matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-10 07:46:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

Optimize container images for startup (#6547)

Browse Source 
* Optimize container images for startup

This change adjusts how to handle runner payloads to support
container builds where we keep them extracted in the filesystem.
This makes it easier to optimize the cpu/cuda vs cpu/rocm images for
size, and should result in faster startup times for container images.

* Refactor payload logic and add buildx support for faster builds

* Move payloads around

* Review comments

* Converge to buildx based helper scripts

* Use docker buildx action for release
This commit is contained in:
Daniel Hiltgen
2024-09-12 12:10:30 -07:00
committed by GitHub
parent fef257c5c5
commit cd5c8f6471
32 changed files with 861 additions and 689 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
llm/generate/gen_common.sh
View File

@@ -31,6 +31,7 @@ init_vars() {
NO_WHOLE_ARCHIVE=""
GCC_ARCH="-arch ${ARCH}"
DIST_BASE=../../dist/darwin-${GOARCH}/
PAYLOAD_BASE=../../build/darwin/${GOARCH}
;;
"Linux")
LIB_EXT="so"
@@ -40,6 +41,7 @@ init_vars() {
# Cross compiling not supported on linux - Use docker
GCC_ARCH=""
DIST_BASE=../../dist/linux-${GOARCH}/
PAYLOAD_BASE=../../build/linux/${GOARCH}
;;
*)
;;
@@ -47,7 +49,8 @@ init_vars() {
if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
fi
GZIP=$(which pigz 2>/dev/null || echo "gzip")
GZIP=$(command -v pigz 2>/dev/null || echo "gzip")
RUNNER_BASE="${DIST_BASE}/lib/ollama/runners"
}
git_module_setup() {
@@ -91,17 +94,34 @@ build() {
rm -f ${BUILD_DIR}/bin/ggml-common.h ${BUILD_DIR}/bin/ggml-metal.metal
}
compress() {
echo "Compressing payloads to reduce overall binary size..."
rm -rf ${BUILD_DIR}/bin/*.gz
dist() {
[ -z "${RUNNER}" ] && exit 1
mkdir -p ${RUNNER_BASE}/${RUNNER}/
for f in ${BUILD_DIR}/bin/* ; do
${GZIP} -n --best -f ${f} &
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} -n --best -f ${f} &
${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
compress_pids+=" $!"
done
fi

12
llm/generate/gen_darwin.sh
View File

@@ -39,7 +39,8 @@ case "${GOARCH}" in
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu"
RUNNER=cpu
BUILD_DIR="../build/darwin/${ARCH}/${RUNNER}"
echo "Building LCD CPU"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
@@ -51,7 +52,8 @@ case "${GOARCH}" in
#
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 ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
RUNNER=cpu_avx
BUILD_DIR="../build/darwin/${ARCH}/${RUNNER}"
echo "Building AVX CPU"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
@@ -63,7 +65,8 @@ case "${GOARCH}" in
#
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 ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
RUNNER=cpu_avx2
BUILD_DIR="../build/darwin/${ARCH}/${RUNNER}"
echo "Building AVX2 CPU"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
build
@@ -84,7 +87,8 @@ case "${GOARCH}" in
if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
init_vars
CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/metal"
RUNNER="metal"
BUILD_DIR="../build/darwin/${ARCH}/${RUNNER}"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
build
sign ${BUILD_DIR}/bin/ollama_llama_server

35
llm/generate/gen_linux.sh
View File

@@ -79,10 +79,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
init_vars
echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu"
RUNNER="cpu"
BUILD_DIR="../build/linux/${ARCH}/${RUNNER}"
echo "Building custom CPU"
build
install
dist
compress
else
# Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
@@ -102,10 +104,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu"
RUNNER=cpu
BUILD_DIR="../build/linux/${ARCH}/${RUNNER}"
echo "Building LCD CPU"
build
install
dist
compress
fi
@@ -120,10 +124,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
RUNNER=cpu_avx
BUILD_DIR="../build/linux/${ARCH}/${RUNNER}"
echo "Building AVX CPU"
build
install
dist
compress
fi
@@ -134,10 +140,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
RUNNER=cpu_avx2
BUILD_DIR="../build/linux/${ARCH}/${RUNNER}"
echo "Building AVX2 CPU"
build
install
dist
compress
fi
fi
@@ -187,11 +195,13 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
fi
export CUDAFLAGS="-t8"
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS} -DGGML_STATIC=off"
BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
RUNNER=cuda${CUDA_VARIANT}
BUILD_DIR="../build/linux/${ARCH}/${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* ${CUDA_LIB_DIR}/libcublasLt.so* ; do
@@ -212,7 +222,8 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
CC=icx
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
BUILD_DIR="../build/linux/${ARCH}/oneapi"
RUNNER=oneapi
BUILD_DIR="../build/linux/${ARCH}/${RUNNER}"
ONEAPI_DIST_DIR="${DIST_BASE}/lib/ollama"
export LLAMA_SERVER_LDFLAGS="-fsycl -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
@@ -231,6 +242,7 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${ONEAPI_DIST_DIR}"
cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${ONEAPI_DIST_DIR}"
install
dist
compress
fi
@@ -259,7 +271,8 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
echo "Building custom ROCM GPU"
fi
BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
RUNNER=rocm${ROCM_VARIANT}
BUILD_DIR="../build/linux/${ARCH}/${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)
@@ -269,13 +282,17 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
# 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
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 -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 ${BUILD_DIR}/..; echo *)"
echo "go generate completed. LLM runners: $(cd ${PAYLOAD_BASE}; echo *)"

4
llm/llm_darwin_arm64.go → llm/llm_darwin.go
View File

@@ -1,11 +1,7 @@
package llm
import (
"embed"
"syscall"
)
//go:embed build/darwin/arm64/*/bin/*
var libEmbed embed.FS
var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

11
llm/llm_darwin_amd64.go
View File

@@ -1,11 +0,0 @@
package llm
import (
"embed"
"syscall"
)
//go:embed build/darwin/x86_64/*/bin/*
var libEmbed embed.FS
var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

4
llm/llm_linux.go
View File

@@ -1,11 +1,7 @@
package llm
import (
"embed"
"syscall"
)
//go:embed build/linux/*/*/bin/*
var libEmbed embed.FS
var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

4
llm/llm_windows.go
View File

@@ -1,13 +1,9 @@
package llm
import (
"embed"
"syscall"
)
// unused on windows
var libEmbed embed.FS
const CREATE_DEFAULT_ERROR_MODE = 0x04000000
var LlamaServerSysProcAttr = &syscall.SysProcAttr{

233
llm/payload.go
View File

@@ -1,233 +0,0 @@
package llm
import (
"compress/gzip"
"errors"
"fmt"
"io"
"io/fs"
"log/slog"
"os"
"path/filepath"
"runtime"
"slices"
"strings"
"golang.org/x/sync/errgroup"
"github.com/ollama/ollama/gpu"
)
var errPayloadMissing = errors.New("expected payloads not included in this build of ollama")
func Init() error {
payloadsDir, err := gpu.PayloadsDir()
if err != nil {
return err
}
if runtime.GOOS != "windows" {
slog.Info("extracting embedded files", "dir", payloadsDir)
binGlob := "build/*/*/*/bin/*"
// extract server libraries
err = extractFiles(payloadsDir, binGlob)
if err != nil {
return fmt.Errorf("extract binaries: %v", err)
}
}
var variants []string
for v := range getAvailableServers() {
variants = append(variants, v)
}
slog.Info(fmt.Sprintf("Dynamic LLM libraries %v", variants))
slog.Debug("Override detection logic by setting OLLAMA_LLM_LIBRARY")
return nil
}
// binary names may contain an optional variant separated by '_'
// For example, "ollama_rocm_v6" and "ollama_rocm_v5" or "ollama_cpu" and "ollama_cpu_avx2"
// Any library without a variant is the lowest common denominator
func getAvailableServers() map[string]string {
payloadsDir, err := gpu.PayloadsDir()
if err != nil {
slog.Error("payload lookup error", "error", err)
return nil
}
// glob payloadsDir for files that start with ollama_
pattern := filepath.Join(payloadsDir, "*", "ollama_*")
files, err := filepath.Glob(pattern)
if err != nil {
slog.Debug("could not glob", "pattern", pattern, "error", err)
return nil
}
servers := make(map[string]string)
for _, file := range files {
slog.Debug("availableServers : found", "file", file)
servers[filepath.Base(filepath.Dir(file))] = filepath.Dir(file)
}
return servers
}
// serversForGpu returns a list of compatible servers give the provided GPU
// info, ordered by performance. assumes Init() has been called
// TODO - switch to metadata based mapping
func serversForGpu(info gpu.GpuInfo) []string {
// glob workDir for files that start with ollama_
availableServers := getAvailableServers()
requested := info.Library
if info.Variant != gpu.CPUCapabilityNone.String() {
requested += "_" + info.Variant
}
servers := []string{}
// exact match first
for a := range availableServers {
if a == requested {
servers = []string{a}
if a == "metal" {
return servers
}
break
}
}
alt := []string{}
// Then for GPUs load alternates and sort the list for consistent load ordering
if info.Library != "cpu" {
for a := range availableServers {
if info.Library == strings.Split(a, "_")[0] && a != requested {
alt = append(alt, a)
}
}
slices.Sort(alt)
servers = append(servers, alt...)
}
if !(runtime.GOOS == "darwin" && runtime.GOARCH == "arm64") {
// Load up the best CPU variant if not primary requested
if info.Library != "cpu" {
variant := gpu.GetCPUCapability()
// If no variant, then we fall back to default
// If we have a variant, try that if we find an exact match
// Attempting to run the wrong CPU instructions will panic the
// process
if variant != gpu.CPUCapabilityNone {
for cmp := range availableServers {
if cmp == "cpu_"+variant.String() {
servers = append(servers, cmp)
break
}
}
} else {
servers = append(servers, "cpu")
}
}
if len(servers) == 0 {
servers = []string{"cpu"}
}
}
return servers
}
// Return the optimal server for this CPU architecture
func serverForCpu() string {
if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
return "metal"
}
variant := gpu.GetCPUCapability()
availableServers := getAvailableServers()
if variant != gpu.CPUCapabilityNone {
for cmp := range availableServers {
if cmp == "cpu_"+variant.String() {
return cmp
}
}
}
return "cpu"
}
// extract extracts the embedded files to the target directory
func extractFiles(targetDir string, glob string) error {
files, err := fs.Glob(libEmbed, glob)
if err != nil || len(files) == 0 {
return errPayloadMissing
}
if err := os.MkdirAll(targetDir, 0o755); err != nil {
return fmt.Errorf("extractFiles could not mkdir %s: %v", targetDir, err)
}
g := new(errgroup.Group)
// build/$OS/$GOARCH/$VARIANT/{bin,lib}/$FILE
for _, file := range files {
filename := file
variant := filepath.Base(filepath.Dir(filepath.Dir(filename)))
slog.Debug("extracting", "variant", variant, "file", filename)
g.Go(func() error {
srcf, err := libEmbed.Open(filename)
if err != nil {
return err
}
defer srcf.Close()
src := io.Reader(srcf)
if strings.HasSuffix(filename, ".gz") {
src, err = gzip.NewReader(src)
if err != nil {
return fmt.Errorf("decompress payload %s: %v", filename, err)
}
filename = strings.TrimSuffix(filename, ".gz")
}
variantDir := filepath.Join(targetDir, variant)
if err := os.MkdirAll(variantDir, 0o755); err != nil {
return fmt.Errorf("extractFiles could not mkdir %s: %v", variantDir, err)
}
base := filepath.Base(filename)
destFilename := filepath.Join(variantDir, base)
_, err = os.Stat(destFilename)
switch {
case errors.Is(err, os.ErrNotExist):
destFile, err := os.OpenFile(destFilename, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
if err != nil {
return fmt.Errorf("write payload %s: %v", filename, err)
}
defer destFile.Close()
if _, err := io.Copy(destFile, src); err != nil {
return fmt.Errorf("copy payload %s: %v", filename, err)
}
case err != nil:
return fmt.Errorf("stat payload %s: %v", filename, err)
}
return nil
})
}
err = g.Wait()
if err != nil {
// If we fail to extract, the payload dir is unusable, so cleanup whatever we extracted
gpu.Cleanup()
return err
}
return nil
}

29
llm/server.go
View File

@@ -24,9 +24,11 @@ import (
"golang.org/x/sync/semaphore"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/build"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format"
"github.com/ollama/ollama/gpu"
"github.com/ollama/ollama/runners"
)
type LlamaServer interface {
@@ -106,7 +108,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
gpus = gpu.GetCPUInfo()
}
if len(gpus) == 1 && gpus[0].Library == "cpu" {
cpuRunner = serverForCpu()
cpuRunner = runners.ServerForCpu()
estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
} else {
estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
@@ -118,7 +120,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
opts.NumGPU = 0
case gpus[0].Library != "metal" && estimate.Layers == 0:
// Don't bother loading into the GPU if no layers can fit
cpuRunner = serverForCpu()
cpuRunner = runners.ServerForCpu()
gpus = gpu.GetCPUInfo()
case opts.NumGPU < 0 && estimate.Layers > 0 && gpus[0].Library != "cpu":
opts.NumGPU = estimate.Layers
@@ -145,25 +147,20 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
}
availableServers := getAvailableServers()
rDir, err := runners.Refresh(build.EmbedFS)
if err != nil {
return nil, err
}
availableServers := runners.GetAvailableServers(rDir)
if len(availableServers) == 0 {
if runtime.GOOS != "windows" {
slog.Warn("llama server binary disappeared, reinitializing payloads")
err = Init()
if err != nil {
slog.Warn("failed to reinitialize payloads", "error", err)
return nil, err
}
availableServers = getAvailableServers()
} else {
return nil, finalErr
}
return nil, finalErr
}
var servers []string
if cpuRunner != "" {
servers = []string{cpuRunner}
} else {
servers = serversForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
servers = runners.ServersForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
}
demandLib := envconfig.LLMLibrary()
if demandLib != "" {
@@ -330,7 +327,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
_, err := os.Stat(server)
if errors.Is(err, os.ErrNotExist) {
slog.Warn("llama server disappeared, reinitializing payloads", "path", server, "error", err)
err = Init()
_, err = runners.Refresh(build.EmbedFS)
if err != nil {
slog.Warn("failed to reinitialize payloads", "error", err)
return nil, err