..1fe6189a 100644
+index a639522d..83b80b59 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
 @@ -217,6 +217,7 @@ enum llm_arch {

llama/patches/0008-conditional-fattn.patch

@@ -1,8 +1,17 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Daniel Hiltgen <daniel@ollama.com>
+Date: Wed, 9 Oct 2024 17:26:23 -0700
+Subject: [PATCH] conditional-fattn
+
+---
+ ggml/src/ggml-cuda.cu | 2 ++
+ 1 file changed, 2 insertions(+)
+
 diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
-index 8a844b02..61d61542 100644
+index 809d6ab1..fe77b81c 100644
 --- a/ggml/src/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda.cu
-@@ -2310,9 +2310,11 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
+@@ -2347,9 +2347,11 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
          case GGML_OP_ARGSORT:
              ggml_cuda_op_argsort(ctx, dst);
              break;

llama/patches/0009-blas.patch

@@ -1,3 +1,12 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Jesse Gross <jesse@ollama.com>
+Date: Mon, 30 Sep 2024 16:31:04 -0700
+Subject: [PATCH] blas
+
+---
+ ggml/src/ggml-blas.cpp | 4 ++++
+ 1 file changed, 4 insertions(+)
+
 diff --git a/ggml/src/ggml-blas.cpp b/ggml/src/ggml-blas.cpp
 index 6d99c6be..8e1ab99d 100644
 --- a/ggml/src/ggml-blas.cpp

llama/patches/0010-add-mllama-support.patch

@@ -0,0 +1,690 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Thu, 17 Oct 2024 15:18:22 -0700
+Subject: [PATCH] add mllama support
+
+mllama adds cross-attention layers to the standard llama architecture
+it also requires a way to input a new tensor: cross_attention_state
+once per generation
+
+cross-attention layers don't change and so they are cached in the
+kv cache once per run
+
+remaining is to implement the cross attention mask
+---
+ include/llama.h |   4 +
+ src/llama.cpp   | 456 ++++++++++++++++++++++++++++++++++++++++++++++--
+ 2 files changed, 447 insertions(+), 13 deletions(-)
+
+diff --git a/include/llama.h b/include/llama.h
+index 7cae1bbe..122e3cf1 100644
+--- a/include/llama.h
++++ b/include/llama.h
+@@ -423,6 +423,10 @@ extern "C" {
+                      struct llama_model * model,
+             struct llama_context_params   params);
+ 
++    // TODO (jmorganca): this should most likely be passed in as part of a batch
++    // and not set on the context for all batches.
++    LLAMA_API void llama_set_cross_attn_state(struct llama_context * ctx, float * cross_attn_state);
++
+     // Frees all allocated memory
+     LLAMA_API void llama_free(struct llama_context * ctx);
+ 
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 83b80b59..b189a19a 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -169,6 +169,7 @@ static std::string format(const char * fmt, ...) {
+ 
+ enum llm_arch {
+     LLM_ARCH_LLAMA,
++    LLM_ARCH_MLLAMA,
+     LLM_ARCH_FALCON,
+     LLM_ARCH_BAICHUAN,
+     LLM_ARCH_GROK,
+@@ -223,6 +224,7 @@ enum llm_arch {
+ 
+ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+     { LLM_ARCH_LLAMA,           "llama"        },
++    { LLM_ARCH_MLLAMA,          "mllama"       },
+     { LLM_ARCH_FALCON,          "falcon"       },
+     { LLM_ARCH_GROK,            "grok"         },
+     { LLM_ARCH_GPT2,            "gpt2"         },
+@@ -330,6 +332,7 @@ enum llm_kv {
+     LLM_KV_ATTENTION_SLIDING_WINDOW,
+     LLM_KV_ATTENTION_SCALE,
+     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
++    LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
+ 
+     LLM_KV_ROPE_DIMENSION_COUNT,
+     LLM_KV_ROPE_FREQ_BASE,
+@@ -439,6 +442,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+     { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"           },
+     { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                    },
+     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection.%d" },
++    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers"   },
+ 
+     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
+     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+@@ -613,6 +617,14 @@ enum llm_tensor {
+     LLM_TENSOR_CLS,
+     LLM_TENSOR_CLS_OUT,
+     LLM_TENSOR_BSKCN_TV,
++    LLM_TENSOR_CROSS_ATTN_K_NORM,
++    LLM_TENSOR_CROSS_ATTN_K_PROJ,
++    LLM_TENSOR_CROSS_ATTN_O_PROJ,
++    LLM_TENSOR_CROSS_ATTN_Q_NORM,
++    LLM_TENSOR_CROSS_ATTN_Q_PROJ,
++    LLM_TENSOR_CROSS_ATTN_V_PROJ,
++    LLM_TENSOR_CROSS_ATTN_ATTN_GATE,
++    LLM_TENSOR_CROSS_ATTN_MLP_GATE,
+ };
+ 
+ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
+@@ -642,6 +654,40 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
+             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+         },
+     },
++    {
++        LLM_ARCH_MLLAMA,
++        {
++            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
++            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
++            { LLM_TENSOR_OUTPUT,          "output" },
++            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
++            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
++            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
++            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
++            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
++            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
++            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
++            { LLM_TENSOR_FFN_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_CROSS_ATTN_K_NORM,    "blk.%d.cross_attn_k_norm" },
++            { LLM_TENSOR_CROSS_ATTN_K_PROJ,    "blk.%d.cross_attn_k_proj" },
++            { LLM_TENSOR_CROSS_ATTN_O_PROJ,    "blk.%d.cross_attn_o_proj" },
++            { LLM_TENSOR_CROSS_ATTN_Q_NORM,    "blk.%d.cross_attn_q_norm" },
++            { LLM_TENSOR_CROSS_ATTN_Q_PROJ,    "blk.%d.cross_attn_q_proj" },
++            { LLM_TENSOR_CROSS_ATTN_V_PROJ,    "blk.%d.cross_attn_v_proj" },
++            { LLM_TENSOR_CROSS_ATTN_ATTN_GATE, "blk.%d.cross_attn_attn_gate" },
++            { LLM_TENSOR_CROSS_ATTN_MLP_GATE,  "blk.%d.cross_attn_mlp_gate" },
++        },
++    },
+     {
+         LLM_ARCH_BAICHUAN,
+         {
+@@ -2390,6 +2436,7 @@ enum e_model {
+     MODEL_40B,
+     MODEL_65B,
+     MODEL_70B,
++    MODEL_90B,
+     MODEL_236B,
+     MODEL_314B,
+     MODEL_SMALL,
+@@ -2434,6 +2481,7 @@ struct llama_hparams {
+     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+ 
+     std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
++    std::array<uint32_t, LLAMA_MAX_LAYERS> cross_attn_layers;
+ 
+     uint32_t n_layer_dense_lead = 0;
+     uint32_t n_lora_q = 0;
+@@ -2502,10 +2550,11 @@ struct llama_hparams {
+         if (this->n_expert      != other.n_expert)      return true;
+         if (this->n_expert_used != other.n_expert_used) return true;
+ 
+-        if (this->n_head_arr    != other.n_head_arr)    return true;
+-        if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
+-        if (this->n_ff_arr      != other.n_ff_arr)      return true;
+-        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
++        if (this->n_head_arr        != other.n_head_arr)        return true;
++        if (this->n_head_kv_arr     != other.n_head_kv_arr)     return true;
++        if (this->n_ff_arr          != other.n_ff_arr)          return true;
++        if (this->n_bskcn_arr       != other.n_bskcn_arr)       return true;
++        if (this->cross_attn_layers != other.cross_attn_layers) return true;
+ 
+         if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
+         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
+@@ -2623,6 +2672,10 @@ struct llama_hparams {
+ 
+         GGML_ABORT("fatal error");
+     }
++
++    bool cross_attention_layer(uint32_t il) const {
++        return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
++    }
+ };
+ 
+ static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+@@ -2806,6 +2859,16 @@ struct llama_layer {
+     struct ggml_tensor * ffn_down_scale;
+ 
+     struct ggml_tensor * bskcn_tv;
++
++    // cross attention
++    struct ggml_tensor * cross_attn_k_norm;
++    struct ggml_tensor * cross_attn_k_proj;
++    struct ggml_tensor * cross_attn_o_proj;
++    struct ggml_tensor * cross_attn_q_norm;
++    struct ggml_tensor * cross_attn_q_proj;
++    struct ggml_tensor * cross_attn_v_proj;
++    struct ggml_tensor * cross_attn_attn_gate;
++    struct ggml_tensor * cross_attn_mlp_gate;
+ };
+ 
+ // very similar to llama_batch,
+@@ -3452,6 +3515,12 @@ struct llama_context {
+     struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
+     struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
+     struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
++
++    // TODO (jmorganca): this should most likely be passed in as part of a batch
++    // and not set on the context for all batches.
++    float * cross_attn_state = nullptr;
++    bool cross_attn_state_first_pass = true;
++    struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061]
+ };
+ 
+ struct llama_lora_weight {
+@@ -3686,6 +3755,18 @@ static bool llama_kv_cache_init(
+     cache.v_l.reserve(n_layer);
+ 
+     for (int i = 0; i < (int) n_layer; i++) {
++        // for cross attention layers
++        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layer(i)) {
++            struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front();
++            ggml_tensor * k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
++            ggml_tensor * v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
++            ggml_format_name(k, "cache_k_l%d", i);
++            ggml_format_name(v, "cache_v_l%d", i);
++            cache.k_l.push_back(k);
++            cache.v_l.push_back(v);
++            continue;
++        }
++
+         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
+         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
+ 
+@@ -5460,12 +5541,14 @@ static void llm_load_hparams(
+     }
+ 
+     // zero-out the per-layer hparams
+-    std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
+-    std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
+-    std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
++    std::fill(hparams.n_head_arr.begin(),             hparams.n_head_arr.end(),        0);
++    std::fill(hparams.n_head_kv_arr.begin(),          hparams.n_head_kv_arr.end(),     0);
++    std::fill(hparams.n_ff_arr.begin(),               hparams.n_ff_arr.end(),          0);
++    std::fill(hparams.cross_attn_layers.begin(),      hparams.cross_attn_layers.end(), -1);
+ 
+-    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff_arr,   hparams.n_layer);
+-    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer);
++    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,       hparams.n_ff_arr,          hparams.n_layer);
++    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT,      hparams.n_head_arr,        hparams.n_layer);
++    ml.get_arr(LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, hparams.cross_attn_layers, false);
+ 
+     // n_head_kv is optional, default to n_head
+     hparams.n_head_kv_arr = hparams.n_head_arr;
+@@ -5514,7 +5597,7 @@ static void llm_load_hparams(
+ 
+         ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
+ 
+-        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
++        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_MLLAMA || model.arch == LLM_ARCH_FALCON) {
+             if (hparams.n_rot != hparams.n_embd_head_k) {
+                 throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
+             }
+@@ -5554,6 +5637,16 @@ static void llm_load_hparams(
+                     }
+                 }
+             } break;
++        case LLM_ARCH_MLLAMA:
++            {
++                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
++
++                switch (hparams.n_layer) {
++                    case 40: model.type = e_model::MODEL_11B; break;
++                    case 100: model.type = e_model::MODEL_90B; break;
++                    default: model.type = e_model::MODEL_UNKNOWN;
++                }
++            } break;
+         case LLM_ARCH_MINICPM:
+             {
+                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+@@ -7249,6 +7342,55 @@ static bool llm_load_tensors(
+                         layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                     }
+                 } break;
++            case LLM_ARCH_MLLAMA:
++                {
++                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8});
++
++                    // output
++                    {
++                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
++                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
++
++                        // if output is NULL, init from the input tok embed
++                        if (model.output == NULL) {
++                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
++                        }
++                    }
++
++                    for (int i = 0; i < n_layer; ++i) {
++                        ggml_context * ctx_layer = ctx_for_layer(i);
++                        ggml_context * ctx_split = ctx_for_layer_split(i);
++
++                        auto & layer = model.layers[i];
++
++                        if (hparams.cross_attention_layer(i)) {
++                            layer.cross_attn_k_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_NORM,   "weight", i), {128});
++                            layer.cross_attn_k_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_PROJ,   "weight", i), {n_embd, 1024});
++                            layer.cross_attn_o_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_O_PROJ,   "weight", i), {n_embd, n_embd});
++                            layer.cross_attn_q_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128});
++                            layer.cross_attn_q_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd});
++                            layer.cross_attn_v_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024});
++                            layer.cross_attn_attn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1});
++                            layer.cross_attn_mlp_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1});
++                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
++                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
++                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
++                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
++                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
++                        } else {
++                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
++                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
++                            layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
++                            layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
++                            layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
++                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
++                            layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
++                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
++                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
++                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
++                        }
++                    }
++                } break;
+             case LLM_ARCH_GROK:
+                 {
+                     if (n_expert == 0) {
+@@ -9093,7 +9235,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
+ 
+         if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
+             model.hparams.n_vocab != model.vocab.id_to_token.size()) {
+-            throw std::runtime_error("vocab size mismatch");
++            LLAMA_LOG_WARN("%s: vocab mismatch %u !- %zu ...\n", __func__, model.hparams.n_vocab, model.vocab.id_to_token.size());
+         }
+ 
+         if (params.vocab_only) {
+@@ -9178,7 +9320,7 @@ static struct ggml_tensor * llm_build_inp_embd(
+ 
+         inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
+     } else {
+-       lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
++        lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
+         inpL = lctx.inp_embd;
+         ggml_set_input(lctx.inp_embd);
+     }
+@@ -9193,6 +9335,22 @@ static struct ggml_tensor * llm_build_inp_embd(
+     return inpL;
+ }
+ 
++static struct ggml_tensor * llm_build_inp_cross_attn_state(
++        struct ggml_context * ctx,
++       struct llama_context & lctx,
++        const llama_hparams & hparams,
++         const llm_build_cb & cb) {
++    const int64_t n_embd = hparams.n_embd;
++
++    struct ggml_tensor * inpCAS;
++    lctx.inp_cross_attn_state = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, 1601, 4);
++    cb(lctx.inp_cross_attn_state, "inp_cross_attn_state", -1);
++    ggml_set_input(lctx.inp_cross_attn_state);
++    inpCAS = lctx.inp_cross_attn_state;
++
++    return inpCAS;
++}
++
+ static void llm_build_kv_store(
+         struct ggml_context * ctx,
+         const llama_hparams & hparams,
+@@ -10167,6 +10325,7 @@ struct llm_build_context {
+         lctx.inp_pos_bucket    = nullptr;
+         lctx.inp_embd_enc      = nullptr;
+         lctx.inp_KQ_mask_cross = nullptr;
++        lctx.inp_cross_attn_state = nullptr;
+     }
+ 
+     void free() {
+@@ -10754,6 +10913,253 @@ struct llm_build_context {
+                 LLM_NORM_RMS, cb, -1);
+         cb(cur, "result_norm", -1);
+ 
++        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
++        cb(cur, "result_output", -1);
++
++        ggml_build_forward_expand(gf, cur);
++
++        return gf;
++    }
++
++    struct ggml_cgraph * build_mllama() {
++        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
++
++        // mutable variable, needed during the last layer of the computation to skip unused tokens
++        int32_t n_tokens = this->n_tokens;
++
++        const int64_t n_embd_head = hparams.n_embd_head_v;
++        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
++        GGML_ASSERT(n_embd_head == hparams.n_rot);
++
++        struct ggml_tensor * cur;
++        struct ggml_tensor * inpL;
++        struct ggml_tensor * inpCAS;
++
++        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
++        inpCAS = llm_build_inp_cross_attn_state(ctx0, lctx, hparams, cb);
++
++        // inp_pos - contains the positions
++        struct ggml_tensor * inp_pos = build_inp_pos();
++
++        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
++        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
++
++        for (int il = 0; il < n_layer; ++il) {
++            struct ggml_tensor * inpSA = inpL;
++
++            // norm
++            cur = llm_build_norm(ctx0, inpL, hparams,
++                    model.layers[il].attn_norm, NULL,
++                    LLM_NORM_RMS, cb, il);
++            cb(cur, "attn_norm", il);
++
++            if (hparams.cross_attention_layer(il)) {
++                if (!lctx.cross_attn_state) {
++                    continue;
++                }
++
++                // cross attention layer
++                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_q_proj, cur);
++                cb(Qcur, "Qcur", il);
++
++                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
++                cb(Qcur, "Qcur", il);
++
++                Qcur = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
++                cb(Qcur, "Qcur", il);
++
++                // TODO: is this required?
++                Qcur = ggml_cont(ctx0, Qcur);
++                cb(Qcur, "Qcur", il);
++
++                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, cb, il);
++                cb(Qcur, "Qcur", il);
++
++                struct ggml_tensor * Kcur;
++                if (lctx.cross_attn_state_first_pass) {
++                    Kcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_k_proj, inpCAS);
++                    cb(Kcur, "Kcur", il);
++
++                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 6404);
++                    cb(Kcur, "Kcur", il);
++
++                    Kcur = ggml_permute(ctx0, Kcur, 0, 2, 1, 3);
++                    cb(Kcur, "Kcur", il);
++
++                    // TODO: is this required?
++                    Kcur = ggml_cont(ctx0, Kcur);
++                    cb(Kcur, "Kcur", il);
++
++                    Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, cb, il);
++                    cb(Kcur, "Kcur", il);
++
++                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self.k_l[il]));
++                } else {
++                    Kcur = ggml_view_tensor(ctx0, kv_self.k_l[il]);
++                    cb(Kcur, "Kcur (view)", il);
++                }
++
++                struct ggml_tensor * Vcur;
++                if (lctx.cross_attn_state_first_pass) {
++                    Vcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_v_proj, inpCAS);
++                    cb(Vcur, "Vcur", il);
++
++                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, 6404);
++                    cb(Vcur, "Vcur", il);
++
++                    Vcur = ggml_permute(ctx0, Vcur, 0, 2, 1, 3);
++                    cb(Vcur, "Vcur", il);
++
++                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, kv_self.v_l[il]));
++                } else {
++                    Vcur = ggml_view_tensor(ctx0, kv_self.v_l[il]);
++                    cb(Vcur, "Vcur (view)", il);
++                }
++
++                struct ggml_tensor * kq = ggml_mul_mat(ctx0, Kcur, Qcur);
++                cb(kq, "kq", il);
++
++                kq = ggml_scale_inplace(ctx0, kq, 1.0f/sqrtf(float(n_embd_head)));
++                cb(kq, "kq_scaled", il);
++
++                // TODO: apply causal masks
++                struct ggml_tensor * kq_soft_max = ggml_soft_max_inplace(ctx0, kq);
++                cb(kq_soft_max, "kq_soft_max", il);
++
++                Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, Vcur));
++                cb(Vcur, "Vcur", il);
++
++                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, Vcur, kq_soft_max);
++                cb(kqv, "kqv", il);
++
++                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
++                cb(kqv_merged, "kqv_merged", il);
++
++                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
++                cb(cur, "kqv_merged_cont", il);
++
++                cur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_o_proj, cur);
++                cb(cur, "cur", il);
++
++                // TODO: do this in place once?
++                cur = ggml_mul(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_attn_gate));
++
++                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
++                cb(ffn_inp, "ffn_inp", il);
++
++                // feed-forward network
++                cur = llm_build_norm(ctx0, ffn_inp, hparams,
++                        model.layers[il].ffn_norm, NULL,
++                        LLM_NORM_RMS, cb, il);
++                cb(cur, "ffn_norm", il);
++
++                cur = llm_build_ffn(ctx0, lctx, cur,
++                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
++                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
++                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
++                        NULL,
++                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
++                cb(cur, "ffn_out", il);
++
++                // TODO: do this inplace once?
++                cur = ggml_add_inplace(ctx0, ggml_mul_inplace(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_mlp_gate)), ffn_inp);
++                cb(cur, "ffn_out", il);
++
++                cur = lctx.cvec.apply_to(ctx0, cur, il);
++                cb(cur, "l_out", il);
++
++                // input for next layer
++                inpL = cur;
++            } else {
++                // self attention layer
++
++                // rope freq factors for llama3; may return nullptr for llama2 and other models
++                struct ggml_tensor * rope_factors = build_rope_factors(il);
++
++                // compute Q and K and RoPE them
++                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
++                cb(Qcur, "Qcur", il);
++                if (model.layers[il].bq) {
++                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
++                    cb(Qcur, "Qcur", il);
++                }
++
++                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
++                cb(Kcur, "Kcur", il);
++                if (model.layers[il].bk) {
++                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
++                    cb(Kcur, "Kcur", il);
++                }
++
++                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
++                cb(Vcur, "Vcur", il);
++                if (model.layers[il].bv) {
++                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
++                    cb(Vcur, "Vcur", il);
++                }
++
++                Qcur = ggml_rope_ext(
++                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
++                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
++                    ext_factor, attn_factor, beta_fast, beta_slow
++                );
++                cb(Qcur, "Qcur", il);
++
++                Kcur = ggml_rope_ext(
++                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
++                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
++                    ext_factor, attn_factor, beta_fast, beta_slow
++                );
++                cb(Kcur, "Kcur", il);
++
++                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
++                        model.layers[il].wo, model.layers[il].bo,
++                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
++
++
++                if (il == n_layer - 1) {
++                    // skip computing output for unused tokens
++                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
++                    n_tokens = n_outputs;
++                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
++                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
++                }
++
++                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
++                cb(ffn_inp, "ffn_inp", il);
++
++                // feed-forward network
++                cur = llm_build_norm(ctx0, ffn_inp, hparams,
++                        model.layers[il].ffn_norm, NULL,
++                        LLM_NORM_RMS, cb, il);
++                cb(cur, "ffn_norm", il);
++
++                cur = llm_build_ffn(ctx0, lctx, cur,
++                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
++                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
++                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
++                        NULL,
++                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
++                cb(cur, "ffn_out", il);
++
++                cur = ggml_add(ctx0, cur, ffn_inp);
++                cb(cur, "ffn_out", il);
++
++                cur = lctx.cvec.apply_to(ctx0, cur, il);
++                cb(cur, "l_out", il);
++
++                // input for next layer
++                inpL = cur;
++            }
++        }
++
++        cur = inpL;
++
++        cur = llm_build_norm(ctx0, cur, hparams,
++                model.output_norm, NULL,
++                LLM_NORM_RMS, cb, -1);
++        cb(cur, "result_norm", -1);
++
+         // lm_head
+         cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+         cb(cur, "result_output", -1);
+@@ -16501,6 +16907,10 @@ static struct ggml_cgraph * llama_build_graph(
+             {
+                 result = llm.build_llama();
+             } break;
++        case LLM_ARCH_MLLAMA:
++            {
++                result = llm.build_mllama();
++            } break;
+         case LLM_ARCH_BAICHUAN:
+             {
+                 result = llm.build_baichuan();
+@@ -16773,6 +17183,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
+         ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
+     }
+ 
++    // TODO (jmorganca): this might copy a lot of data on every request of a
++    // single generation even though it doesn't change, so we should
++    // find a way to not set this more than one time per image
++    if (lctx.inp_cross_attn_state &&
++        lctx.inp_cross_attn_state->buffer) {
++        ggml_backend_tensor_set(lctx.inp_cross_attn_state, lctx.cross_attn_state, 0, hparams.n_embd * 1601 * 4 * ggml_element_size(lctx.inp_cross_attn_state));
++    }
++
+     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
+         GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs");
+         const int64_t n_tokens = batch.n_tokens;
+@@ -17455,6 +17873,10 @@ static int llama_decode_internal(
+ 
+         llama_set_inputs(lctx, ubatch);
+ 
++        // TODO: replace with something better to find out if its
++        // our first actual pass
++        lctx.cross_attn_state_first_pass = false;
++
+         llama_graph_compute(lctx, gf, n_threads, threadpool);
+ 
+         // update the kv ring buffer
+@@ -18648,7 +19070,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
+         if (llama_model_has_encoder(&model)) {
+             n_attn_layer *= 3;
+         }
+-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
++        if (qs.n_attention_wv != n_attn_layer) {
++            LLAMA_LOG_WARN("%s: n_attention_wv is unexpected, expected: %d, found: %d\n", __func__, n_attn_layer, qs.n_attention_wv);
++        }
+     }
+ 
+     size_t total_size_org = 0;
+@@ -19744,6 +20168,11 @@ struct llama_context * llama_new_context_with_model(
+     return ctx;
+ }
+ 
++void llama_set_cross_attn_state(struct llama_context * ctx, float * cross_attn_state) {
++    ctx->cross_attn_state_first_pass = true;
++    ctx->cross_attn_state = cross_attn_state;
++}
++
+ void llama_free(struct llama_context * ctx) {
+     delete ctx;
+ }
+@@ -19814,6 +20243,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
+ 
+         // use what we call a normal RoPE, operating on pairs of consecutive head values
+         case LLM_ARCH_LLAMA:
++        case LLM_ARCH_MLLAMA:
+         case LLM_ARCH_BAICHUAN:
+         case LLM_ARCH_STARCODER:
+         case LLM_ARCH_PLAMO:

llama/patches/0011-add-unpad-operator.patch

@@ -0,0 +1,409 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Thu, 17 Oct 2024 17:19:25 -0700
+Subject: [PATCH] add unpad operator
+
+---
+ ggml/include/ggml.h        | 10 ++++
+ ggml/src/ggml-cuda.cu      |  4 ++
+ ggml/src/ggml-cuda/pad.cu  | 46 +++++++++++++++++++
+ ggml/src/ggml-cuda/pad.cuh |  1 +
+ ggml/src/ggml-metal.m      | 33 ++++++++++++++
+ ggml/src/ggml-metal.metal  | 45 ++++++++++++++++++
+ ggml/src/ggml.c            | 93 +++++++++++++++++++++++++++++++++++++-
+ 7 files changed, 230 insertions(+), 2 deletions(-)
+
+diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
+index ce3d92cb..962cb5f7 100644
+--- a/ggml/include/ggml.h
++++ b/ggml/include/ggml.h
+@@ -506,6 +506,7 @@ extern "C" {
+         GGML_OP_POOL_2D_BACK,
+         GGML_OP_UPSCALE, // nearest interpolate
+         GGML_OP_PAD,
++        GGML_OP_UNPAD,
+         GGML_OP_ARANGE,
+         GGML_OP_TIMESTEP_EMBEDDING,
+         GGML_OP_ARGSORT,
+@@ -1764,6 +1765,15 @@ extern "C" {
+             int                  p2,
+             int                  p3);
+ 
++    // unpad each dimension: [x, ..., x, y, ..., y] -> [x, ..., x]
++    GGML_API struct ggml_tensor * ggml_unpad(
++            struct ggml_context * ctx,
++            struct ggml_tensor  * a,
++            int                  p0,
++            int                  p1,
++            int                  p2,
++            int                  p3);
++
+     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
+     // timesteps: [N,]
+     // return: [N, dim]
+diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
+index fe77b81c..6e84af56 100644
+--- a/ggml/src/ggml-cuda.cu
++++ b/ggml/src/ggml-cuda.cu
+@@ -2270,6 +2270,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
+         case GGML_OP_PAD:
+             ggml_cuda_op_pad(ctx, dst);
+             break;
++        case GGML_OP_UNPAD:
++            ggml_cuda_op_unpad(ctx, dst);
++            break;
+         case GGML_OP_ARANGE:
+             ggml_cuda_op_arange(ctx, dst);
+             break;
+@@ -2992,6 +2995,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
+         case GGML_OP_GROUP_NORM:
+         case GGML_OP_UPSCALE:
+         case GGML_OP_PAD:
++        case GGML_OP_UNPAD:
+         case GGML_OP_ARANGE:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_LEAKY_RELU:
+diff --git a/ggml/src/ggml-cuda/pad.cu b/ggml/src/ggml-cuda/pad.cu
+index aba539e8..39fd4b16 100644
+--- a/ggml/src/ggml-cuda/pad.cu
++++ b/ggml/src/ggml-cuda/pad.cu
+@@ -47,3 +47,49 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+ }
++
++static __global__ void unpad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
++    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
++    // blockIdx.y: idx of ne1
++    // blockIDx.x: idx of ne0 / BLOCK_SIZE
++    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
++    if (nidx >= ne0) {
++        return;
++    }
++
++    // operation
++    int offset_dst =
++        nidx +
++        blockIdx.y * ne0 +
++        blockIdx.z * ne0 * gridDim.y;
++    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
++        int offset_src =
++            nidx +
++            blockIdx.y * ne00 +
++            blockIdx.z * ne00 * ne01;
++        dst[offset_dst] = x[offset_src];
++    }
++}
++
++static void unpad_f32_cuda(const float * x, float * dst,
++    const int ne00, const int ne01, const int ne02, const int ne03,
++    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
++    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
++    dim3 gridDim(num_blocks, ne1, ne2*ne3);
++    unpad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
++}
++
++void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
++    const ggml_tensor * src0 = dst->src[0];
++    const float * src0_d = (const float *)src0->data;
++    float * dst_d = (float *)dst->data;
++    cudaStream_t stream = ctx.stream();
++
++    GGML_ASSERT(src0->type == GGML_TYPE_F32);
++    GGML_ASSERT(dst->type == GGML_TYPE_F32);
++    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
++
++    unpad_f32_cuda(src0_d, dst_d,
++        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
++        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
++}
+diff --git a/ggml/src/ggml-cuda/pad.cuh b/ggml/src/ggml-cuda/pad.cuh
+index 8fd386b0..e2ededc3 100644
+--- a/ggml/src/ggml-cuda/pad.cuh
++++ b/ggml/src/ggml-cuda/pad.cuh
+@@ -3,3 +3,4 @@
+ #define CUDA_PAD_BLOCK_SIZE 256
+ 
+ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
++void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
+index 829c5e39..25702d85 100644
+--- a/ggml/src/ggml-metal.m
++++ b/ggml/src/ggml-metal.m
+@@ -193,6 +193,7 @@
+     GGML_METAL_KERNEL_TYPE_IM2COL_F32,
+     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
+     GGML_METAL_KERNEL_TYPE_PAD_F32,
++    GGML_METAL_KERNEL_TYPE_UNPAD_F32,
+     GGML_METAL_KERNEL_TYPE_ARANGE_F32,
+     GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
+     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
+@@ -689,6 +690,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                       pad_f32,                        true);
++        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UNPAD_F32,                     unpad_f32,                        true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,        timestep_embedding_f32,         true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                    arange_f32,                     true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,           argsort_f32_i32_asc,            true);
+@@ -846,6 +848,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
+             return false;
+         case GGML_OP_UPSCALE:
+         case GGML_OP_PAD:
++        case GGML_OP_UNPAD:
+         case GGML_OP_ARANGE:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_ARGSORT:
+@@ -2655,6 +2658,36 @@ static void ggml_metal_encode_node(
+ 
+                 const int nth = MIN(1024, ne0);
+ 
++                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
++            } break;
++        case GGML_OP_UNPAD:
++            {
++                GGML_ASSERT(src0->type == GGML_TYPE_F32);
++
++                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UNPAD_F32].pipeline;
++
++                [encoder setComputePipelineState:pipeline];
++                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
++                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
++                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
++                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
++                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
++                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
++                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
++                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
++                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
++                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
++                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
++                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
++                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
++                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
++                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
++                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
++                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
++                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
++
++                const int nth = MIN(1024, ne0);
++
+                 [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+             } break;
+         case GGML_OP_ARANGE:
+diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal
+index 2b200032..09887511 100644
+--- a/ggml/src/ggml-metal.metal
++++ b/ggml/src/ggml-metal.metal
+@@ -2029,6 +2029,51 @@ kernel void kernel_pad_f32(
+     }
+ }
+ 
++kernel void kernel_unpad_f32(
++    device  const char * src0,
++    device        char * dst,
++    constant   int64_t & ne00,
++    constant   int64_t & ne01,
++    constant   int64_t & ne02,
++    constant   int64_t & ne03,
++    constant  uint64_t & nb00,
++    constant  uint64_t & nb01,
++    constant  uint64_t & nb02,
++    constant  uint64_t & nb03,
++    constant   int64_t & ne0,
++    constant   int64_t & ne1,
++    constant   int64_t & ne2,
++    constant   int64_t & ne3,
++    constant  uint64_t & nb0,
++    constant  uint64_t & nb1,
++    constant  uint64_t & nb2,
++    constant  uint64_t & nb3,
++    uint3 tgpig[[threadgroup_position_in_grid]],
++    uint3 tpitg[[thread_position_in_threadgroup]],
++    uint3   ntg[[threads_per_threadgroup]]) {
++
++    const int64_t i3 = tgpig.z;
++    const int64_t i2 = tgpig.y;
++    const int64_t i1 = tgpig.x;
++
++    const int64_t i03 = i3;
++    const int64_t i02 = i2;
++    const int64_t i01 = i1;
++
++    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
++    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
++
++    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
++        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
++            if (i0 < ne00) {
++                dst_ptr[i0] = src0_ptr[i0];
++            }
++        }
++
++        return;
++    }
++}
++
+ kernel void kernel_arange_f32(
+     device        char * dst,
+     constant   int64_t & ne0,
+diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
+index bcbc32d9..f4864ac8 100644
+--- a/ggml/src/ggml.c
++++ b/ggml/src/ggml.c
+@@ -2997,6 +2997,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+     "POOL_2D_BACK",
+     "UPSCALE",
+     "PAD",
++    "UNPAD",
+     "ARANGE",
+     "TIMESTEP_EMBEDDING",
+     "ARGSORT",
+@@ -3030,7 +3031,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+     "OPT_STEP_ADAMW",
+ };
+ 
+-static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80");
++static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+ 
+ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "none",
+@@ -3091,6 +3092,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "pool_2d_back(x)",
+     "upscale(x)",
+     "pad(x)",
++    "unpad(x)",
+     "arange(start, stop, step)",
+     "timestep_embedding(timesteps, dim, max_period)",
+     "argsort(x)",
+@@ -3124,7 +3126,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "adamw(x)",
+ };
+ 
+-static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80");
++static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+ 
+ static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
+ 
+@@ -6955,6 +6957,32 @@ struct ggml_tensor * ggml_pad(
+     return result;
+ }
+ 
++// ggml_unpad
++
++struct ggml_tensor * ggml_unpad(
++    struct ggml_context * ctx,
++    struct ggml_tensor  * a,
++    int p0, int p1, int p2, int p3) {
++    bool is_node = false;
++
++    if (a->grad) {
++        GGML_ABORT("fatal error"); // TODO: implement backward
++        is_node = true;
++    }
++
++    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
++            a->ne[0] - p0,
++            a->ne[1] - p1,
++            a->ne[2] - p2,
++            a->ne[3] - p3);
++
++    result->op = GGML_OP_UNPAD;
++    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
++    result->src[0] = a;
++
++    return result;
++}
++
+ // ggml_arange
+ 
+ struct ggml_tensor * ggml_arange(
+@@ -15312,6 +15340,58 @@ static void ggml_compute_forward_pad(
+     }
+ }
+ 
++static void ggml_compute_forward_unpad_f32(
++    const struct ggml_compute_params *params,
++    struct ggml_tensor *dst) {
++
++    const struct ggml_tensor * src0 = dst->src[0];
++
++    GGML_ASSERT(src0->nb[0] == sizeof(float));
++    GGML_ASSERT( dst->nb[0] == sizeof(float));
++
++    const int ith = params->ith;
++    const int nth = params->nth;
++
++    GGML_TENSOR_UNARY_OP_LOCALS
++
++    float * dst_ptr = (float *) dst->data;
++
++    // TODO: optimize
++
++    for (int64_t i2 = 0; i2 < ne2; ++i2) {
++        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
++            for (int64_t i0 = 0; i0 < ne0; ++i0) {
++                for (int64_t i3 = 0; i3 < ne3; ++i3) {
++                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
++
++                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
++
++                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
++                        dst_ptr[dst_idx] = *src_ptr;
++                    }
++                }
++            }
++        }
++    }
++}
++
++static void ggml_compute_forward_unpad(
++    const struct ggml_compute_params * params,
++    struct ggml_tensor * dst) {
++
++    const struct ggml_tensor * src0 = dst->src[0];
++
++    switch (src0->type) {
++        case GGML_TYPE_F32:
++            {
++                ggml_compute_forward_unpad_f32(params, dst);
++            } break;
++        default:
++            {
++                GGML_ABORT("fatal error");
++            }
++    }
++}
+ 
+ // ggml_compute_forward_arange
+ 
+@@ -17294,6 +17374,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
+             {
+                 ggml_compute_forward_pad(params, tensor);
+             } break;
++        case GGML_OP_UNPAD:
++            {
++                ggml_compute_forward_unpad(params, tensor);
++            } break;
+         case GGML_OP_ARANGE:
+             {
+                 ggml_compute_forward_arange(params, tensor);
+@@ -18369,6 +18453,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
+             {
+                 GGML_ABORT("fatal error"); // TODO: not implemented
+             }
++        case GGML_OP_UNPAD:
++            {
++                GGML_ABORT("fatal error"); // TODO: not implemented
++            }
+         case GGML_OP_ARANGE:
+             {
+                 GGML_ABORT("fatal error"); // TODO: not implemented
+@@ -19165,6 +19253,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
+             } break;
+         case GGML_OP_UPSCALE:
+         case GGML_OP_PAD:
++        case GGML_OP_UNPAD:
+         case GGML_OP_ARANGE:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_ARGSORT:

llama/patches/02-pretokenizer.diff

@@ -1,32 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 4c0a1bb6..800dfb95 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -6287,16 +6287,7 @@ static void llm_load_vocab(
-         if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
-             vocab.tokenizer_add_space_prefix = false;
-             vocab.tokenizer_clean_spaces = true;
--            if (tokenizer_pre.empty()) {
--                LLAMA_LOG_WARN("%s: missing pre-tokenizer type, using: 'default'\n", __func__);
--                LLAMA_LOG_WARN("%s:                                             \n", __func__);
--                LLAMA_LOG_WARN("%s: ************************************        \n", __func__);
--                LLAMA_LOG_WARN("%s: GENERATION QUALITY WILL BE DEGRADED!        \n", __func__);
--                LLAMA_LOG_WARN("%s: CONSIDER REGENERATING THE MODEL             \n", __func__);
--                LLAMA_LOG_WARN("%s: ************************************        \n", __func__);
--                LLAMA_LOG_WARN("%s:                                             \n", __func__);
--                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
--            } else if (tokenizer_pre == "default") {
-+            if (tokenizer_pre == "default") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
-             } else if (
-                     tokenizer_pre == "llama3"   ||
-@@ -6398,7 +6389,8 @@ static void llm_load_vocab(
-                 vocab.tokenizer_add_bos = true;
-                 vocab.tokenizer_clean_spaces = false;
-             } else {
--                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
-+                LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
-+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
-             }
-         } else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
-             vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;

llama/patches/03-metal.diff

@@ -1,45 +0,0 @@
-diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
-index 9da08fe2..3a433703 100644
---- a/ggml/src/ggml-metal.m
-+++ b/ggml/src/ggml-metal.m
-@@ -1720,27 +1720,23 @@ static void ggml_metal_encode_node(
-                 // to the matrix-vector kernel
-                 int ne11_mm_min = 1;
- 
--#if 0
-                 // the numbers below are measured on M2 Ultra for 7B and 13B models
-                 // these numbers do not translate to other devices or model sizes
-                 // TODO: need to find a better approach
--                        if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
--                            switch (src0t) {
--                                case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
--                                case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
--                                case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
--                                case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
--                                case GGML_TYPE_Q4_0:
--                                case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
--                                case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
--                                case GGML_TYPE_Q5_0:                          // not tested yet
--                                case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
--                                case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
--                                case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
--                                default:             ne11_mm_min = 1;  break;
--                            }
-+                        switch (src0t) {
-+                            case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
-+                            case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
-+                            case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
-+                            case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
-+                            case GGML_TYPE_Q4_0:
-+                            case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
-+                            case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
-+                            case GGML_TYPE_Q5_0:                          // not tested yet
-+                            case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
-+                            case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
-+                            case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
-+                            default:             ne11_mm_min = 1;  break;
-                         }
--#endif
- 
-                         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
-                         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel

llama/patches/05-embeddings.diff

@@ -1,43 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 4c0a1bb6..17e5bc2a 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -16928,7 +16928,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
-     const auto n_embd  = hparams.n_embd;
-
-     // TODO: use a per-batch flag for logits presence instead
--    const bool has_logits = !cparams.embeddings;
-+    const bool has_logits =  cparams.causal_attn;
-     const bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
-
-     const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
-@@ -17200,20 +17200,23 @@ static int llama_decode_internal(
-             // no output
-             res  = nullptr;
-             embd = nullptr;
--        } else if (cparams.embeddings) {
--            res  = nullptr; // do not extract logits for embedding case
--            embd = nullptr;
--            for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) {
-+        }
-+
-+        if (cparams.embeddings) {
-+            for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) {
-+                embd = ggml_graph_node(gf, i);
-                 if (strcmp(ggml_graph_node(gf, i)->name, "result_embd_pooled") == 0) {
--                    embd = ggml_graph_node(gf, i);
-                     break;
-                 }
-             }
--            GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
-         } else {
-             embd = nullptr; // do not extract embeddings when not needed
-             GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
-         }
-+
-+        if (!cparams.causal_attn) {
-+            res = nullptr; // do not extract logits when not needed
-+        }
-         // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
-
-         ggml_backend_sched_alloc_graph(lctx.sched, gf);

llama/runner/runner.go

@@ -18,6 +18,7 @@ import (
 	"strings"
 	"sync"
 	"time"
+	"unicode/utf8"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/llama"
@@ -206,6 +207,26 @@ func (s *Server) inputs(prompt string, images []ImageData) ([]input, error) {
 		}
 	}
 
+	if s.clip.cc != nil {
+		var embed [][]float32
+
+		if s.clip.cc.IsMllama && len(images) >= 1 {
+			hash := s.cache.HashImage(images[0].Data)
+
+			s.clip.mu.Lock()
+			var err error
+			embed, err = s.cache.FindImage(hash)
+			if err != nil {
+				embed = llama.NewMllamaImageEmbed(s.lc, s.clip.cc, images[0].Data, images[0].AspectRatioID)
+				s.cache.AddImage(hash, embed)
+			}
+			s.clip.mu.Unlock()
+		}
+		s.mu.Lock()
+		llama.MllamaSetCrossAttn(s.lc, s.clip.cc, embed)
+		s.mu.Unlock()
+	}
+
 	return inputs, nil
 }
 
@@ -273,17 +294,29 @@ func (s *Server) shiftContext(seq *Sequence) {
 }
 
 func flushPending(seq *Sequence) bool {
-	for _, p := range seq.pendingResponses {
-		select {
-		case seq.responses <- p:
-		case <-seq.quit:
-			seq.pendingResponses = []string{}
-			return false
-		}
+	joined := strings.Join(seq.pendingResponses, "")
+	seq.pendingResponses = []string{}
+
+	// Check if there are any partial UTF-8 characters remaining.
+	// We already check and queue as we are generating but some may
+	// still make it here:
+	// - Sequence is ending, e.g. generation limit has been hit
+	// - Invalid characters in the middle of a string
+	// This is a stricter check to ensure we never output invalid Unicode.
+	for !utf8.ValidString(joined) {
+		joined = joined[:len(joined)-1]
 	}
 
-	seq.pendingResponses = []string{}
-	return true
+	if len(joined) == 0 {
+		return true
+	}
+
+	select {
+	case seq.responses <- joined:
+		return true
+	case <-seq.quit:
+		return false
+	}
 }
 
 func (s *Server) removeSequence(seqIndex int, reason string) {
@@ -294,6 +327,9 @@ func (s *Server) removeSequence(seqIndex int, reason string) {
 	close(seq.responses)
 	close(seq.embedding)
 	seq.cache.InUse = false
+	if s.clip.cc != nil {
+		llama.MllamaSetCrossAttn(s.lc, s.clip.cc, nil)
+	}
 	s.seqs[seqIndex] = nil
 }
 
@@ -517,8 +553,9 @@ type Options struct {
 }
 
 type ImageData struct {
-	Data []byte `json:"data"`
-	ID   int    `json:"id"`
+	Data          []byte `json:"data"`
+	ID            int    `json:"id"`
+	AspectRatioID int    `json:"aspect_ratio_id"`
 }
 
 type CompletionRequest struct {
@@ -770,7 +807,11 @@ func (s *Server) loadModel(
 	}
 
 	if ppath != "" {
-		s.clip.cc = llama.NewClipContext(ppath)
+		var err error
+		s.clip.cc, err = llama.NewClipContext(ppath)
+		if err != nil {
+			panic(err)
+		}
 	}
 
 	s.cache = NewInputCache(s.lc, kvSize, s.parallel, multiUserCache)

llama/sync.sh

@@ -1,138 +0,0 @@
-#!/bin/bash
-
-set -e
-
-# Run in the llama directory
-
-# Set the source directory
-# TODO in the future: src_dir=$1
-src_dir=../llm/llama.cpp
-
-if [ -z "$src_dir" ]; then
-  echo "Usage: $0 LLAMA_CPP_DIR"
-  exit 1
-fi
-
-# Set the destination directory
-dst_dir=$(pwd)
-
-# TODO remove once we no longer use the submodule
-if [ -z "${OLLAMA_SKIP_PATCHING}" ]; then
-  (cd ../ && git submodule init && git submodule update --force ./llm/llama.cpp)
-
-  # apply patches
-  for patch in $dst_dir/patches/*.diff; do
-    echo "Applying $patch"
-    git -C $src_dir apply "$patch"
-  done
-else
-  echo "Skipping patching"
-fi
-
-# llama.cpp
-cp $src_dir/src/unicode.cpp $dst_dir/unicode.cpp
-cp $src_dir/src/unicode.h $dst_dir/unicode.h
-cp $src_dir/src/unicode-data.cpp $dst_dir/unicode-data.cpp
-cp $src_dir/src/unicode-data.h $dst_dir/unicode-data.h
-cp $src_dir/src/llama.cpp $dst_dir/llama.cpp
-cp $src_dir/src/llama-impl.h $dst_dir/llama-impl.h
-cp $src_dir/src/llama-vocab.cpp $dst_dir/llama-vocab.cpp
-cp $src_dir/src/llama-vocab.h $dst_dir/llama-vocab.h
-cp $src_dir/src/llama-grammar.cpp $dst_dir/llama-grammar.cpp
-cp $src_dir/src/llama-grammar.h $dst_dir/llama-grammar.h
-cp $src_dir/src/llama-sampling.cpp $dst_dir/llama-sampling.cpp
-cp $src_dir/src/llama-sampling.h $dst_dir/llama-sampling.h
-cp $src_dir/include/llama.h $dst_dir/llama.h
-cp $src_dir/ggml/src/llamafile/sgemm.cpp $dst_dir/sgemm.cpp
-cp $src_dir/ggml/src/llamafile/sgemm.h $dst_dir/sgemm.h
-mkdir -p $dst_dir/llamafile
-cp $src_dir/ggml/src/llamafile/sgemm.h $dst_dir/llamafile/sgemm.h
-
-# ggml
-cp $src_dir/ggml/src/ggml.c $dst_dir/ggml.c
-cp $src_dir/ggml/include/ggml.h $dst_dir/ggml.h
-cp $src_dir/ggml/src/ggml-quants.c $dst_dir/ggml-quants.c
-cp $src_dir/ggml/src/ggml-quants.h $dst_dir/ggml-quants.h
-cp $src_dir/ggml/src/ggml-metal.metal $dst_dir/ggml-metal.metal
-cp $src_dir/ggml/include/ggml-metal.h $dst_dir/ggml-metal.h
-cp $src_dir/ggml/src/ggml-metal.m $dst_dir/ggml-metal_darwin_arm64.m
-cp $src_dir/ggml/src/ggml-impl.h $dst_dir/ggml-impl.h
-cp $src_dir/ggml/include/ggml-cuda.h $dst_dir/ggml-cuda.h
-cp $src_dir/ggml/src/ggml-cuda.cu $dst_dir/ggml-cuda.cu
-cp $src_dir/ggml/src/ggml-common.h $dst_dir/ggml-common.h
-cp $src_dir/ggml/include/ggml-backend.h $dst_dir/ggml-backend.h
-cp $src_dir/ggml/src/ggml-backend.c $dst_dir/ggml-backend.c
-cp $src_dir/ggml/src/ggml-backend-impl.h $dst_dir/ggml-backend-impl.h
-cp $src_dir/ggml/include/ggml-alloc.h $dst_dir/ggml-alloc.h
-cp $src_dir/ggml/src/ggml-alloc.c $dst_dir/ggml-alloc.c
-cp $src_dir/ggml/src/ggml-aarch64.h $dst_dir/ggml-aarch64.h
-cp $src_dir/ggml/src/ggml-aarch64.c $dst_dir/ggml-aarch64.c
-cp $src_dir/ggml/src/ggml-cpu-impl.h $dst_dir/ggml-cpu-impl.h
-cp $src_dir/ggml/include/ggml-blas.h $dst_dir/ggml-blas.h
-cp $src_dir/ggml/src/ggml-blas.cpp $dst_dir/ggml-blas.cpp
-
-# ggml-cuda
-mkdir -p $dst_dir/ggml-cuda/template-instances
-mkdir -p $dst_dir/ggml-cuda/vendors
-cp $src_dir/ggml/src/ggml-cuda/*.cu $dst_dir/ggml-cuda/
-cp $src_dir/ggml/src/ggml-cuda/*.cuh $dst_dir/ggml-cuda/
-cp $src_dir/ggml/src/ggml-cuda/template-instances/*.cu $dst_dir/ggml-cuda/template-instances/
-cp $src_dir/ggml/src/ggml-cuda/vendors/*.h $dst_dir/ggml-cuda/vendors/
-
-# llava
-cp $src_dir/examples/llava/clip.cpp $dst_dir/clip.cpp
-cp $src_dir/examples/llava/clip.h $dst_dir/clip.h
-cp $src_dir/examples/llava/llava.cpp $dst_dir/llava.cpp
-cp $src_dir/examples/llava/llava.h $dst_dir/llava.h
-cp $src_dir/common/log.h $dst_dir/log.h
-cp $src_dir/common/log.cpp $dst_dir/log.cpp
-cp $src_dir/common/stb_image.h $dst_dir/stb_image.h
-
-# These files are mostly used by the llava code
-# and shouldn't be necessary once we use clip.cpp directly
-cp $src_dir/common/common.cpp $dst_dir/common.cpp
-cp $src_dir/common/common.h $dst_dir/common.h
-cp $src_dir/common/sampling.cpp $dst_dir/sampling.cpp
-cp $src_dir/common/sampling.h $dst_dir/sampling.h
-cp $src_dir/common/json.hpp $dst_dir/json.hpp
-cp $src_dir/common/json-schema-to-grammar.cpp $dst_dir/json-schema-to-grammar.cpp
-cp $src_dir/common/json-schema-to-grammar.h $dst_dir/json-schema-to-grammar.h
-cp $src_dir/common/base64.hpp $dst_dir/base64.hpp
-cat <<EOF > $dst_dir/build-info.cpp
-int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "$sha1";
-char const *LLAMA_COMPILER = "";
-char const *LLAMA_BUILD_TARGET = "";
-EOF
-
-# add licenses
-sha1=$(git -C $src_dir rev-parse @)
-
-TEMP_LICENSE=$(mktemp)
-cleanup() {
-    rm -f $TEMP_LICENSE
-}
-trap cleanup 0
-
-cat <<EOF | sed 's/ *$//' >$TEMP_LICENSE
-/**
- * llama.cpp - commit $sha1 - do not edit this file
- *
-$(sed 's/^/ * /' <$src_dir/LICENSE)
- */
-
-EOF
-
-LICENSE_FILES=$(find $dst_dir -type f \( -name "*.c" -o -name "*.h" -o -name "*.cpp" -o -name "*.m" -o -name "*.metal" -o -name "*.cu" -o -name "*.cuh" \))
-EXCLUDED_FILES=("sgemm.cpp" "sgemm.h" "sampling_ext.cpp" "sampling_ext.h" "stb_image.h" "json.hpp" "llama_darwin.c")
-
-for IN in $LICENSE_FILES; do
-    for EXCLUDED in "${EXCLUDED_FILES[@]}"; do
-        if [[ "$IN" == *"$EXCLUDED" ]]; then
-            continue 2
-        fi
-    done
-    TMP=$(mktemp)
-    cat $TEMP_LICENSE $IN >$TMP
-    mv $TMP $IN
-done

llama/vendoring.env

@@ -0,0 +1 @@
+LLAMACPP_BASE_COMMIT=3f1ae2e32cde00c39b96be6d01c2997c29bae555

llm/ext_server/CMakeLists.txt

@@ -1,15 +0,0 @@
-set(TARGET ollama_llama_server)
-option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-set(LLAMA_SERVER_LDFLAGS $ENV{LLAMA_SERVER_LDFLAGS})
-include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp httplib.h)
-install(TARGETS ${TARGET} RUNTIME)
-target_compile_definitions(${TARGET} PRIVATE
-    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
-)
-target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_SERVER_LDFLAGS})
-if (WIN32)
-    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
-    target_link_options(${TARGET} PRIVATE -municode -Wl,/subsystem:console)
-endif()
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

llm/ext_server/utils.hpp

@@ -1,661 +0,0 @@
-// MIT License
-
-// Copyright (c) 2023 Georgi Gerganov
-
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-
-// The above copyright notice and this permission notice shall be included in all
-// copies or substantial portions of the Software.
-
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-// SOFTWARE.
-
-#pragma once
-
-#include <string>
-#include <vector>
-#include <set>
-#include <mutex>
-#include <condition_variable>
-#include <unordered_map>
-#include <random>
-#include <iostream>
-#include <thread>
-
-#include "json.hpp"
-
-#include "../llava/clip.h"
-
-using json = nlohmann::json;
-
-extern bool server_verbose;
-extern bool server_log_json;
-
-#ifndef SERVER_VERBOSE
-#define SERVER_VERBOSE 1
-#endif
-
-#if SERVER_VERBOSE != 1
-#define LOG_VERBOSE(MSG, ...)
-#else
-#define LOG_VERBOSE(MSG, ...)                                            \
-    do                                                                   \
-    {                                                                    \
-        if (server_verbose)                                              \
-        {                                                                \
-            server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
-        }                                                                \
-    } while (0)
-#endif
-
-#define LOG_ERROR(  MSG, ...) server_log("ERROR",  __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_INFO(   MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_DEBUG(  MSG, ...) server_log("DEBUG", __func__, __LINE__, MSG, __VA_ARGS__)
-
-enum server_state {
-    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
-    SERVER_STATE_READY,          // Server is ready and model is loaded
-    SERVER_STATE_ERROR           // An error occurred, load_model failed
-};
-
-enum task_type {
-    TASK_TYPE_COMPLETION,
-    TASK_TYPE_CANCEL,
-    TASK_TYPE_NEXT_RESPONSE,
-    TASK_TYPE_METRICS
-};
-
-struct task_server {
-    int id = -1; // to be filled by llama_server_queue
-    int target_id;
-    task_type type;
-    json data;
-    bool infill_mode = false;
-    bool embedding_mode = false;
-    int multitask_id = -1;
-};
-
-struct task_result {
-    int id;
-    int multitask_id = -1;
-    bool stop;
-    bool error;
-    json result_json;
-};
-
-struct task_multi {
-    int id;
-    std::set<int> subtasks_remaining{};
-    std::vector<task_result> results{};
-};
-
-// completion token output with probabilities
-struct completion_token_output {
-    struct token_prob
-    {
-        llama_token tok;
-        float prob;
-    };
-
-    std::vector<token_prob> probs;
-    llama_token tok;
-    std::string text_to_send;
-};
-
-struct token_translator {
-    llama_context * ctx;
-    std::string operator()(llama_token tok)                    const { return llama_token_to_piece(ctx, tok); }
-    std::string operator()(const completion_token_output &cto) const { return (*this)(cto.tok); }
-};
-
-static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra) {
-    std::stringstream ss_tid;
-    ss_tid << std::this_thread::get_id();
-    json log = nlohmann::ordered_json{
-        {"tid", ss_tid.str()},
-        {"timestamp", time(nullptr)},
-    };
-
-    if (strncmp("DEBUG", level, strlen(level)) == 0 && !server_verbose) {
-        return;
-    }
-
-    if (server_log_json) {
-        log.merge_patch(
-                {
-                        {"level",     level},
-                        {"function",  function},
-                        {"line",      line},
-                        {"msg",       message},
-                });
-        if (!extra.empty()) {
-            log.merge_patch(extra);
-        }
-
-        std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush;
-    } else {
-        if (!extra.empty()) {
-            log.merge_patch(extra);
-        }
-
-        std::stringstream ss;
-        ss << level << " [" << function << "] " << message << " |";
-        for (const auto& el : log.items())
-        {
-            const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
-            ss << " " << el.key() << "=" << value;
-        }
-
-        const std::string str = ss.str();
-        printf("%.*s\n", (int)str.size(), str.data());
-        fflush(stdout);
-    }
-}
-
-//
-// server utils
-//
-
-template <typename T>
-static T json_value(const json &body, const std::string &key, const T &default_value) {
-    // Fallback null to default value
-    return body.contains(key) && !body.at(key).is_null()
-        ? body.value(key, default_value)
-        : default_value;
-}
-
-// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-inline bool verify_custom_template(const std::string & tmpl) {
-    llama_chat_message chat[] = {{"user", "test"}};
-    std::vector<char> buf(1);
-    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, buf.data(), buf.size());
-    return res >= 0;
-}
-
-// Format given chat. If tmpl is empty, we take the template from model metadata
-inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
-    size_t alloc_size = 0;
-    // vector holding all allocated string to be passed to llama_chat_apply_template
-    std::vector<std::string> str(messages.size() * 2);
-    std::vector<llama_chat_message> chat(messages.size());
-
-    for (size_t i = 0; i < messages.size(); ++i) {
-        auto &curr_msg = messages[i];
-        str[i*2 + 0]    = json_value(curr_msg, "role",    std::string(""));
-        str[i*2 + 1]    = json_value(curr_msg, "content", std::string(""));
-        alloc_size     += str[i*2 + 1].length();
-        chat[i].role    = str[i*2 + 0].c_str();
-        chat[i].content = str[i*2 + 1].c_str();
-    }
-
-    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
-    std::vector<char> buf(alloc_size * 2);
-
-    // run the first time to get the total output length
-    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
-
-    // if it turns out that our buffer is too small, we resize it
-    if ((size_t) res > buf.size()) {
-        buf.resize(res);
-        res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
-    }
-
-    std::string formatted_chat(buf.data(), res);
-    LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
-
-    return formatted_chat;
-}
-
-//
-// work queue utils
-//
-
-struct llama_server_queue {
-    int id = 0;
-    std::mutex mutex_tasks;
-    bool running;
-    // queues
-    std::vector<task_server> queue_tasks;
-    std::vector<task_server> queue_tasks_deferred;
-    std::vector<task_multi> queue_multitasks;
-    std::condition_variable condition_tasks;
-    // callback functions
-    std::function<void(task_server&)> callback_new_task;
-    std::function<void(task_multi&)> callback_finish_multitask;
-    std::function<void(void)> callback_run_slots;
-
-    // Add a new task to the end of the queue
-    int post(task_server task) {
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        if (task.id == -1) {
-            task.id = id++;
-            LOG_VERBOSE("new task id", {{"new_id", task.id}});
-        }
-        queue_tasks.push_back(std::move(task));
-        condition_tasks.notify_one();
-        return task.id;
-    }
-
-    // Add a new task, but defer until one slot is available
-    void defer(task_server task) {
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        queue_tasks_deferred.push_back(std::move(task));
-    }
-
-    // Get the next id for creating anew task
-    int get_new_id() {
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        int new_id = id++;
-        LOG_VERBOSE("new task id", {{"new_id", new_id}});
-        return new_id;
-    }
-
-    // Register function to process a new task
-    void on_new_task(std::function<void(task_server&)> callback) {
-        callback_new_task = callback;
-    }
-
-    // Register function to process a multitask when it is finished
-    void on_finish_multitask(std::function<void(task_multi&)> callback) {
-        callback_finish_multitask = callback;
-    }
-
-    // Register the function to be called when all slots data is ready to be processed
-    void on_run_slots(std::function<void(void)> callback) {
-        callback_run_slots = callback;
-    }
-
-    // Call when the state of one slot is changed
-    void notify_slot_changed() {
-        // move deferred tasks back to main loop
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        for (auto & task : queue_tasks_deferred) {
-            queue_tasks.push_back(std::move(task));
-        }
-        queue_tasks_deferred.clear();
-    }
-
-    // end the start_loop routine
-    void terminate() {
-        {
-            std::unique_lock<std::mutex> lock(mutex_tasks);
-            running = false;
-        }
-        condition_tasks.notify_all();
-    }
-
-    /**
-     * Main loop consists of these steps:
-     * - Wait until a new task arrives
-     * - Process the task (i.e. maybe copy data into slot)
-     * - Check if multitask is finished
-     * - Run all slots
-     */
-    void start_loop() {
-        running = true;
-        while (true) {
-            LOG_VERBOSE("new task may arrive", {});
-            {
-                while (true)
-                {
-                    std::unique_lock<std::mutex> lock(mutex_tasks);
-                    if (queue_tasks.empty()) {
-                        lock.unlock();
-                        break;
-                    }
-                    task_server task = queue_tasks.front();
-                    queue_tasks.erase(queue_tasks.begin());
-                    lock.unlock();
-                    LOG_VERBOSE("callback_new_task", {{"task_id", task.id}});
-                    callback_new_task(task);
-                }
-                LOG_VERBOSE("update_multitasks", {});
-                // check if we have any finished multitasks
-                auto queue_iterator = queue_multitasks.begin();
-                while (queue_iterator != queue_multitasks.end())
-                {
-                    if (queue_iterator->subtasks_remaining.empty())
-                    {
-                        // all subtasks done == multitask is done
-                        task_multi current_multitask = *queue_iterator;
-                        callback_finish_multitask(current_multitask);
-                        // remove this multitask
-                        queue_iterator = queue_multitasks.erase(queue_iterator);
-                    }
-                    else
-                    {
-                        ++queue_iterator;
-                    }
-                }
-                // all tasks in the current loop is processed, slots data is now ready
-                LOG_VERBOSE("callback_run_slots", {});
-                callback_run_slots();
-            }
-            LOG_VERBOSE("wait for new task", {});
-            // wait for new task
-            {
-                std::unique_lock<std::mutex> lock(mutex_tasks);
-                if (queue_tasks.empty()) {
-                    if (!running) {
-                        LOG_VERBOSE("ending start_loop", {});
-                        return;
-                    }
-                    condition_tasks.wait(lock, [&]{
-                        return (!queue_tasks.empty() || !running);
-                    });
-                }
-            }
-        }
-    }
-
-    //
-    // functions to manage multitasks
-    //
-
-    // add a multitask by specifying the id of all subtask (subtask is a task_server)
-    void add_multitask(int multitask_id, std::vector<int>& sub_ids)
-    {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
-        task_multi multi;
-        multi.id = multitask_id;
-        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
-        queue_multitasks.push_back(multi);
-    }
-
-    // updatethe remaining subtasks, while appending results to multitask
-    void update_multitask(int multitask_id, int subtask_id, task_result& result)
-    {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
-        for (auto& multitask : queue_multitasks)
-        {
-            if (multitask.id == multitask_id)
-            {
-                multitask.subtasks_remaining.erase(subtask_id);
-                multitask.results.push_back(result);
-            }
-        }
-    }
-};
-
-struct llama_server_response {
-    typedef std::function<void(int, int, task_result&)> callback_multitask_t;
-    callback_multitask_t callback_update_multitask;
-    // for keeping track of all tasks waiting for the result
-    std::set<int> waiting_task_ids;
-    // the main result queue
-    std::vector<task_result> queue_results;
-    std::mutex mutex_results;
-    std::condition_variable condition_results;
-
-    // add the task_id to the list of tasks waiting for response
-    void add_waiting_task_id(int task_id) {
-        LOG_VERBOSE("waiting for task id", {{"task_id", task_id}});
-        std::unique_lock<std::mutex> lock(mutex_results);
-        waiting_task_ids.insert(task_id);
-    }
-
-    // when the request is finished, we can remove task associated with it
-    void remove_waiting_task_id(int task_id) {
-        LOG_VERBOSE("remove waiting for task id", {{"task_id", task_id}});
-        std::unique_lock<std::mutex> lock(mutex_results);
-        waiting_task_ids.erase(task_id);
-    }
-
-    // This function blocks the thread until there is a response for this task_id
-    task_result recv(int task_id) {
-        while (true)
-        {
-            std::unique_lock<std::mutex> lock(mutex_results);
-            condition_results.wait(lock, [&]{
-                return !queue_results.empty();
-            });
-
-            for (int i = 0; i < (int) queue_results.size(); i++)
-            {
-                if (queue_results[i].id == task_id)
-                {
-                    assert(queue_results[i].multitask_id == -1);
-                    task_result res = queue_results[i];
-                    queue_results.erase(queue_results.begin() + i);
-                    return res;
-                }
-            }
-        }
-
-        // should never reach here
-    }
-
-    // Register the function to update multitask
-    void on_multitask_update(callback_multitask_t callback) {
-        callback_update_multitask = callback;
-    }
-
-    // Send a new result to a waiting task_id
-    void send(task_result result) {
-        std::unique_lock<std::mutex> lock(mutex_results);
-        LOG_VERBOSE("send new result", {{"task_id", result.id}});
-        for (auto& task_id : waiting_task_ids) {
-            // LOG_TEE("waiting task id %i \n", task_id);
-            // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
-            if (result.multitask_id == task_id)
-            {
-                LOG_VERBOSE("callback_update_multitask", {{"task_id", task_id}});
-                callback_update_multitask(task_id, result.id, result);
-                continue;
-            }
-
-            if (result.id == task_id)
-            {
-                LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}});
-                queue_results.push_back(result);
-                condition_results.notify_all();
-                return;
-            }
-        }
-    }
-};
-
-//
-// base64 utils (TODO: move to common in the future)
-//
-
-static const std::string base64_chars =
-             "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
-             "abcdefghijklmnopqrstuvwxyz"
-             "0123456789+/";
-
-static inline bool is_base64(uint8_t c)
-{
-    return (isalnum(c) || (c == '+') || (c == '/'));
-}
-
-static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string)
-{
-    int i = 0;
-    int j = 0;
-    int in_ = 0;
-
-    int in_len = encoded_string.size();
-
-    uint8_t char_array_4[4];
-    uint8_t char_array_3[3];
-
-    std::vector<uint8_t> ret;
-
-    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_]))
-    {
-        char_array_4[i++] = encoded_string[in_]; in_++;
-        if (i == 4)
-        {
-            for (i = 0; i <4; i++)
-            {
-                char_array_4[i] = base64_chars.find(char_array_4[i]);
-            }
-
-            char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
-            char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
-            char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
-
-            for (i = 0; (i < 3); i++)
-            {
-                ret.push_back(char_array_3[i]);
-            }
-            i = 0;
-        }
-    }
-
-    if (i)
-    {
-        for (j = i; j <4; j++)
-        {
-            char_array_4[j] = 0;
-        }
-
-        for (j = 0; j <4; j++)
-        {
-            char_array_4[j] = base64_chars.find(char_array_4[j]);
-        }
-
-        char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
-        char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
-        char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
-
-        for (j = 0; (j < i - 1); j++)
-        {
-            ret.push_back(char_array_3[j]);
-        }
-    }
-
-    return ret;
-}
-
-//
-// random string / id
-//
-
-static std::string random_string()
-{
-    static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
-
-    std::random_device rd;
-    std::mt19937 generator(rd());
-
-    std::string result(32, ' ');
-
-    for (int i = 0; i < 32; ++i) {
-        result[i] = str[generator() % str.size()];
-    }
-
-    return result;
-}
-
-static std::string gen_chatcmplid()
-{
-    std::stringstream chatcmplid;
-    chatcmplid << "chatcmpl-" << random_string();
-    return chatcmplid.str();
-}
-
-//
-// other common utils
-//
-
-static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
-{
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++)
-    {
-    }
-    return i;
-}
-
-static bool ends_with(const std::string &str, const std::string &suffix)
-{
-    return str.size() >= suffix.size() &&
-           0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
-}
-
-static size_t find_partial_stop_string(const std::string &stop,
-                                       const std::string &text)
-{
-    if (!text.empty() && !stop.empty())
-    {
-        const char text_last_char = text.back();
-        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--)
-        {
-            if (stop[char_index] == text_last_char)
-            {
-                const std::string current_partial = stop.substr(0, char_index + 1);
-                if (ends_with(text, current_partial))
-                {
-                    return text.size() - char_index - 1;
-                }
-            }
-        }
-    }
-    return std::string::npos;
-}
-
-// TODO: reuse llama_detokenize
-template <class Iter>
-static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
-{
-    std::string ret;
-    for (; begin != end; ++begin)
-    {
-        ret += llama_token_to_piece(ctx, *begin);
-    }
-    return ret;
-}
-
-// format incomplete utf-8 multibyte character for output
-static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
-{
-    std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
-    // if the size is 1 and first bit is 1, meaning it's a partial character
-    //   (size > 1 meaning it's already a known token)
-    if (out.size() == 1 && (out[0] & 0x80) == 0x80)
-    {
-        std::stringstream ss;
-        ss << std::hex << (out[0] & 0xff);
-        std::string res(ss.str());
-        out = "byte: \\x" + res;
-    }
-    return out;
-}
-
-// convert a vector of completion_token_output to json
-static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> &probs)
-{
-    json out = json::array();
-    for (const auto &prob : probs)
-    {
-        json probs_for_token = json::array();
-        for (const auto &p : prob.probs)
-        {
-            std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
-            probs_for_token.push_back(json
-            {
-                {"tok_str", tok_str},
-                {"prob",    p.prob},
-            });
-        }
-        std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
-        out.push_back(json{
-            {"content", tok_str},
-            {"probs",   probs_for_token},
-        });
-    }
-    return out;
-}

llm/generate/gen_common.sh

@@ -1,137 +0,0 @@
-# common logic across linux and darwin
-
-init_vars() {
-    case "${GOARCH}" in
-    "amd64")
-        ARCH="x86_64"
-        ;;
-    "arm64")
-        ARCH="arm64"
-        ;;
-    *)
-        echo "GOARCH must be set"
-        echo "this script is meant to be run from within go generate"
-        exit 1
-        ;;
-    esac
-
-    LLAMACPP_DIR=../llama.cpp
-    CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on"
-    CMAKE_TARGETS="--target ollama_llama_server"
-    if echo "${CGO_CFLAGS}" | grep -- '-g' >/dev/null; then
-        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on -DLLAMA_SERVER_VERBOSE=on ${CMAKE_DEFS}"
-    else
-        # TODO - add additional optimization flags...
-        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=Release -DLLAMA_SERVER_VERBOSE=off ${CMAKE_DEFS}"
-    fi
-    case $(uname -s) in
-    "Darwin")
-        LIB_EXT="dylib"
-        WHOLE_ARCHIVE="-Wl,-force_load"
-        NO_WHOLE_ARCHIVE=""
-        GCC_ARCH="-arch ${ARCH}"
-        DIST_BASE=../../dist/darwin-${GOARCH}/
-        PAYLOAD_BASE=../../build/darwin/${GOARCH}
-        ;;
-    "Linux")
-        LIB_EXT="so"
-        WHOLE_ARCHIVE="-Wl,--whole-archive"
-        NO_WHOLE_ARCHIVE="-Wl,--no-whole-archive"
-
-        # Cross compiling not supported on linux - Use docker
-        GCC_ARCH=""
-        DIST_BASE=../../dist/linux-${GOARCH}/
-        PAYLOAD_BASE=../../build/linux/${GOARCH}
-        ;;
-    *)
-        ;;
-    esac
-    if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
-        CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
-    fi
-    GZIP=$(command -v pigz 2>/dev/null || echo "gzip")
-    RUNNER_BASE="${DIST_BASE}/lib/ollama/runners"
-}
-
-git_module_setup() {
-    if [ -n "${OLLAMA_SKIP_PATCHING}" ]; then
-        echo "Skipping submodule initialization"
-        return
-    fi
-    # Make sure the tree is clean after the directory moves
-    if [ -d "${LLAMACPP_DIR}/gguf" ]; then
-        echo "Cleaning up old submodule"
-        rm -rf ${LLAMACPP_DIR}
-    fi
-    git submodule init
-    git submodule update --force ${LLAMACPP_DIR}
-
-}
-
-apply_patches() {
-    # apply temporary patches until fix is upstream
-    for patch in ../patches/*.patch; do
-        git -c 'user.name=nobody' -c 'user.email=<>' -C ${LLAMACPP_DIR} am ${patch}
-    done
-}
-
-build() {
-    cmake -S ${LLAMACPP_DIR} -B ${BUILD_DIR} ${CMAKE_DEFS}
-    cmake --build ${BUILD_DIR} ${CMAKE_TARGETS} -j8
-    # remove unnecessary build artifacts
-    rm -f ${BUILD_DIR}/bin/ggml-common.h ${BUILD_DIR}/bin/ggml-metal.metal
-}
-
-dist() {
-    [ -z "${RUNNER}" ] && exit 1
-    mkdir -p ${RUNNER_BASE}/${RUNNER}/
-    for f in ${BUILD_DIR}/bin/* ; do
-        cp ${f} ${RUNNER_BASE}/${RUNNER}/
-    done
-    # check for lib directory
-    if [ -d ${BUILD_DIR}/lib ]; then
-        for f in ${BUILD_DIR}/lib/* ; do
-            cp ${f} ${RUNNER_BASE}/${RUNNER}/
-        done
-    fi
-}
-
-# Compress from the build $BUILD_DIR into the $PAYLOAD_BASE/$RUNNER dir
-compress() {
-    [ -z "${RUNNER}" ] && exit 1
-    echo "Compressing payloads with ${GZIP} to reduce overall binary size..."
-    rm -rf "${PAYLOAD_BASE}/${RUNNER}/"
-    mkdir -p "${PAYLOAD_BASE}/${RUNNER}/"
-    for f in ${BUILD_DIR}/bin/* ; do
-        ${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
-        compress_pids+=" $!"
-    done
-    # check for lib directory
-    if [ -d ${BUILD_DIR}/lib ]; then
-        for f in ${BUILD_DIR}/lib/* ; do
-            ${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
-            compress_pids+=" $!"
-        done
-    fi
-    echo
-}
-
-wait_for_compress() {
-    for pid in ${compress_pids}; do
-        wait $pid
-    done
-    echo "Finished compression"
-}
-
-install() {
-    echo "Installing libraries to bin dir ${BUILD_DIR}/bin/"
-    for lib in $(find ${BUILD_DIR} -name \*.${LIB_EXT} | grep -v "${BUILD_DIR}/bin/" ); do
-        rm -f "${BUILD_DIR}/bin/$(basename ${lib})"
-        cp -af "${lib}" "${BUILD_DIR}/bin/"
-    done
-}
-
-# Keep the local tree clean after we're done with the build
-cleanup() {
-    git submodule update --force ${LLAMACPP_DIR}
-}

llm/generate/gen_darwin.sh

@@ -1,91 +0,0 @@
-#!/bin/bash
-# This script is intended to run inside the go generate
-# working directory must be ./llm/generate/
-
-# TODO - add hardening to detect missing tools (cmake, etc.)
-
-set -ex
-set -o pipefail
-compress_pids=""
-echo "Starting darwin generate script"
-source $(dirname $0)/gen_common.sh
-init_vars
-git_module_setup
-apply_patches
-
-sign() {
-    if [ -n "$APPLE_IDENTITY" ]; then
-        codesign -f --timestamp --deep --options=runtime --sign "$APPLE_IDENTITY" --identifier ai.ollama.ollama $1
-    fi
-}
-
-COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DGGML_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
-
-case "${GOARCH}" in
-"amd64")
-    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DGGML_METAL=off -DGGML_NATIVE=off"
-
-    if [ -z "$OLLAMA_SKIP_CPU_GENERATE" ]; then
-        #
-        # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
-        #
-        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-        RUNNER=cpu
-        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
-        echo "Building LCD CPU"
-        build
-        sign ${BUILD_DIR}/bin/ollama_llama_server
-        compress
-
-        #
-        # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
-        # Approximately 400% faster than LCD on same CPU
-        #
-        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-        RUNNER=cpu_avx
-        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
-        echo "Building AVX CPU"
-        build
-        sign ${BUILD_DIR}/bin/ollama_llama_server
-        compress
-
-        #
-        # ~2013 CPU Dynamic library
-        # Approximately 10% faster than AVX on same CPU
-        #
-        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
-        RUNNER=cpu_avx2
-        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
-        echo "Building AVX2 CPU"
-        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
-        build
-        sign ${BUILD_DIR}/bin/ollama_llama_server
-        compress
-    fi
-    ;;
-"arm64")
-
-    if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
-        init_vars
-        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
-        RUNNER="metal"
-        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
-        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
-        build
-        sign ${BUILD_DIR}/bin/ollama_llama_server
-        compress
-    fi
-    ;;
-*)
-    echo "GOARCH must be set"
-    echo "this script is meant to be run from within go generate"
-    exit 1
-    ;;
-esac
-
-cleanup
-wait_for_compress
-echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_linux.sh

@@ -1,285 +0,0 @@
-#!/bin/bash
-# This script is intended to run inside the go generate
-# working directory must be llm/generate/
-
-# First we build one or more CPU based LLM libraries
-#
-# Then if we detect CUDA, we build a CUDA dynamic library, and carry the required
-# library dependencies
-#
-# Then if we detect ROCm, we build a dynamically loaded ROCm lib.  The ROCM
-# libraries are quite large, and also dynamically load data files at runtime
-# which in turn are large, so we don't attempt to cary them as payload
-
-set -ex
-set -o pipefail
-compress_pids=""
-
-# See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
-amdGPUs() {
-    if [ -n "${AMDGPU_TARGETS}" ]; then
-        echo "${AMDGPU_TARGETS}"
-        return
-    fi
-    GPU_LIST=(
-        "gfx900"
-        "gfx906:xnack-"
-        "gfx908:xnack-"
-        "gfx90a:xnack+"
-        "gfx90a:xnack-"
-        "gfx940"
-        "gfx941"
-        "gfx942"
-        "gfx1010"
-        "gfx1012"
-        "gfx1030"
-        "gfx1100"
-        "gfx1101"
-        "gfx1102"
-    )
-    (
-        IFS=$';'
-        echo "'${GPU_LIST[*]}'"
-    )
-}
-
-echo "Starting linux generate script"
-if [ -z "${CUDACXX}" ]; then
-    if [ -x /usr/local/cuda/bin/nvcc ]; then
-        export CUDACXX=/usr/local/cuda/bin/nvcc
-    else
-        # Try the default location in case it exists
-        export CUDACXX=$(command -v nvcc)
-    fi
-fi
-COMMON_CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
-source $(dirname $0)/gen_common.sh
-init_vars
-git_module_setup
-apply_patches
-
-init_vars
-if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
-    # Users building from source can tune the exact flags we pass to cmake for configuring
-    # llama.cpp, and we'll build only 1 CPU variant in that case as the default.
-    if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
-        init_vars
-        echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
-        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
-        RUNNER="cpu"
-        BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
-        echo "Building custom CPU"
-        build
-        install
-        dist
-        compress
-    else
-        # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
-        # -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-        # -DGGML_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
-        # -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-        # -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
-        # Note: the following seem to yield slower results than AVX2 - ymmv
-        # -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
-        # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
-        # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
-
-        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
-        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
-            #
-            # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
-            #
-            init_vars
-            CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-            RUNNER=cpu
-            BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
-            echo "Building LCD CPU"
-            build
-            install
-            dist
-            compress
-        fi
-
-        if [ "${ARCH}" == "x86_64" ]; then
-            #
-            # ARM chips in M1/M2/M3-based MACs and NVidia Tegra devices do not currently support avx extensions.
-            #
-            if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx" ]; then
-                #
-                # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
-                # Approximately 400% faster than LCD on same CPU
-                #
-                init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-                RUNNER=cpu_avx
-                BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
-                echo "Building AVX CPU"
-                build
-                install
-                dist
-                compress
-            fi
-
-            if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx2" ]; then
-                #
-                # ~2013 CPU Dynamic library
-                # Approximately 10% faster than AVX on same CPU
-                #
-                init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
-                RUNNER=cpu_avx2
-                BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
-                echo "Building AVX2 CPU"
-                build
-                install
-                dist
-                compress
-            fi
-        fi
-    fi
-else
-    echo "Skipping CPU generation step as requested"
-fi
-
-# If needed, look for the default CUDA toolkit location
-if [ -z "${CUDA_LIB_DIR}" ] && [ -d /usr/local/cuda/lib64 ]; then
-    CUDA_LIB_DIR=/usr/local/cuda/lib64
-fi
-
-# If needed, look for CUDA on Arch Linux
-if [ -z "${CUDA_LIB_DIR}" ] && [ -d /opt/cuda/targets/x86_64-linux/lib ]; then
-    CUDA_LIB_DIR=/opt/cuda/targets/x86_64-linux/lib
-fi
-
-# Allow override in case libcudart is in the wrong place
-if [ -z "${CUDART_LIB_DIR}" ]; then
-    CUDART_LIB_DIR="${CUDA_LIB_DIR}"
-fi
-
-if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
-    echo "CUDA libraries detected - building dynamic CUDA library"
-    init_vars
-    CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
-    if [ -n "${CUDA_MAJOR}" -a -z "${CUDA_VARIANT}" ]; then
-        CUDA_VARIANT=_v${CUDA_MAJOR}
-    fi
-    if [ "${ARCH}" == "arm64" ]; then
-        echo "ARM CPU detected - disabling unsupported AVX instructions"
-
-        # ARM-based CPUs such as M1 and Tegra do not support AVX extensions.
-        #
-        # CUDA compute < 6.0 lacks proper FP16 support on ARM.
-        # Disabling has minimal performance effect while maintaining compatibility.
-        ARM64_DEFS="-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_CUDA_F16=off"
-    fi
-    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
-    if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
-        echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
-        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
-        echo "Building custom CUDA GPU"
-    else
-        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
-    fi
-    export CUDAFLAGS="-t8"
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS} -DGGML_STATIC=off"
-    RUNNER=cuda${CUDA_VARIANT}
-    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
-    export LLAMA_SERVER_LDFLAGS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
-    CUDA_DIST_DIR="${CUDA_DIST_DIR:-${DIST_BASE}/lib/ollama}"
-    build
-    install
-    dist
-    echo "Installing CUDA dependencies in ${CUDA_DIST_DIR}"
-    mkdir -p "${CUDA_DIST_DIR}"
-    for lib in ${CUDA_LIB_DIR}/libcudart.so* ${CUDA_LIB_DIR}/libcublas.so* ${CUDA_LIB_DIR}/libcublasLt.so* ; do
-        cp -a "${lib}" "${CUDA_DIST_DIR}"
-    done
-    compress
-
-fi
-
-if [ -z "${ONEAPI_ROOT}" ]; then
-    # Try the default location in case it exists
-    ONEAPI_ROOT=/opt/intel/oneapi
-fi
-
-if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
-    echo "OneAPI libraries detected - building dynamic OneAPI library"
-    init_vars
-    source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
-    CC=icx
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
-    RUNNER=oneapi
-    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
-    ONEAPI_DIST_DIR="${DIST_BASE}/lib/ollama"
-    export LLAMA_SERVER_LDFLAGS="-fsycl -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
-    DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
-    build
-
-    # copy oneAPI dependencies
-    mkdir -p "${ONEAPI_DIST_DIR}"
-    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e sycl -e mkl -e tbb); do
-        cp -a "${dep}" "${ONEAPI_DIST_DIR}"
-    done
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${ONEAPI_DIST_DIR}"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${ONEAPI_DIST_DIR}"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${ONEAPI_DIST_DIR}"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${ONEAPI_DIST_DIR}"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${ONEAPI_DIST_DIR}"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${ONEAPI_DIST_DIR}"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${ONEAPI_DIST_DIR}"
-    install
-    dist
-    compress
-fi
-
-if [ -z "${ROCM_PATH}" ]; then
-    # Try the default location in case it exists
-    ROCM_PATH=/opt/rocm
-fi
-
-if [ -z "${CLBlast_DIR}" ]; then
-    # Try the default location in case it exists
-    if [ -d /usr/lib/cmake/CLBlast ]; then
-        export CLBlast_DIR=/usr/lib/cmake/CLBlast
-    fi
-fi
-
-if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
-    echo "ROCm libraries detected - building dynamic ROCm library"
-    if [ -f ${ROCM_PATH}/lib/librocblas.so.*.*.????? ]; then
-        ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
-    fi
-    init_vars
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
-    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
-    if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
-        echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
-        CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
-        echo "Building custom ROCM GPU"
-    fi
-    RUNNER=rocm${ROCM_VARIANT}
-    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
-    # ROCm dependencies are too large to fit into a unified bundle
-    ROCM_DIST_DIR="${DIST_BASE}/../linux-${GOARCH}-rocm/lib/ollama"
-    # TODO figure out how to disable runpath (rpath)
-    # export CMAKE_HIP_FLAGS="-fno-rtlib-add-rpath" # doesn't work
-    export LLAMA_SERVER_LDFLAGS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
-    build
-
-    # copy the ROCM dependencies
-    mkdir -p "${ROCM_DIST_DIR}"
-    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -v "${GOARCH}/rocm${ROCM_VARIANT}" | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf ); do
-        cp -a "${dep}"* "${ROCM_DIST_DIR}"
-        if [ $(readlink -f "${dep}") != "${dep}" ] ; then
-            cp $(readlink -f "${dep}") "${ROCM_DIST_DIR}"
-        fi
-    done
-    install
-    dist
-    compress
-fi
-
-cleanup
-wait_for_compress
-echo "go generate completed.  LLM runners: $(cd ${PAYLOAD_BASE}; echo *)"

llm/generate/gen_windows.ps1

@@ -1,403 +0,0 @@
-#!powershell
-
-$ErrorActionPreference = "Stop"
-
-function amdGPUs {
-    if ($env:AMDGPU_TARGETS) {
-        return $env:AMDGPU_TARGETS
-    }
-    # Current supported rocblas list from ROCm v6.1.2 on windows
-    # https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html#windows-supported-gpus
-    $GPU_LIST = @(
-        "gfx1030"
-        "gfx1100"
-        "gfx1101"
-        "gfx1102"
-    )
-    $GPU_LIST -join ';'
-}
-
-
-function init_vars {
-    write-host "Checking for cmake..."
-    get-command cmake
-    write-host "Checking for ninja..."
-    $d=(get-command -ea 'silentlycontinue' ninja).path
-    if ($null -eq $d) {
-        $MSVC_INSTALL=(Get-CimInstance MSFT_VSInstance -Namespace root/cimv2/vs)[0].InstallLocation
-        $matches=(gci -path $MSVC_INSTALL -r -fi ninja.exe)
-        if ($matches.count -eq 0) {
-            throw "Unable to locate ninja"
-        }
-        $ninjaDir=($matches[0].FullName | split-path -parent)
-        $env:PATH="$env:PATH;$ninjaDir"
-    }
-    if (!$script:SRC_DIR) {
-        $script:SRC_DIR = $(resolve-path "..\..\")
-    }
-    if (!$script:llamacppDir) {
-        $script:llamacppDir = "../llama.cpp"
-    }
-    if (!$script:cmakeTargets) {
-        $script:cmakeTargets = @("ollama_llama_server")
-    }
-    $script:cmakeDefs = @(
-        "-DBUILD_SHARED_LIBS=on",
-        "-DGGML_NATIVE=off",
-        "-DGGML_OPENMP=off"
-        )
-    $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
-    $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
-    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\runners"
-    md "$script:DIST_BASE" -ea 0 > $null
-    if ($env:CGO_CFLAGS -contains "-g") {
-        $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
-        $script:config = "RelWithDebInfo"
-    } else {
-        $script:cmakeDefs += @("-DLLAMA_SERVER_VERBOSE=off", "-DCMAKE_BUILD_TYPE=Release")
-        $script:config = "Release"
-    }
-    if ($null -ne $env:CMAKE_SYSTEM_VERSION) {
-        $script:cmakeDefs += @("-DCMAKE_SYSTEM_VERSION=${env:CMAKE_SYSTEM_VERSION}")
-    }
-    # Try to find the CUDA dir
-    if ($env:CUDA_LIB_DIR -eq $null) {
-        $d=(get-command -ea 'silentlycontinue' nvcc).path
-        if ($d -ne $null) {
-            $script:CUDA_LIB_DIR=($d| split-path -parent)
-            $script:CUDA_INCLUDE_DIR=($script:CUDA_LIB_DIR|split-path -parent)+"\include"
-        }
-    } else {
-        $script:CUDA_LIB_DIR=$env:CUDA_LIB_DIR
-    }
-    $script:DUMPBIN=(get-command -ea 'silentlycontinue' dumpbin).path
-    if ($null -eq $env:CMAKE_CUDA_ARCHITECTURES) {
-        $script:CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
-    } else {
-        $script:CMAKE_CUDA_ARCHITECTURES=$env:CMAKE_CUDA_ARCHITECTURES
-    }
-    # Note: Windows Kits 10 signtool crashes with GCP's plugin
-    if ($null -eq $env:SIGN_TOOL) {
-        ${script:SignTool}="C:\Program Files (x86)\Windows Kits\8.1\bin\x64\signtool.exe"
-    } else {
-        ${script:SignTool}=${env:SIGN_TOOL}
-    }
-    if ("${env:KEY_CONTAINER}") {
-        ${script:OLLAMA_CERT}=$(resolve-path "${script:SRC_DIR}\ollama_inc.crt")
-    }
-}
-
-function git_module_setup {
-    # TODO add flags to skip the init/patch logic to make it easier to mod llama.cpp code in-repo
-    & git submodule init
-    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    & git submodule update --force "${script:llamacppDir}"
-    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-}
-
-function apply_patches {
-    # Apply temporary patches until fix is upstream
-    foreach ($patch in $(Get-ChildItem "../patches/*.patch")) {
-        git -c 'user.name=nobody' -c 'user.email=<>' -C "${script:llamacppDir}" am $patch.FullName
-    }
-}
-
-function build {
-    write-host "generating config with: cmake -S ${script:llamacppDir} -B $script:buildDir $script:cmakeDefs"
-    & cmake --version
-    & cmake -S "${script:llamacppDir}" -B $script:buildDir $script:cmakeDefs
-    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    if ($cmakeDefs -contains "-G") {
-        $extra=@("-j8")
-    } else {
-        $extra= @("--", "/maxCpuCount:8")
-    }
-    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ }) $extra"
-    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ }) $extra
-    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    # Rearrange output to be consistent between different generators
-    if ($null -ne ${script:config} -And (test-path -path "${script:buildDir}/bin/${script:config}" ) ) {
-        mv -force "${script:buildDir}/bin/${script:config}/*" "${script:buildDir}/bin/"
-        remove-item "${script:buildDir}/bin/${script:config}"
-    }
-}
-
-function sign {
-    if ("${env:KEY_CONTAINER}") {
-        write-host "Signing ${script:buildDir}/bin/*.exe  ${script:buildDir}/bin/*.dll"
-        foreach ($file in @(get-childitem "${script:buildDir}/bin/*.exe") + @(get-childitem "${script:buildDir}/bin/*.dll")){
-            & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
-                /csp "Google Cloud KMS Provider" /kc "${env:KEY_CONTAINER}" $file
-            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-        }
-    }
-}
-
-function install {
-    write-host "Installing binaries to dist dir ${script:distDir}"
-    mkdir ${script:distDir} -ErrorAction SilentlyContinue
-    $binaries = dir "${script:buildDir}/bin/*.exe"
-    foreach ($file in $binaries) {
-        copy-item -Path $file -Destination ${script:distDir} -Force
-    }
-
-    write-host "Installing dlls to dist dir ${script:distDir}"
-    $dlls = dir "${script:buildDir}/bin/*.dll"
-    foreach ($file in $dlls) {
-        copy-item -Path $file -Destination ${script:distDir} -Force
-    }
-}
-
-function cleanup {
-    $patches = Get-ChildItem "../patches/*.diff"
-    foreach ($patch in $patches) {
-        # Extract file paths from the patch file
-        $filePaths = Get-Content $patch.FullName | Where-Object { $_ -match '^\+\+\+ ' } | ForEach-Object {
-            $parts = $_ -split ' '
-            ($parts[1] -split '/', 2)[1]
-        }
-
-        # Checkout each file
-        foreach ($file in $filePaths) {
-            git -C "${script:llamacppDir}" checkout $file
-        }
-        git -C "${script:llamacppDir}" checkout CMakeLists.txt
-    }
-}
-
-
-# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
-
-
-function build_cpu_x64 {
-    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
-        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
-        $script:buildDir="../build/windows/${script:ARCH}/cpu"
-        $script:distDir="$script:DIST_BASE\cpu"
-        write-host "Building LCD CPU"
-        build
-        sign
-        install
-    } else {
-        write-host "Skipping CPU generation step as requested"
-    }
-}
-
-function build_cpu_arm64 {
-    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
-        init_vars
-        write-host "Checking for clang..."
-        get-command clang
-        $env:CFLAGS="-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only"
-        $env:CXXFLAGS="$env:CFLAGS"
-        $env:LDFLAGS="-static-libstdc++"
-        $script:cmakeDefs = $script:commonCpuDefs + @(
-            "-DCMAKE_VERBOSE_MAKEFILE=on",
-            "-DCMAKE_C_COMPILER=clang.exe",
-            "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DMSVC_RUNTIME_LIBRARY=MultiThreaded"
-        ) + $script:cmakeDefs
-        $script:buildDir="../build/windows/${script:ARCH}/cpu"
-        $script:distDir="$script:DIST_BASE\cpu"
-        write-host "Building LCD CPU"
-        build
-        sign
-        install
-    } else {
-        write-host "Skipping CPU generation step as requested"
-    }
-}
-
-
-function build_cpu_avx() {
-    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
-        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
-        $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
-        $script:distDir="$script:DIST_BASE\cpu_avx"
-        write-host "Building AVX CPU"
-        build
-        sign
-        install
-    } else {
-        write-host "Skipping CPU AVX generation step as requested"
-    }
-}
-
-function build_cpu_avx2() {
-    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
-        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=on", "-DGGML_AVX512=off", "-DGGML_FMA=on", "-DGGML_F16C=on") + $script:cmakeDefs
-        $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
-        $script:distDir="$script:DIST_BASE\cpu_avx2"
-        write-host "Building AVX2 CPU"
-        build
-        sign
-        install
-    } else {
-        write-host "Skipping CPU AVX2 generation step as requested"
-    }
-}
-
-function build_cuda() {
-    if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${script:CUDA_LIB_DIR}")) {
-        # Then build cuda as a dynamically loaded library
-        $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
-        $script:CUDA_VERSION=((get-item ($nvcc | split-path | split-path)).Basename -Split "\.")[0]
-        if ($null -ne $script:CUDA_VERSION) {
-            $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
-        }
-        init_vars
-        $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
-        $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
-        $script:cmakeDefs += @(
-            "-A", "x64",
-            "-DGGML_CUDA=ON",
-            "-DGGML_AVX=on",
-            "-DGGML_AVX2=off",
-            "-DCMAKE_CUDA_FLAGS=-t6",
-            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}",
-            "-DCMAKE_CUDA_COMPILER_TOOLKIT_ROOT=$env:CUDA_PATH"
-            )
-        if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
-            write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
-            $script:cmakeDefs +=@("${env:OLLAMA_CUSTOM_CUDA_DEFS}")
-            write-host "building custom CUDA GPU"
-        }
-        build
-        sign
-        install
-
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ea 0 > $null
-        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    } else {
-        write-host "Skipping CUDA generation step"
-    }
-}
-
-function build_oneapi() {
-  if ((-not "${env:OLLAMA_SKIP_ONEAPI_GENERATE}") -and ("${env:ONEAPI_ROOT}"))  {
-    # Get oneAPI version
-    $script:ONEAPI_VERSION = icpx --version
-    $script:ONEAPI_VERSION = [regex]::Match($script:ONEAPI_VERSION, '(?<=oneAPI DPC\+\+/C\+\+ Compiler )(?<version>\d+\.\d+\.\d+)').Value
-    if ($null -ne $script:ONEAPI_VERSION) {
-      $script:ONEAPI_VARIANT = "_v" + $script:ONEAPI_VERSION
-    }
-    init_vars
-    $script:buildDir = "../build/windows/${script:ARCH}/oneapi$script:ONEAPI_VARIANT"
-    $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
-    $script:cmakeDefs += @(
-      "-G", "MinGW Makefiles",
-      "-DGGML_SYCL=ON",
-      "-DCMAKE_C_COMPILER=icx",
-      "-DCMAKE_CXX_COMPILER=icx",
-      "-DCMAKE_BUILD_TYPE=Release"
-    )
-
-    Write-Host "Building oneAPI"
-    build
-    # Ninja doesn't prefix with config name
-    if ($null -ne $script:DUMPBIN) {
-      & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | Select-String ".dll"
-    }
-    sign
-    install
-
-    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ea 0 > $null
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-  } else {
-    Write-Host "Skipping oneAPI generation step"
-  }
-}
-
-function build_rocm() {
-    if ((-not "${env:OLLAMA_SKIP_ROCM_GENERATE}") -and ("${env:HIP_PATH}")) {
-        $script:ROCM_VERSION=(get-item $env:HIP_PATH).Basename
-        if ($null -ne $script:ROCM_VERSION) {
-            $script:ROCM_VARIANT="_v"+$script:ROCM_VERSION
-        }
-
-        init_vars
-        $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
-        $script:distDir="$script:DIST_BASE\rocm$script:ROCM_VARIANT"
-        $script:cmakeDefs += @(
-            "-G", "Ninja",
-            "-DCMAKE_C_COMPILER=clang.exe",
-            "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DGGML_HIPBLAS=on",
-            "-DHIP_PLATFORM=amd",
-            "-DGGML_AVX=on",
-            "-DGGML_AVX2=off",
-            "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
-            "-DAMDGPU_TARGETS=$(amdGPUs)",
-            "-DGPU_TARGETS=$(amdGPUs)"
-            )
-
-        # Make sure the ROCm binary dir is first in the path
-        $env:PATH="$env:HIP_PATH\bin;$env:PATH"
-
-        # We have to clobber the LIB var from the developer shell for clang to work properly
-        $env:LIB=""
-        if ($null -ne $env:OLLAMA_CUSTOM_ROCM_DEFS) {
-            write-host "OLLAMA_CUSTOM_ROCM_DEFS=`"${env:OLLAMA_CUSTOM_ROCM_DEFS}`""
-            $script:cmakeDefs += @("${env:OLLAMA_CUSTOM_ROCM_DEFS}")
-            write-host "building custom ROCM GPU"
-        }
-        write-host "Building ROCm"
-        build
-        # Ninja doesn't prefix with config name
-        ${script:config}=""
-        if ($null -ne $script:DUMPBIN) {
-            & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | select-string ".dll"
-        }
-        sign
-        install
-
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\" -ea 0 > $null
-        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
-        # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
-        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\"
-    } else {
-        write-host "Skipping ROCm generation step"
-    }
-}
-
-init_vars
-if ($($args.count) -eq 0) {
-    git_module_setup
-    apply_patches
-    if ($script:ARCH -eq "arm64") {
-        build_cpu_arm64
-    } else { # amd64
-        build_cpu_x64
-        build_cpu_avx
-        build_cpu_avx2
-        build_cuda
-        build_oneapi
-        build_rocm
-    }
-
-    cleanup
-    write-host "`ngo generate completed.  LLM runners: $(get-childitem -path $script:DIST_BASE)"
-} else {
-    for ( $i = 0; $i -lt $args.count; $i++ ) {
-        write-host "performing $($args[$i])"
-        & $($args[$i])
-    }
-}

llm/generate/generate_darwin.go

@@ -1,3 +0,0 @@
-package generate
-
-//go:generate bash ./gen_darwin.sh

llm/generate/generate_linux.go

@@ -1,3 +0,0 @@
-package generate
-
-//go:generate bash ./gen_linux.sh

llm/generate/generate_windows.go

@@ -1,3 +0,0 @@
-package generate
-
-//go:generate powershell -ExecutionPolicy Bypass -File ./gen_windows.ps1

llm/ggml_test.go

@@ -1 +0,0 @@
-package llm

llm/patches/0000-cmakelist.patch

@@ -1,22 +0,0 @@
-From 7a3555098d4591c9b329c677654497ed8cee07ec Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Fri, 23 Aug 2024 11:27:48 -0700
-Subject: [PATCH] patch cmakelist
-
----
- CMakeLists.txt | 2 ++
- 1 file changed, 2 insertions(+)
-
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 415743c2..aaadd13e 100644
---- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -210,3 +210,5 @@ if (LLAMA_BUILD_EXAMPLES)
-     add_subdirectory(examples)
-     add_subdirectory(pocs)
- endif()
-+
-+add_subdirectory(../ext_server ext_server) # ollama
--- 
-2.39.3 (Apple Git-146)
-

llm/patches/0001-load-progress.patch

@@ -1,44 +0,0 @@
-From c97ed60c3369294d5551ba099a88ddc509687df1 Mon Sep 17 00:00:00 2001
-From: Gabe Goodhart <ghart@us.ibm.com>
-Date: Thu, 19 Sep 2024 16:55:15 -0600
-Subject: [PATCH] patch load progress
-
----
- common/common.cpp | 2 ++
- common/common.h   | 7 +++++++
- 2 files changed, 9 insertions(+)
-
-diff --git a/common/common.cpp b/common/common.cpp
-index 8d0ed4f9..a09e8a53 100644
---- a/common/common.cpp
-+++ b/common/common.cpp
-@@ -955,6 +955,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
-     mparams.use_mmap        = params.use_mmap;
-     mparams.use_mlock       = params.use_mlock;
-     mparams.check_tensors   = params.check_tensors;
-+    mparams.progress_callback = params.progress_callback;
-+    mparams.progress_callback_user_data = params.progress_callback_user_data;
-     if (params.kv_overrides.empty()) {
-         mparams.kv_overrides = NULL;
-     } else {
-diff --git a/common/common.h b/common/common.h
-index cb87c447..818a4a4a 100644
---- a/common/common.h
-+++ b/common/common.h
-@@ -266,6 +266,13 @@ struct gpt_params {
-     std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
-     std::vector<std::string> image; // path to image file(s)
- 
-+    // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
-+    // If the provided progress_callback returns true, model loading continues.
-+    // If it returns false, model loading is immediately aborted.
-+    llama_progress_callback progress_callback = NULL;
-+    // context pointer passed to the progress callback
-+    void * progress_callback_user_data;
-+
-     // embedding
-     bool embedding         = false; // get only sentence embedding
-     int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
--- 
-2.39.3 (Apple Git-146)
-

llm/patches/0002-clip-log.patch

@@ -1,24 +0,0 @@
-From 6fdf4268e13e56f0050fa6a29b029cbd54be49d2 Mon Sep 17 00:00:00 2001
-From: Gabe Goodhart <ghart@us.ibm.com>
-Date: Thu, 19 Sep 2024 16:58:03 -0600
-Subject: [PATCH] clip log
-
----
- examples/llava/clip.cpp | 1 +
- 1 file changed, 1 insertion(+)
-
-diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index 8aa7b075..b8941c74 100644
---- a/examples/llava/clip.cpp
-+++ b/examples/llava/clip.cpp
-@@ -3,6 +3,7 @@
- // I'll gradually clean and extend it
- // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
- #include "clip.h"
-+#include "common.h"
- #include "ggml.h"
- #include "ggml-alloc.h"
- #include "ggml-backend.h"
--- 
-2.39.3 (Apple Git-146)
-

llm/patches/0003-load_exception.patch

@@ -1,57 +0,0 @@
-From 4f2b9cd0f012c49f40d0784454864ad41ca418b2 Mon Sep 17 00:00:00 2001
-From: Gabe Goodhart <ghart@us.ibm.com>
-Date: Thu, 19 Sep 2024 17:00:28 -0600
-Subject: [PATCH] load exception
-
----
- src/llama.cpp | 25 ++++++++++++++++---------
- 1 file changed, 16 insertions(+), 9 deletions(-)
-
-diff --git a/src/llama.cpp b/src/llama.cpp
-index af8afd84..4d1db3d5 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -8871,7 +8871,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
-         }
-     } catch (const std::exception & err) {
-         LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
--        return -1;
-+        throw;
-     }
- 
-     // loading time will be recalculate after the first eval, so
-@@ -18675,16 +18675,23 @@ struct llama_model * llama_load_model_from_file(
-         }
-         model->rpc_servers.push_back(servers);
-     }
--    int status = llama_model_load(path_model, *model, params);
--    GGML_ASSERT(status <= 0);
--    if (status < 0) {
--        if (status == -1) {
--            LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
--        } else if (status == -2) {
--            LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
-+
-+    try {
-+        int status = llama_model_load(path_model, *model, params);
-+        GGML_ASSERT(status <= 0);
-+        if (status < 0) {
-+            if (status == -1) {
-+                LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
-+            } else if (status == -2) {
-+                LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
-+            }
-+            delete model;
-+            return nullptr;
-         }
-+    } catch (...) {
-+        LLAMA_LOG_ERROR("%s: exception loading model\n", __func__);
-         delete model;
--        return nullptr;
-+        throw;
-     }
- 
-     return model;
--- 
-2.39.3 (Apple Git-146)
-

llm/patches/0007-clip-unicode.patch

@@ -1,54 +0,0 @@
-From 01c42149cbdc194644a2f138598029938e0dd447 Mon Sep 17 00:00:00 2001
-From: Gabe Goodhart <ghart@us.ibm.com>
-Date: Thu, 19 Sep 2024 17:09:57 -0600
-Subject: [PATCH] clip unicode
-
----
- examples/llava/clip.cpp | 23 +++++++++++++++++++++++
- 1 file changed, 23 insertions(+)
-
-diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index b8941c74..3a735f17 100644
---- a/examples/llava/clip.cpp
-+++ b/examples/llava/clip.cpp
-@@ -40,6 +40,14 @@
- #include <cinttypes>
- #include <limits>
- 
-+#if defined(_WIN32)
-+#define WIN32_LEAN_AND_MEAN
-+#ifndef NOMINMAX
-+    #define NOMINMAX
-+#endif
-+#include <windows.h>
-+#endif
-+
- #define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
- #define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
- #define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-@@ -1227,7 +1235,22 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
-             return nullptr;
-         }
- 
-+#ifdef _WIN32
-+        int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
-+        if (!wlen) {
-+            return NULL;
-+        }
-+        wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
-+        wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wbuf, wlen);
-+        if (!wlen) {
-+            free(wbuf);
-+            return NULL;
-+        }
-+        auto fin = std::ifstream(wbuf, std::ios::binary);
-+        free(wbuf);
-+#else
-         auto fin = std::ifstream(fname, std::ios::binary);
-+#endif
-         if (!fin) {
-             LOG_ERR("cannot open model file for loading tensors\n");
-             clip_free(new_clip);
--- 
-2.39.3 (Apple Git-146)
-

llm/patches/0008-solar-pro.patch

@@ -1,412 +0,0 @@
-From a8fe40fa7b026d2db9bb6aeecd24fcd2027110ec Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Mon, 16 Sep 2024 15:53:16 -0700
-Subject: [PATCH] add solar-pro support
-
-solar-pro introduces block skip connections where blocks are connected
-to other, non-sequential blocks with a scale multiple
-
-this change adds 4 new keys to store the skip connections and one new
-tensor to store the scalar. the scalar is implemented a 1-dimensional
-tensor with 2 elements dervied from the model's bskcn_tv configuration.
-in general, the values are (bskcn_tv, 1 - bskcn_tv)
----
- src/llama.cpp | 270 +++++++++++++++++++++++++++++++++++++++++++++++---
- 1 file changed, 255 insertions(+), 15 deletions(-)
-
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 4c0a1bb6..c6fc0c3f 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -217,6 +217,7 @@ enum llm_arch {
-     LLM_ARCH_GRANITE,
-     LLM_ARCH_GRANITE_MOE,
-     LLM_ARCH_CHAMELEON,
-+    LLM_ARCH_SOLAR,
-     LLM_ARCH_UNKNOWN,
- };
- 
-@@ -270,6 +271,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
-     { LLM_ARCH_GRANITE,         "granite"      },
-     { LLM_ARCH_GRANITE_MOE,     "granitemoe"   },
-     { LLM_ARCH_CHAMELEON,       "chameleon"    },
-+    { LLM_ARCH_SOLAR,           "solar"        },
-     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
- };
- 
-@@ -327,6 +329,7 @@ enum llm_kv {
-     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
-     LLM_KV_ATTENTION_SLIDING_WINDOW,
-     LLM_KV_ATTENTION_SCALE,
-+    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
- 
-     LLM_KV_ROPE_DIMENSION_COUNT,
-     LLM_KV_ROPE_FREQ_BASE,
-@@ -421,20 +424,21 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
-     { LLM_KV_RESIDUAL_SCALE,                    "%s.residual_scale"                    },
-     { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },
- 
--    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
--    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
--    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"         },
--    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"              },
--    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"             },
--    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"           },
--    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"     },
--    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon" },
--    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                 },
--    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"            },
--    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
--    { 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_HEAD_COUNT,             "%s.attention.head_count"               },
-+    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"            },
-+    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"           },
-+    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"                },
-+    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"               },
-+    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"             },
-+    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"       },
-+    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon"   },
-+    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                   },
-+    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"              },
-+    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"             },
-+    { 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_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection.%d" },
- 
-     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
-     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
-@@ -608,6 +612,7 @@ enum llm_tensor {
-     LLM_TENSOR_ENC_OUTPUT_NORM,
-     LLM_TENSOR_CLS,
-     LLM_TENSOR_CLS_OUT,
-+    LLM_TENSOR_BSKCN_TV,
- };
- 
- static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
-@@ -1527,6 +1532,25 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
-             { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
-         },
-     },
-+
-+    {
-+        LLM_ARCH_SOLAR,
-+        {
-+            { 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_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_BSKCN_TV,        "bskcn_tv" },
-+        },
-+    },
-     {
-         LLM_ARCH_UNKNOWN,
-         {
-@@ -2360,6 +2384,7 @@ enum e_model {
-     MODEL_15B,
-     MODEL_16B,
-     MODEL_20B,
-+    MODEL_22B,
-     MODEL_30B,
-     MODEL_34B,
-     MODEL_35B,
-@@ -2409,6 +2434,8 @@ struct llama_hparams {
-     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
-     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
- 
-+    std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
-+
-     uint32_t n_layer_dense_lead = 0;
-     uint32_t n_lora_q = 0;
-     uint32_t n_lora_kv = 0;
-@@ -2479,6 +2506,7 @@ struct llama_hparams {
-         if (this->n_head_arr    != other.n_head_arr)    return true;
-         if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
-         if (this->n_ff_arr      != other.n_ff_arr)      return true;
-+        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
- 
-         if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
-         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
-@@ -2588,6 +2616,14 @@ struct llama_hparams {
-             return ssm_d_state * ssm_d_inner;
-         }
-     }
-+
-+    bool n_bskcn(uint32_t n, uint32_t il = 0) const {
-+        if (il < n_layer) {
-+            return n_bskcn_arr[n][il] > 0;
-+        }
-+
-+        GGML_ABORT("fatal error");
-+    }
- };
- 
- static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
-@@ -2769,6 +2805,8 @@ struct llama_layer {
-     struct ggml_tensor * ffn_gate_scale;
-     struct ggml_tensor * ffn_up_scale;
-     struct ggml_tensor * ffn_down_scale;
-+
-+    struct ggml_tensor * bskcn_tv;
- };
- 
- // very similar to llama_batch,
-@@ -6134,6 +6172,21 @@ static void llm_load_hparams(
-                     default: model.type = e_model::MODEL_UNKNOWN;
-                }
-             } break;
-+        case LLM_ARCH_SOLAR:
-+            {
-+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-+
-+                for (int i = 0; i < hparams.n_bskcn_arr.max_size(); ++i) {
-+                    auto & bskcn = hparams.n_bskcn_arr.at(i);
-+                    bskcn.fill(0);
-+                    ml.get_key_or_arr(::format(LLM_KV_NAMES.at(LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION), LLM_ARCH_NAMES.at(ml.llm_kv.arch), i), bskcn, hparams.n_layer, false);
-+                }
-+
-+                switch (hparams.n_layer) {
-+                    case 64: model.type = e_model::MODEL_22B; break;
-+                    default: model.type = e_model::MODEL_UNKNOWN;
-+                }
-+            }
-         default: (void)0;
-     }
- 
-@@ -8839,6 +8892,37 @@ static bool llm_load_tensors(
- 
-                         layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
- 
-+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
-+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
-+                    }
-+                } break;
-+            case LLM_ARCH_SOLAR:
-+                {
-+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-+
-+                    // output
-+                    {
-+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
-+                    }
-+
-+                    for (int i = 0; i < n_layer; ++i) {
-+                        ggml_context * ctx_layer = ctx_for_layer(i);
-+                        ggml_context * ctx_split = ctx_for_layer_split(i);
-+
-+                        auto & layer = model.layers[i];
-+
-+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
-+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
-+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
-+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
-+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
-+
-+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-+
-+                        layer.bskcn_tv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_BSKCN_TV, "weight"), {2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
-+
-                         layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
-                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
-@@ -16009,7 +16093,6 @@ struct llm_build_context {
- 
-         return gf;
-     }
--
-     // ref: https://github.com/facebookresearch/chameleon
-     // based on the original build_llama() function, changes:
-     //   * qk-norm
-@@ -16187,6 +16270,158 @@ struct llm_build_context {
- 
-         return gf;
-     }
-+
-+    ggml_cgraph * build_solar() {
-+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
-+
-+        // mutable variable, needed during the last layer of the computation to skip unused tokens
-+        int32_t n_tokens = this->n_tokens;
-+
-+        const int64_t n_embd_head = hparams.n_embd_head_v;
-+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-+        GGML_ASSERT(n_embd_head == hparams.n_rot);
-+
-+        struct ggml_tensor * cur;
-+        struct ggml_tensor * inpL;
-+
-+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
-+
-+        // inp_pos - contains the positions
-+        struct ggml_tensor * inp_pos = build_inp_pos();
-+
-+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-+
-+        struct ggml_tensor * bskcn_1;
-+        struct ggml_tensor * bskcn_2;
-+
-+        for (int il = 0; il < n_layer; ++il) {
-+            struct ggml_tensor * inpSA = inpL;
-+
-+            if (hparams.n_bskcn(0, il)) {
-+                bskcn_1 = inpSA;
-+            }
-+
-+            if (hparams.n_bskcn(1, il)) {
-+                bskcn_2 = inpSA;
-+            }
-+
-+            if (hparams.n_bskcn(2, il)) {
-+                inpSA = ggml_add(
-+                   ctx0,
-+                   ggml_mul(ctx0, bskcn_1, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
-+                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
-+            }
-+
-+            if (hparams.n_bskcn(3, il)) {
-+                inpSA = ggml_add(
-+                   ctx0,
-+                   ggml_mul(ctx0, bskcn_2, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
-+                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
-+            }
-+
-+            // norm
-+            cur = llm_build_norm(ctx0, inpL, hparams,
-+                    model.layers[il].attn_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "attn_norm", il);
-+
-+            // self-attention
-+            {
-+                // rope freq factors for llama3; may return nullptr for llama2 and other models
-+                struct ggml_tensor * rope_factors = build_rope_factors(il);
-+
-+                // compute Q and K and RoPE them
-+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-+                cb(Qcur, "Qcur", il);
-+                if (model.layers[il].bq) {
-+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-+                    cb(Qcur, "Qcur", il);
-+                }
-+
-+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-+                cb(Kcur, "Kcur", il);
-+                if (model.layers[il].bk) {
-+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-+                    cb(Kcur, "Kcur", il);
-+                }
-+
-+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-+                cb(Vcur, "Vcur", il);
-+                if (model.layers[il].bv) {
-+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-+                    cb(Vcur, "Vcur", il);
-+                }
-+
-+                Qcur = ggml_rope_ext(
-+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
-+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-+                    ext_factor, attn_factor, beta_fast, beta_slow
-+                );
-+                cb(Qcur, "Qcur", il);
-+
-+                Kcur = ggml_rope_ext(
-+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
-+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-+                    ext_factor, attn_factor, beta_fast, beta_slow
-+                );
-+                cb(Kcur, "Kcur", il);
-+
-+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-+                        model.layers[il].wo, model.layers[il].bo,
-+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-+            }
-+
-+            if (il == n_layer - 1) {
-+                // skip computing output for unused tokens
-+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-+                n_tokens = n_outputs;
-+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-+            }
-+
-+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-+            cb(ffn_inp, "ffn_inp", il);
-+
-+            // feed-forward network
-+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-+                    model.layers[il].ffn_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "ffn_norm", il);
-+
-+            cur = llm_build_ffn(ctx0, lctx, cur,
-+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-+                    NULL,
-+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-+            cb(cur, "ffn_out", il);
-+
-+            cur = ggml_add(ctx0, cur, ffn_inp);
-+            cb(cur, "ffn_out", il);
-+
-+            cur = lctx.cvec.apply_to(ctx0, cur, il);
-+            cb(cur, "l_out", il);
-+
-+            // input for next layer
-+            inpL = cur;
-+        }
-+
-+        cur = inpL;
-+
-+        cur = llm_build_norm(ctx0, cur, hparams,
-+                model.output_norm, NULL,
-+                LLM_NORM_RMS, cb, -1);
-+        cb(cur, "result_norm", -1);
-+
-+        // lm_head
-+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-+        cb(cur, "result_output", -1);
-+
-+        ggml_build_forward_expand(gf, cur);
-+
-+        return gf;
-+    }
- };
- 
- static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
-@@ -16451,6 +16686,10 @@ static struct ggml_cgraph * llama_build_graph(
-             {
-                 result = llm.build_chameleon();
-             } break;
-+        case LLM_ARCH_SOLAR:
-+            {
-+                result = llm.build_solar();
-+            } break;
-         default:
-             GGML_ABORT("fatal error");
-     }
-@@ -19594,6 +19833,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
-         case LLM_ARCH_GRANITE:
-         case LLM_ARCH_GRANITE_MOE:
-         case LLM_ARCH_CHAMELEON:
-+        case LLM_ARCH_SOLAR:
-             return LLAMA_ROPE_TYPE_NORM;
- 
-         // the pairs of head values are offset by n_rot/2
--- 
-2.39.3 (Apple Git-146)
-

runners/README.md

@@ -0,0 +1,3 @@
+# `runners`
+
+Ollama uses a subprocess model to run one or more child processes to load the LLM.  On some platforms (Linux non-containerized, MacOS) these executables are carried as payloads inside the main executable via the ../build package.  Extraction and discovery of these runners at runtime is implemented in this package.  This package also provides the abstraction to communicate with these subprocesses. 

runners/common.go

@@ -2,6 +2,7 @@ package runners
 
 import (
 	"compress/gzip"
+	"context"
 	"errors"
 	"fmt"
 	"io"
@@ -15,9 +16,11 @@ import (
 	"strings"
 	"sync"
 	"syscall"
+	"time"
 
 	"golang.org/x/sync/errgroup"
 
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/discover"
 	"github.com/ollama/ollama/envconfig"
 )
@@ -31,6 +34,36 @@ var (
 	runnersDir = ""
 )
 
+type CompletionRequest struct {
+	Prompt  string
+	Format  string
+	Images  []ImageData
+	Options *api.Options
+}
+
+type CompletionResponse struct {
+	Content            string
+	DoneReason         string
+	Done               bool
+	PromptEvalCount    int
+	PromptEvalDuration time.Duration
+	EvalCount          int
+	EvalDuration       time.Duration
+}
+
+type LLMServer interface {
+	Ping(ctx context.Context) error
+	WaitUntilRunning(ctx context.Context) error
+	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
+	Embedding(ctx context.Context, input string) ([]float32, error)
+	Tokenize(ctx context.Context, content string) ([]int, error)
+	Detokenize(ctx context.Context, tokens []int) (string, error)
+	Close() error
+	EstimatedVRAM() uint64 // Total VRAM across all GPUs
+	EstimatedTotal() uint64
+	EstimatedVRAMByGPU(gpuID string) uint64
+}
+
 // Return the location where runners are stored
 // If runners are payloads, this will either extract them
 // or refresh them if any have disappeared due to tmp cleaners

runners/llama-server.go

@@ -1,4 +1,4 @@
-package llm
+package runners
 
 import (
 	"bufio"
@@ -28,24 +28,11 @@ import (
 	"github.com/ollama/ollama/build"
 	"github.com/ollama/ollama/discover"
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/fileutils"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/llama"
-	"github.com/ollama/ollama/runners"
 )
 
-type LlamaServer interface {
-	Ping(ctx context.Context) error
-	WaitUntilRunning(ctx context.Context) error
-	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
-	Embedding(ctx context.Context, input string) ([]float32, error)
-	Tokenize(ctx context.Context, content string) ([]int, error)
-	Detokenize(ctx context.Context, tokens []int) (string, error)
-	Close() error
-	EstimatedVRAM() uint64 // Total VRAM across all GPUs
-	EstimatedTotal() uint64
-	EstimatedVRAMByGPU(gpuID string) uint64
-}
-
 // llmServer is an instance of the llama.cpp server
 type llmServer struct {
 	port        int
@@ -58,7 +45,7 @@ type llmServer struct {
 	modelLock   sync.Mutex   // Temporary until we switch fully to Go server
 	model       *llama.Model // If non-nil, the runner is a new Go server
 
-	estimate    MemoryEstimate
+	estimate    fileutils.MemoryEstimate
 	totalLayers uint64
 	// gpuCount     int
 	gpus         discover.GpuInfoList // Recorded just before the model loaded, free space will be incorrect
@@ -68,32 +55,12 @@ type llmServer struct {
 	sem *semaphore.Weighted
 }
 
-// LoadModel will load a model from disk. The model must be in the GGML format.
-//
-// It collects array values for arrays with a size less than or equal to
-// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
-// the maxArraySize is negative, all arrays are collected.
-func LoadModel(model string, maxArraySize int) (*GGML, error) {
-	if _, err := os.Stat(model); err != nil {
-		return nil, err
-	}
-
-	f, err := os.Open(model)
-	if err != nil {
-		return nil, err
-	}
-	defer f.Close()
-
-	ggml, _, err := DecodeGGML(f, maxArraySize)
-	return ggml, err
-}
-
 // NewLlamaServer will run a server for the given GPUs
 // The gpu list must be a single family.
-func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) {
+func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *fileutils.GGML, adapters, projectors []string, opts api.Options, numParallel int) (LLMServer, error) {
 	var err error
 	var cpuRunner string
-	var estimate MemoryEstimate
+	var estimate fileutils.MemoryEstimate
 	var systemTotalMemory uint64
 	var systemFreeMemory uint64
 	var systemSwapFreeMemory uint64
@@ -109,10 +76,10 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 		gpus = discover.GetCPUInfo()
 	}
 	if len(gpus) == 1 && gpus[0].Library == "cpu" {
-		cpuRunner = runners.ServerForCpu()
-		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
+		cpuRunner = ServerForCpu()
+		estimate = fileutils.EstimateGPULayers(gpus, ggml, projectors, opts)
 	} else {
-		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
+		estimate = fileutils.EstimateGPULayers(gpus, ggml, projectors, opts)
 
 		switch {
 		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:
@@ -121,7 +88,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 			opts.NumGPU = 0
 		case gpus[0].Library != "metal" && estimate.Layers == 0:
 			// Don't bother loading into the GPU if no layers can fit
-			cpuRunner = runners.ServerForCpu()
+			cpuRunner = ServerForCpu()
 			gpus = discover.GetCPUInfo()
 		case opts.NumGPU < 0 && estimate.Layers > 0 && gpus[0].Library != "cpu":
 			opts.NumGPU = estimate.Layers
@@ -139,7 +106,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 		}
 	}
 
-	estimate.log()
+	estimate.Log()
 
 	// Loop through potential servers
 	finalErr := errors.New("no suitable llama servers found")
@@ -148,12 +115,12 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
 	}
 
-	rDir, err := runners.Refresh(build.EmbedFS)
+	rDir, err := Refresh(build.EmbedFS)
 	if err != nil {
 		return nil, err
 	}
 
-	availableServers := runners.GetAvailableServers(rDir)
+	availableServers := GetAvailableServers(rDir)
 	if len(availableServers) == 0 {
 		return nil, finalErr
 	}
@@ -161,7 +128,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 	if cpuRunner != "" {
 		servers = []string{cpuRunner}
 	} else {
-		servers = runners.ServersForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
+		servers = ServersForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
 	}
 	demandLib := envconfig.LLMLibrary()
 	if demandLib != "" {
@@ -325,7 +292,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 		_, err := os.Stat(server)
 		if errors.Is(err, os.ErrNotExist) {
 			slog.Warn("llama server disappeared, reinitializing payloads", "path", server, "error", err)
-			_, err = runners.Refresh(build.EmbedFS)
+			_, err = Refresh(build.EmbedFS)
 			if err != nil {
 				slog.Warn("failed to reinitialize payloads", "error", err)
 				return nil, err
@@ -442,26 +409,6 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 	return nil, finalErr
 }
 
-func projectorMemoryRequirements(filename string) uint64 {
-	file, err := os.Open(filename)
-	if err != nil {
-		return 0
-	}
-	defer file.Close()
-
-	ggml, _, err := DecodeGGML(file, 0)
-	if err != nil {
-		return 0
-	}
-
-	var mem uint64
-	for _, layer := range ggml.Tensors().Layers() {
-		mem += layer.size()
-	}
-
-	return mem
-}
-
 type ServerStatus int
 
 const ( // iota is reset to 0
@@ -673,8 +620,9 @@ ws ::= ([ \t\n] ws)?
 const maxBufferSize = 512 * format.KiloByte
 
 type ImageData struct {
-	Data []byte `json:"data"`
-	ID   int    `json:"id"`
+	Data          []byte `json:"data"`
+	ID            int    `json:"id"`
+	AspectRatioID int    `json:"aspect_ratio_id"`
 }
 
 type completion struct {
@@ -692,23 +640,6 @@ type completion struct {
 	}
 }
 
-type CompletionRequest struct {
-	Prompt  string
-	Format  string
-	Images  []ImageData
-	Options *api.Options
-}
-
-type CompletionResponse struct {
-	Content            string
-	DoneReason         string
-	Done               bool
-	PromptEvalCount    int
-	PromptEvalDuration time.Duration
-	EvalCount          int
-	EvalDuration       time.Duration
-}
-
 func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {
 	if err := s.sem.Acquire(ctx, 1); err != nil {
 		slog.Error("Failed to acquire semaphore", "error", err)