Skip to content

GitLab

Unverified Commit 652c273f authored by Daniel Hiltgen's avatar Daniel Hiltgen Committed by GitHub
Browse files

Merge pull request #5049 from dhiltgen/cuda_v12 

Cuda v12
parents 88e77050 f9e31da9
Show whitespace changes
Inline Side-by-side
Showing with 346 additions and 190 deletions
.github/workflows/release.yaml
View file @ 652c273f
......@@ -187,6 +187,13 @@ jobs:
generate-windows-cuda:
environment: release
runs-on: windows
strategy:
matrix:
cuda:
- version: "11"
url: 'https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe'
- version: "12"
url: 'https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe'
env:
KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
steps:
......@@ -220,11 +227,11 @@ jobs:
with:
go-version-file: go.mod
cache: true
- name: 'Install CUDA'
- name: 'Install CUDA ${{ matrix.cuda.version }}'
run: |
$ErrorActionPreference = "Stop"
write-host "downloading CUDA Installer"
Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
Invoke-WebRequest -Uri "${{ matrix.cuda.url }}" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
write-host "Installing CUDA"
Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
write-host "Completed CUDA"
......@@ -256,7 +263,7 @@ jobs:
cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
- uses: actions/upload-artifact@v4
with:
name: generate-windows-cuda
name: generate-windows-cuda-${{ matrix.cuda.version }}
path: |
llm/build/**/bin/*
dist/windows-amd64/**
......@@ -265,6 +272,7 @@ jobs:
name: windows-cuda-deps
path: dist/deps/*
# Import the prior generation steps and build the final windows assets
build-windows:
environment: release
......@@ -314,7 +322,10 @@ jobs:
name: generate-windows-cpu
- uses: actions/download-artifact@v4
with:
name: generate-windows-cuda
name: generate-windows-cuda-11
- uses: actions/download-artifact@v4
with:
name: generate-windows-cuda-12
- uses: actions/download-artifact@v4
with:
name: windows-cuda-deps
......@@ -363,7 +374,6 @@ jobs:
- run: |
./scripts/build_linux.sh
./scripts/build_docker.sh
mv dist/deps/* dist/
- uses: actions/upload-artifact@v4
with:
name: dist-linux-amd64
......
Dockerfile
View file @ 652c273f
ARG GOLANG_VERSION=1.22.5
ARG CMAKE_VERSION=3.22.1
# this CUDA_VERSION corresponds with the one specified in docs/gpu.md
ARG CUDA_VERSION=11.3.1
ARG CUDA_VERSION_11=11.3.1
ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
ARG CUDA_VERSION_12=12.4.0
ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
ARG ROCM_VERSION=6.1.2
# Copy the minimal context we need to run the generate scripts
......@@ -10,7 +12,7 @@ COPY .git .git
COPY .gitmodules .gitmodules
COPY llm llm
FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION-devel-centos7 AS cuda-build-amd64
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
......@@ -18,9 +20,34 @@ ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
COPY --from=llm-code / /go/src/github.com/ollama/ollama/
WORKDIR /go/src/github.com/ollama/ollama/llm/generate
ARG CGO_CFLAGS
RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION-devel-rockylinux8 AS cuda-build-arm64
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-server-arm64
ARG CMAKE_VERSION
COPY ./scripts/rh_linux_deps.sh /
RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
......@@ -28,7 +55,32 @@ ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
COPY --from=llm-code / /go/src/github.com/ollama/ollama/
WORKDIR /go/src/github.com/ollama/ollama/llm/generate
ARG CGO_CFLAGS
RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
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-server-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
......@@ -40,15 +92,11 @@ 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
RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
RUN mkdir /tmp/scratch && \
for dep in $(zcat /go/src/github.com/ollama/ollama/llm/build/linux/x86_64/rocm*/bin/deps.txt.gz) ; do \
cp ${dep} /tmp/scratch/ || exit 1 ; \
done && \
(cd /opt/rocm/lib && tar cf - rocblas/library) | (cd /tmp/scratch/ && tar xf - ) && \
mkdir -p /go/src/github.com/ollama/ollama/dist/deps/ && \
(cd /tmp/scratch/ && tar czvf /go/src/github.com/ollama/ollama/dist/deps/ollama-linux-amd64-rocm.tgz . )
ENV GOARCH amd64
RUN --mount=type=cache,target=/root/.ccache \
OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
RUN mkdir -p ../../dist/linux-amd64/lib/ollama && \
(cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64/lib/ollama && tar xf - )
FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
ARG CMAKE_VERSION
......@@ -59,16 +107,21 @@ ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$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
FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
RUN --mount=type=cache,target=/root/.ccache \
OLLAMA_CPU_TARGET="static" bash gen_linux.sh
FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
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 OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
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 OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
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
ARG CMAKE_VERSION
......@@ -79,12 +132,15 @@ 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
FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
RUN --mount=type=cache,target=/root/.ccache \
OLLAMA_CPU_TARGET="static" bash gen_linux.sh
FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
RUN --mount=type=cache,target=/root/.ccache \
OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
# Intermediate stage used for ./scripts/build_linux.sh
......@@ -95,12 +151,16 @@ COPY . .
COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
COPY --from=cuda-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
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/llm/build/linux/ llm/build/linux/
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/llm/build/linux/ llm/build/linux/
COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/deps/ ./dist/deps/
ARG GOFLAGS
ARG CGO_CFLAGS
RUN go build -trimpath .
RUN --mount=type=cache,target=/root/.ccache \
go build -trimpath -o dist/linux-amd64/bin/ollama .
# Intermediate stage used for ./scripts/build_linux.sh
FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
......@@ -109,23 +169,36 @@ ARG GOLANG_VERSION
WORKDIR /go/src/github.com/ollama/ollama
COPY . .
COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
COPY --from=cuda-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
COPY --from=cuda-11-build-server-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
COPY --from=cuda-11-build-server-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
COPY --from=cuda-12-build-server-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
COPY --from=cuda-12-build-server-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
ARG GOFLAGS
ARG CGO_CFLAGS
RUN go build -trimpath .
RUN --mount=type=cache,target=/root/.ccache \
go build -trimpath -o dist/linux-arm64/bin/ollama .
# Strip out ROCm dependencies to keep the primary image lean
FROM --platform=linux/amd64 ubuntu:22.04 as amd64-libs-without-rocm
COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /scratch/
RUN cd /scratch/ollama/ && rm -rf rocblas libamd* libdrm* libroc* libhip* libhsa*
# Runtime stages
FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
COPY --from=amd64-libs-without-rocm /scratch/ /lib/
RUN apt-get update && apt-get install -y ca-certificates
COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
FROM --platform=linux/arm64 ubuntu:22.04 as runtime-arm64
COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
RUN apt-get update && apt-get install -y ca-certificates
COPY --from=build-arm64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
# Radeon images are much larger so we keep it distinct from the CPU/CUDA image
FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete as runtime-rocm
RUN update-pciids
COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
RUN ln -s /opt/rocm/lib /lib/ollama
EXPOSE 11434
ENV OLLAMA_HOST 0.0.0.0
......
app/ollama.iss
View file @ 652c273f
......@@ -87,20 +87,11 @@ DialogFontSize=12
[Files]
Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
Source: "..\dist\windows-{#ARCH}\ollama_runners\*"; DestDir: "{app}\ollama_runners"; Flags: ignoreversion 64bit recursesubdirs
Source: "..\ollama.exe"; DestDir: "{app}\bin"; Flags: ignoreversion 64bit
Source: "..\dist\windows-{#ARCH}\lib\ollama\runners\*"; DestDir: "{app}\lib\ollama\runners"; Flags: ignoreversion 64bit recursesubdirs
Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
#if DirExists("..\dist\windows-amd64\cuda")
Source: "..\dist\windows-amd64\cuda\*"; DestDir: "{app}\cuda\"; Flags: ignoreversion recursesubdirs
#endif
#if DirExists("..\dist\windows-amd64\oneapi")
Source: "..\dist\windows-amd64\oneapi\*"; DestDir: "{app}\oneapi\"; Flags: ignoreversion recursesubdirs
#endif
#if DirExists("..\dist\windows-amd64\rocm")
Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
#endif
Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Flags: ignoreversion recursesubdirs
[Icons]
Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
......@@ -108,7 +99,7 @@ Name: "{userstartup}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilen
Name: "{userprograms}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
[Run]
Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
Filename: "{cmd}"; Parameters: "/C set PATH={app}\bin;%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
[UninstallRun]
; Filename: "{cmd}"; Parameters: "/C ""taskkill /im ''{#MyAppExeName}'' /f /t"; Flags: runhidden
......@@ -143,8 +134,8 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi
[Registry]
Root: HKCU; Subkey: "Environment"; \
ValueType: expandsz; ValueName: "Path"; ValueData: "{olddata};{app}"; \
Check: NeedsAddPath('{app}')
ValueType: expandsz; ValueName: "Path"; ValueData: "{olddata};{app}\bin"; \
Check: NeedsAddPath('{app}\bin')
[Code]
......
docs/linux.md
View file @ 652c273f
......@@ -20,13 +20,12 @@ GPU.
## Manual install
### Download the `ollama` binary
### Download `ollama`
Ollama is distributed as a self-contained binary. Download it to a directory in your PATH:
Download and extract the Linux package:
```bash
sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
sudo chmod +x /usr/bin/ollama
curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
```
### Adding Ollama as a startup service (recommended)
......@@ -96,8 +95,7 @@ curl -fsSL https://ollama.com/install.sh | sh
Or by downloading the ollama binary:
```bash
sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
sudo chmod +x /usr/bin/ollama
curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
```
## Installing specific versions
......
envconfig/config.go
View file @ 652c273f
......@@ -174,7 +174,7 @@ func RunnersDir() (p string) {
defer func() {
if p == "" {
slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama_runners'")
slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama/runners'")
}
}()
......@@ -190,17 +190,17 @@ func RunnersDir() (p string) {
}
var paths []string
for _, root := range []string{filepath.Dir(exe), cwd} {
for _, root := range []string{filepath.Dir(exe), filepath.Join(filepath.Dir(exe), ".."), cwd} {
paths = append(paths,
root,
filepath.Join(root, "windows-"+runtime.GOARCH),
filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH),
filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH),
)
}
// Try a few variations to improve developer experience when building from source in the local tree
for _, path := range paths {
candidate := filepath.Join(path, "ollama_runners")
candidate := filepath.Join(path, "lib", "ollama", "runners")
if _, err := os.Stat(candidate); err == nil {
p = candidate
break
......
gpu/amd_common.go
View file @ 652c273f
......@@ -54,7 +54,7 @@ func commonAMDValidateLibDir() (string, error) {
// Installer payload location if we're running the installed binary
exe, err := os.Executable()
if err == nil {
rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
rocmTargetDir := filepath.Join(filepath.Dir(exe), "..", "lib", "ollama")
if rocmLibUsable(rocmTargetDir) {
slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
return rocmTargetDir, nil
......
gpu/amd_windows.go
View file @ 652c273f
......@@ -153,7 +153,7 @@ func AMDValidateLibDir() (string, error) {
// Installer payload (if we're running from some other location)
localAppData := os.Getenv("LOCALAPPDATA")
appDir := filepath.Join(localAppData, "Programs", "Ollama")
rocmTargetDir := filepath.Join(appDir, "rocm")
rocmTargetDir := filepath.Join(appDir, "..", "lib", "ollama")
if rocmLibUsable(rocmTargetDir) {
slog.Debug("detected ollama installed ROCm at " + rocmTargetDir)
return rocmTargetDir, nil
......
gpu/cuda_common.go
View file @ 652c273f
......@@ -4,9 +4,17 @@ package gpu
import (
"log/slog"
"os"
"regexp"
"runtime"
"strconv"
"strings"
)
// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
var CudaTegra string = os.Getenv("JETSON_JETPACK")
func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
ids := []string{}
for _, info := range gpuInfo {
......@@ -19,3 +27,38 @@ func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
}
return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
}
func cudaVariant(gpuInfo CudaGPUInfo) string {
if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
if CudaTegra != "" {
ver := strings.Split(CudaTegra, ".")
if len(ver) > 0 {
return "jetpack" + ver[0]
}
} else if data, err := os.ReadFile("/etc/nv_tegra_release"); err == nil {
r := regexp.MustCompile(` R(\d+) `)
m := r.FindSubmatch(data)
if len(m) != 2 {
slog.Info("Unexpected format for /etc/nv_tegra_release. Set JETSON_JETPACK to select version")
} else {
if l4t, err := strconv.Atoi(string(m[1])); err == nil {
// Note: mapping from L4t -> JP is inconsistent (can't just subtract 30)
// https://developer.nvidia.com/embedded/jetpack-archive
switch l4t {
case 35:
return "jetpack5"
case 36:
return "jetpack6"
default:
slog.Info("unsupported L4T version", "nv_tegra_release", string(data))
}
}
}
}
}
if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 {
return "v11"
}
return "v12"
}
gpu/gpu.go
View file @ 652c273f
......@@ -64,10 +64,6 @@ var RocmComputeMin = 9
// TODO find a better way to detect iGPU instead of minimum memory
const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
var CudaTegra string = os.Getenv("JETSON_JETPACK")
// Note: gpuMutex must already be held
func initCudaHandles() *cudaHandles {
// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
......@@ -215,7 +211,7 @@ func GetGPUInfo() GpuInfoList {
GpuInfo: GpuInfo{
memInfo: mem,
Library: "cpu",
Variant: cpuCapability,
Variant: cpuCapability.String(),
ID: "0",
},
},
......@@ -229,11 +225,7 @@ func GetGPUInfo() GpuInfoList {
return GpuInfoList{cpus[0].GpuInfo}
}
// On windows we bundle the nvidia library one level above the runner dir
depPath := ""
if runtime.GOOS == "windows" && envconfig.RunnersDir() != "" {
depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir()), "cuda")
}
depPath := LibraryDir()
// Load ALL libraries
cHandles = initCudaHandles()
......@@ -269,11 +261,23 @@ func GetGPUInfo() GpuInfoList {
gpuInfo.FreeMemory = uint64(memInfo.free)
gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
gpuInfo.computeMajor = int(memInfo.major)
gpuInfo.computeMinor = int(memInfo.minor)
gpuInfo.MinimumMemory = cudaMinimumMemory
variant := cudaVariant(gpuInfo)
if depPath != "" {
gpuInfo.DependencyPath = depPath
// Check for variant specific directory
if variant != "" {
if _, err := os.Stat(filepath.Join(depPath, "cuda_"+variant)); err == nil {
gpuInfo.DependencyPath = filepath.Join(depPath, "cuda_"+variant)
}
}
}
gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
gpuInfo.DriverMajor = driverMajor
gpuInfo.DriverMinor = driverMinor
gpuInfo.Variant = variant
// query the management library as well so we can record any skew between the two
// which represents overhead on the GPU we must set aside on subsequent updates
......@@ -306,13 +310,6 @@ func GetGPUInfo() GpuInfoList {
if envconfig.IntelGPU() {
oHandles = initOneAPIHandles()
if oHandles != nil && oHandles.oneapi != nil {
// On windows we bundle the oneapi library one level above the runner dir
depPath = ""
if runtime.GOOS == "windows" && envconfig.RunnersDir() != "" {
depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir()), "oneapi")
}
for d := range oHandles.oneapi.num_drivers {
if oHandles.oneapi == nil {
// shouldn't happen
......@@ -467,10 +464,12 @@ func GetGPUInfo() GpuInfoList {
func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
var ldPaths []string
var patterns []string
gpuLibPaths := []string{}
slog.Debug("Searching for GPU library", "name", baseLibName)
// Start with our bundled libraries
patterns := []string{filepath.Join(LibraryDir(), baseLibName)}
switch runtime.GOOS {
case "windows":
ldPaths = strings.Split(os.Getenv("PATH"), ";")
......@@ -479,13 +478,14 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
default:
return gpuLibPaths
}
// Start with whatever we find in the PATH/LD_LIBRARY_PATH
// Then with whatever we find in the PATH/LD_LIBRARY_PATH
for _, ldPath := range ldPaths {
d, err := filepath.Abs(ldPath)
if err != nil {
continue
}
patterns = append(patterns, filepath.Join(d, baseLibName+"*"))
patterns = append(patterns, filepath.Join(d, baseLibName))
}
patterns = append(patterns, defaultPatterns...)
slog.Debug("gpu library search", "globs", patterns)
......@@ -641,3 +641,31 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
return "", ""
}
}
func LibraryDir() string {
// On Windows/linux we bundle the dependencies at the same level as the executable
appExe, err := os.Executable()
if err != nil {
slog.Warn("failed to lookup executable path", "error", err)
}
cwd, err := os.Getwd()
if err != nil {
slog.Warn("failed to lookup working directory", "error", err)
}
// Scan for any of our dependeices, and pick first match
for _, root := range []string{filepath.Dir(appExe), filepath.Join(filepath.Dir(appExe), ".."), cwd} {
libDep := filepath.Join("lib", "ollama")
if _, err := os.Stat(filepath.Join(root, libDep)); err == nil {
return filepath.Join(root, libDep)
}
// Developer mode, local build
if _, err := os.Stat(filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
return filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)
}
if _, err := os.Stat(filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
return filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)
}
}
slog.Warn("unable to locate gpu dependency libraries")
return ""
}
gpu/gpu_darwin.go
View file @ 652c273f
......@@ -25,7 +25,7 @@ func GetGPUInfo() GpuInfoList {
return []GpuInfo{
{
Library: "cpu",
Variant: GetCPUCapability(),
Variant: GetCPUCapability().String(),
memInfo: mem,
},
}
......@@ -48,7 +48,7 @@ func GetCPUInfo() GpuInfoList {
return []GpuInfo{
{
Library: "cpu",
Variant: GetCPUCapability(),
Variant: GetCPUCapability().String(),
memInfo: mem,
},
}
......
gpu/gpu_linux.go
View file @ 652c273f
......@@ -47,7 +47,7 @@ var (
CudartMgmtName = "libcudart.so*"
NvcudaMgmtName = "libcuda.so*"
NvmlMgmtName = "" // not currently wired on linux
OneapiMgmtName = "libze_intel_gpu.so"
OneapiMgmtName = "libze_intel_gpu.so*"
)
func GetCPUMem() (memInfo, error) {
......
gpu/types.go
View file @ 652c273f
......@@ -19,7 +19,7 @@ type GpuInfo struct {
Library string `json:"library,omitempty"`
// Optional variant to select (e.g. versions, cpu feature flags)
Variant CPUCapability `json:"variant"`
Variant string `json:"variant"`
// MinimumMemory represents the minimum memory required to use the GPU
MinimumMemory uint64 `json:"-"`
......@@ -55,6 +55,8 @@ type CudaGPUInfo struct {
GpuInfo
OSOverhead uint64 // Memory overhead between the driver library and management library
index int //nolint:unused,nolintlint
computeMajor int //nolint:unused,nolintlint
computeMinor int //nolint:unused,nolintlint
}
type CudaGPUInfoList []CudaGPUInfo
......@@ -81,8 +83,8 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
for _, info := range l {
found := false
requested := info.Library
if info.Variant != CPUCapabilityNone {
requested += "_" + info.Variant.String()
if info.Variant != CPUCapabilityNone.String() {
requested += "_" + info.Variant
}
for i, lib := range libs {
if lib == requested {
......@@ -105,6 +107,7 @@ func (l GpuInfoList) LogDetails() {
slog.Info("inference compute",
"id", g.ID,
"library", g.Library,
"variant", g.Variant,
"compute", g.Compute,
"driver", fmt.Sprintf("%d.%d", g.DriverMajor, g.DriverMinor),
"name", g.Name,
......
llm/ext_server/CMakeLists.txt
View file @ 652c273f
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 json.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 ${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_SERVER_LDFLAGS})
if (WIN32)
TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
endif()
......
llm/generate/gen_common.sh
View file @ 652c273f
......@@ -9,11 +9,14 @@ init_vars() {
ARCH="arm64"
;;
*)
ARCH=$(uname -m | sed -e "s/aarch64/arm64/g")
echo "GOARCH must be set"
echo "this script is meant to be run from within go generate"
exit 1
;;
esac
LLAMACPP_DIR=../llama.cpp
CMAKE_DEFS=""
CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on"
CMAKE_TARGETS="--target ollama_llama_server"
if echo "${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}"
......@@ -27,6 +30,7 @@ init_vars() {
WHOLE_ARCHIVE="-Wl,-force_load"
NO_WHOLE_ARCHIVE=""
GCC_ARCH="-arch ${ARCH}"
DIST_BASE=../../dist/darwin-${GOARCH}/
;;
"Linux")
LIB_EXT="so"
......@@ -35,6 +39,7 @@ init_vars() {
# Cross compiling not supported on linux - Use docker
GCC_ARCH=""
DIST_BASE=../../dist/linux-${GOARCH}/
;;
*)
;;
......@@ -42,6 +47,7 @@ init_vars() {
if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
fi
GZIP=$(which pigz 2>/dev/null || echo "gzip")
}
git_module_setup() {
......@@ -85,26 +91,36 @@ build() {
compress() {
echo "Compressing payloads to reduce overall binary size..."
pids=""
rm -rf ${BUILD_DIR}/bin/*.gz
for f in ${BUILD_DIR}/bin/* ; do
gzip -n --best -f ${f} &
pids+=" $!"
${GZIP} -n --best -f ${f} &
compress_pids+=" $!"
done
# check for lib directory
if [ -d ${BUILD_DIR}/lib ]; then
for f in ${BUILD_DIR}/lib/* ; do
gzip -n --best -f ${f} &
pids+=" $!"
${GZIP} -n --best -f ${f} &
compress_pids+=" $!"
done
fi
echo
for pid in ${pids}; do
}
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}); 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() {
(cd ${LLAMACPP_DIR}/ && git checkout CMakeLists.txt)
......
llm/generate/gen_darwin.sh
View file @ 652c273f
......@@ -6,6 +6,7 @@
set -ex
set -o pipefail
compress_pids=""
echo "Starting darwin generate script"
source $(dirname $0)/gen_common.sh
init_vars
......@@ -98,4 +99,5 @@ case "${GOARCH}" in
esac
cleanup
wait_for_compress
echo "go generate completed. LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
llm/generate/gen_linux.sh
View file @ 652c273f
......@@ -13,6 +13,7 @@
set -ex
set -o pipefail
compress_pids=""
# See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
amdGPUs() {
......@@ -51,7 +52,7 @@ if [ -z "${CUDACXX}" ]; then
export CUDACXX=$(command -v nvcc)
fi
fi
COMMON_CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -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"
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 $0)/gen_common.sh
init_vars
git_module_setup
......@@ -77,10 +78,11 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
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=off -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu"
echo "Building custom CPU"
build
install
compress
else
# Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
......@@ -93,7 +95,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
# -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
# -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
#
# CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
......@@ -103,6 +105,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
BUILD_DIR="../build/linux/${ARCH}/cpu"
echo "Building LCD CPU"
build
install
compress
fi
......@@ -120,6 +123,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
echo "Building AVX CPU"
build
install
compress
fi
......@@ -133,6 +137,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
echo "Building AVX2 CPU"
build
install
compress
fi
fi
......@@ -160,7 +165,7 @@ if [ -z "${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}" ]; then
if [ -n "${CUDA_MAJOR}" -a -z "${CUDA_VARIANT}" ]; then
CUDA_VARIANT=_v${CUDA_MAJOR}
fi
if [ "${ARCH}" == "arm64" ]; then
......@@ -178,29 +183,19 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
echo "Building custom CUDA GPU"
else
CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
fi
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
export CUDAFLAGS="-t8"
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS} -DGGML_STATIC=off"
BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
export LLAMA_SERVER_LDFLAGS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
CUDA_DIST_DIR="${CUDA_DIST_DIR:-${DIST_BASE}/lib/ollama}"
build
# Carry the CUDA libs as payloads to help reduce dependency burden on users
#
# TODO - in the future we may shift to packaging these separately and conditionally
# downloading them in the install script.
DEPS="$(ldd ${BUILD_DIR}/bin/ollama_llama_server )"
for lib in libcudart.so libcublas.so libcublasLt.so ; do
DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/bin/"
elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/bin/"
elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/bin/"
else
cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/bin/"
fi
install
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}" "${CUDA_DIST_DIR}"
done
compress
......@@ -218,21 +213,24 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
CC=icx
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
BUILD_DIR="../build/linux/${ARCH}/oneapi"
EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
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 "${dep}" "${BUILD_DIR}/bin/"
cp -a "${dep}" "${ONEAPI_DIST_DIR}"
done
cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${BUILD_DIR}/bin/"
cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${BUILD_DIR}/bin/"
cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${BUILD_DIR}/bin/"
cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${BUILD_DIR}/bin/"
cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${BUILD_DIR}/bin/"
cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${BUILD_DIR}/bin/"
cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${BUILD_DIR}/bin/"
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
compress
fi
......@@ -262,23 +260,21 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
echo "Building custom ROCM GPU"
fi
BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
ROCM_DIST_DIR="${DIST_BASE}/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${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
build
# Record the ROCM dependencies
rm -f "${BUILD_DIR}/bin/deps.txt"
touch "${BUILD_DIR}/bin/deps.txt"
for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
# 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 "${ARCH}/rocm${ROCM_VARIANT}" | grep -e rocm -e amdgpu -e libtinfo ); do
cp -a "${dep}"* "${ROCM_DIST_DIR}"
done
# bomb out if for some reason we didn't get a few deps
if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
cat "${BUILD_DIR}/bin/deps.txt"
echo "ERROR: deps file short"
exit 1
fi
install
compress
fi
cleanup
wait_for_compress
echo "go generate completed. LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
llm/generate/gen_windows.ps1
View file @ 652c273f
......@@ -35,7 +35,7 @@ function init_vars {
)
$script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
$script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
$script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\ollama_runners"
$script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\runners"
md "$script:DIST_BASE" -ea 0 > $null
if ($env:CGO_CFLAGS -contains "-g") {
$script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
......@@ -117,7 +117,7 @@ function build {
if ($cmakeDefs -contains "-G") {
$extra=@("-j8")
} else {
$extra= @("--", "/p:CL_MPcount=8")
$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", $_ }) $extra
......@@ -261,7 +261,7 @@ function build_cuda() {
if ((-not "${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
$script:CUDA_VERSION=((get-item ($nvcc | split-path | split-path)).Basename -Split "\.")[0]
if ($null -ne $script:CUDA_VERSION) {
$script:CUDA_VARIANT="_"+$script:CUDA_VERSION
}
......@@ -273,9 +273,9 @@ function build_cuda() {
"-DGGML_CUDA=ON",
"-DGGML_AVX=on",
"-DGGML_AVX2=off",
"-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
"-DCMAKE_CUDA_FLAGS=-t8",
"-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
"-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}`""
......@@ -286,12 +286,11 @@ function build_cuda() {
sign
install
rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
md "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\" -ea 0 > $null
write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
md "${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-${script:ARCH}\lib\ollama\"
} else {
write-host "Skipping CUDA generation step"
}
......@@ -325,18 +324,17 @@ function build_oneapi() {
sign
install
rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
md "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\" -ea 0 > $null
cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
md "${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-${script:ARCH}\lib\ollama\"
} else {
Write-Host "Skipping oneAPI generation step"
}
......@@ -386,12 +384,11 @@ function build_rocm() {
sign
install
rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
md "${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}\rocm\rocblas\library\"
cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\"
} else {
write-host "Skipping ROCm generation step"
}
......
llm/payload.go
View file @ 652c273f
......@@ -82,8 +82,8 @@ func serversForGpu(info gpu.GpuInfo) []string {
// glob workDir for files that start with ollama_
availableServers := getAvailableServers()
requested := info.Library
if info.Variant != gpu.CPUCapabilityNone {
requested += "_" + info.Variant.String()
if info.Variant != gpu.CPUCapabilityNone.String() {
requested += "_" + info.Variant
}
servers := []string{}
......
llm/server.go
View file @ 652c273f
......@@ -306,20 +306,18 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
if runtime.GOOS == "windows" {
pathEnv = "PATH"
}
// prepend the server directory to LD_LIBRARY_PATH/PATH and the parent dir for common dependencies
libraryPaths := []string{dir, filepath.Dir(dir)}
// Start with the server directory for the LD_LIBRARY_PATH/PATH
libraryPaths := []string{dir}
if libraryPath, ok := os.LookupEnv(pathEnv); ok {
// Append our runner directory to the path
// This will favor system libraries over our bundled library dependencies
// favor our bundled library dependencies over system libraries
libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
}
// Note: we always put the dependency path first
// since this was the exact version we verified for AMD GPUs
// and we favor what the user had in their path
// since this was the exact version we compiled/linked against
if gpus[0].DependencyPath != "" {
// TODO refine for multi-gpu support
// assume gpus from the same library have the same dependency path
libraryPaths = append([]string{gpus[0].DependencyPath}, libraryPaths...)
}
......
scripts/build_linux.sh
View file @ 652c273f
......@@ -4,6 +4,7 @@ set -eu
export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
GZIP=$(which pigz 2>/dev/null || echo "gzip")
BUILD_ARCH=${BUILD_ARCH:-"amd64 arm64"}
export AMDGPU_TARGETS=${AMDGPU_TARGETS:=""}
......@@ -21,11 +22,10 @@ for TARGETARCH in ${BUILD_ARCH}; do
-t builder:$TARGETARCH \
.
docker create --platform linux/$TARGETARCH --name builder-$TARGETARCH builder:$TARGETARCH
docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/ollama ./dist/ollama-linux-$TARGETARCH
if [ "$TARGETARCH" = "amd64" ]; then
docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/dist/deps/ ./dist/
fi
rm -rf ./dist/linux-$TARGETARCH
docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/dist/linux-$TARGETARCH ./dist
docker rm builder-$TARGETARCH
echo "Compressing final linux bundle..."
rm -f ./dist/ollama-linux-$TARGETARCH.tgz
(cd dist/linux-$TARGETARCH && tar cf - . | ${GZIP} --best > ../ollama-linux-$TARGETARCH.tgz )
done
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment