Add v12 + v13 cuda support (#12000)

* Add support for upcoming NVIDIA Jetsons The latest Jetsons with JetPack 7 are moving to an SBSA compatible model and will not require building a JetPack specific variant. * cuda: bring back dual versions This adds back dual CUDA versions for our releases, with v11 and v13 to cover a broad set of GPUs and driver versions. * win: break up native builds in build_windows.ps1 * v11 build working on windows and linux * switch to cuda v12.8 not JIT * Set CUDA compression to size * enhance manual install linux docs
2025-09-26 00:47:05 +02:00 · 2025-09-10 12:05:18 -07:00
parent 8d6fffaead
commit 17a023f34b
8 changed files with 146 additions and 25 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -67,12 +67,21 @@ jobs:
            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
            cuda-version: '12.8'
            flags: ''
+            runner_dir: 'cuda_v12'
+          - os: windows
+            arch: amd64
+            preset: 'CUDA 13'
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
+            cuda-version: '13.0'
+            flags: ''
+            runner_dir: 'cuda_v13'
          - os: windows
            arch: amd64
            preset: 'ROCm 6'
            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
            rocm-version: '6.2'
            flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
+            runner_dir: ''
    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
    environment: release
    env:
@@ -138,7 +147,7 @@ jobs:
        run: |
          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
+          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} -DOLLAMA_RUNNER_DIR="${{ matrix.runner_dir }}"
          cmake --build --parallel --preset "${{ matrix.preset }}"
          cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || 'CPU' }}" --strip --parallel 8
        env:
@@ -232,7 +241,7 @@ jobs:
            case "$COMPONENT" in
              bin/ollama)                echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_sbsa)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
              lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
              lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
              lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -46,7 +46,7 @@ jobs:
        include:
          - preset: CPU
          - preset: CUDA
-            container: nvidia/cuda:12.8.1-devel-ubuntu22.04
+            container: nvidia/cuda:13.0.0-devel-ubuntu22.04
            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
          - preset: ROCm
            container: rocm/dev-ubuntu-22.04:6.1.2
@@ -78,7 +78,7 @@ jobs:
        include:
          - preset: CPU
          - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
          - preset: ROCm
            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
@@ -102,7 +102,7 @@ jobs:
          $ErrorActionPreference = "Stop"
          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_12.8", "nvcc_12.8", "cublas_12.8", "cublas_dev_12.8")) -NoNewWindow -Wait
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_13.0", "nvcc_13.0", "cublas_13.0", "cublas_dev_13.0")) -NoNewWindow -Wait
          }

          $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -25,7 +25,7 @@ set(GGML_LLAMAFILE ON)
 set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
 set(GGML_CUDA_GRAPHS ON)
 set(GGML_CUDA_FA ON)
-set(GGML_CUDA_COMPRESSION_MODE default)
+set(GGML_CUDA_COMPRESSION_MODE size)

 if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
    OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
@@ -38,7 +38,7 @@ if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
 endif()

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

 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
@@ -81,7 +81,7 @@ if(CMAKE_CUDA_COMPILER)
    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
    install(TARGETS ggml-cuda
        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_LIBRARY_DIR}
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
            PRE_INCLUDE_REGEXES cublas cublasLt cudart
            PRE_EXCLUDE_REGEXES ".*"
        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
--- a/CMakePresets.json
+++ b/CMakePresets.json
@@ -18,6 +18,14 @@
      "name": "CUDA",
      "inherits": [ "Default" ]
    },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "50-virtual;60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
+      }
+    },
    {
      "name": "CUDA 12",
      "inherits": [ "CUDA" ],
@@ -26,6 +34,14 @@
        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
      }
    },
+    {
+      "name": "CUDA 13",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;110-virtual;120-virtual;121-virtual",
+        "CMAKE_CUDA_FLAGS": "-t 2"
+      }
+    },
    {
      "name": "JetPack 5",
      "inherits": [ "CUDA" ],
@@ -72,11 +88,21 @@
      "configurePreset": "CUDA",
      "targets": [ "ggml-cuda" ]
    },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 11"
+    },
    {
      "name": "CUDA 12",
      "inherits": [ "CUDA" ],
      "configurePreset": "CUDA 12"
    },
+    {
+      "name": "CUDA 13",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 13"
+    },
    {
      "name": "JetPack 5",
      "inherits": [ "CUDA" ],
--- a/30
+++ b/30
@@ -39,15 +39,35 @@ RUN --mount=type=cache,target=/root/.ccache \
        && cmake --build --parallel --preset 'CPU' \
        && cmake --install build --component CPU --strip --parallel 8

+FROM base AS cuda-11
+ARG CUDA11VERSION=11.8
+RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
+ENV PATH=/usr/local/cuda-11/bin:$PATH
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 11' -DOLLAMA_RUNNER_DIR="cuda_v11" \
+        && cmake --build --parallel --preset 'CUDA 11' \
+        && cmake --install build --component CUDA --strip --parallel 8
+
 FROM base AS cuda-12
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
 ENV PATH=/usr/local/cuda-12/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 12' \
+    cmake --preset 'CUDA 12' -DOLLAMA_RUNNER_DIR="cuda_v12"\
        && cmake --build --parallel --preset 'CUDA 12' \
        && cmake --install build --component CUDA --strip --parallel 8

+
+FROM base AS cuda-13
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-}
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 13' -DOLLAMA_RUNNER_DIR="cuda_v13" \
+        && cmake --build --parallel --preset 'CUDA 13' \
+        && cmake --install build --component CUDA --strip --parallel 8
+
+
 FROM base AS rocm-6
 ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
@@ -92,10 +112,14 @@ RUN --mount=type=cache,target=/root/.cache/go-build \
    go build -trimpath -buildmode=pie -o /bin/ollama .

 FROM --platform=linux/amd64 scratch AS amd64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama
+# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/

 FROM --platform=linux/arm64 scratch AS arm64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/cuda_sbsa
+# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
 COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/cuda_jetpack5
 COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/cuda_jetpack6

--- a/discover/cuda_common.go
+++ b/discover/cuda_common.go
@@ -43,14 +43,15 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 				}
 			}
 		}
-		return "sbsa"
 	}

-	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
-	if gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
-		// The detected driver is older than Feb 2023
-		slog.Warn("old CUDA driver detected - please upgrade to a newer driver", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
-		return "v11"
+	if gpuInfo.DriverMajor < 13 {
+		// The detected driver is older than 580 (Aug 2025)
+		// Warn if their CC is compatible with v13 and they should upgrade their driver to get better performance
+		if gpuInfo.computeMajor > 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor >= 5) {
+			slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
+		}
+		return "v12"
 	}
-	return "v12"
+	return "v13"
 }
--- a/docs/linux.md
+++ b/docs/linux.md
@@ -11,12 +11,13 @@ curl -fsSL https://ollama.com/install.sh | sh
 ## Manual install

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

 Download and extract the package:

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

--- a/scripts/build_windows.ps1
+++ b/scripts/build_windows.ps1
@@ -78,7 +78,7 @@ function checkEnv() {
 }


-function buildOllama() {
+function buildCPU() {
    mkdir -Force -path "${script:DIST_DIR}\"
    if ($script:ARCH -ne "arm64") {
        Remove-Item -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}"
@@ -90,20 +90,72 @@ function buildOllama() {
        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
        & cmake --install build --component CPU --strip
        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    }
+}

+function buildCUDA11() {
+    # CUDA v11 claims to be compatible with MSVC 2022, but the latest updates are no longer compatible
+    # 19.40 is the last compiler version that works, but recent udpates are 19.43
+    # So this pins to MSVC 2019 for best compatibility
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
        $hashEnv = @{}
        Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
-        if ("$script:CUDA_DIRS".Contains("v12")) {
-            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V12")) { $v12="$_" }}
-            $env:CUDAToolkit_ROOT=$hashEnv[$v12]
-            write-host "Building CUDA v12 backend libraries"
-            & cmake --fresh --preset "CUDA 12" --install-prefix $script:DIST_DIR
+        if ("$script:CUDA_DIRS".Contains("v11")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V11")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            write-host "Building CUDA v11 backend libraries $cuda"
+            $env:CUDAToolkit_ROOT=$cuda
+            & cmake --fresh --preset "CUDA 11" -T cuda="$cuda" -DCMAKE_CUDA_COMPILER="$cuda\bin\nvcc.exe" -G "Visual Studio 16 2019" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v11"
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --build --preset "CUDA 11"  --config Release --parallel $script:JOBS
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --install build --component "CUDA" --strip
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
+    }
+}
+
+function buildCUDA12() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
+        $hashEnv = @{}
+        Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
+        if ("$script:CUDA_DIRS".Contains("v12.8")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V12_8")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            write-host "Building CUDA v12 backend libraries $cuda"
+            $env:CUDAToolkit_ROOT=$cuda
+            & cmake --fresh --preset "CUDA 12" -T cuda="$cuda" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v12"
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
            & cmake --build --preset "CUDA 12"  --config Release --parallel $script:JOBS
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
            & cmake --install build --component "CUDA" --strip
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
        }
+    }
+}
+
+function buildCUDA13() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
+        $hashEnv = @{}
+        Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
+        if ("$script:CUDA_DIRS".Contains("v13")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V13")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            $env:CUDAToolkit_ROOT=$cuda
+            write-host "Building CUDA v13 backend libraries $cuda"
+            & cmake --fresh --preset "CUDA 13" -T cuda="$cuda" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v13"
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --build --preset "CUDA 13"  --config Release --parallel $script:JOBS
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --install build --component "CUDA" --strip
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
+    }
+}
+
+function buildROCm() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
        if ($env:HIP_PATH) {
            write-host "Building ROCm backend libraries"
            if (-Not (get-command -ErrorAction silent ninja)) {
@@ -129,6 +181,10 @@ function buildOllama() {
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
        }
    }
+}
+
+function buildOllama() {
+    mkdir -Force -path "${script:DIST_DIR}\"
    write-host "Building ollama CLI"
    & go build -trimpath -ldflags "-s -w -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" .
    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
@@ -236,6 +292,10 @@ function distZip() {
 checkEnv
 try {
    if ($($args.count) -eq 0) {
+        buildCPU
+        buildCUDA12
+        buildCUDA13
+        buildROCm
        buildOllama
        buildApp
        gatherDependencies