mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 14:26:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

DRY out the runner lifecycle code (#12540)

Browse Source 
* DRY out the runner lifecycle code

Now that discovery uses the runners as well, this unifies the runner spawning code
into a single place.  This also unifies GPU discovery types with the newer ml.DeviceInfo

* win: make incremental builds better

Place build artifacts in discrete directories so incremental builds don't have to start fresh

* Adjust sort order to consider iGPUs

* handle cpu inference oom scenarios

* review comments
This commit is contained in:
Daniel Hiltgen
2025-10-23 11:20:02 -07:00
committed by GitHub
parent 1c093e97af
commit 3258a89b6e
16 changed files with 720 additions and 924 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
discover/cpu_linux_test.go
View File

@@ -2065,12 +2065,6 @@ power management:
cpus := linuxCPUDetails(buf)
slog.Info("example", "scenario", k, "cpus", cpus)
si := SystemInfo{
System: CPUInfo{
CPUs: cpus,
},
}
threadCount := si.GetOptimalThreadCount()
if len(v.expCPUs) != len(cpus) {
t.Fatalf("incorrect number of sockets: expected:%v got:%v", v.expCPUs, cpus)
}
@@ -2085,10 +2079,6 @@ power management:
t.Fatalf("incorrect number of threads: expected:%v got:%v", v.expCPUs[i], c)
}
}
if threadCount != v.expThreadCount {
t.Fatalf("incorrect thread count expected:%d got:%d", v.expThreadCount, threadCount)
}
})
}
}

164
discover/gpu.go
View File

@@ -1,16 +1,13 @@
package discover
import (
"context"
"log/slog"
"os"
"path/filepath"
"regexp"
"runtime"
"strconv"
"strings"
"github.com/ollama/ollama/format"
"github.com/ollama/ollama/ml"
)
@@ -18,159 +15,28 @@ import (
// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
var CudaTegra string = os.Getenv("JETSON_JETPACK")
func GetCPUInfo() GpuInfo {
mem, err := GetCPUMem()
// GetSystemInfo returns the last cached state of the GPUs on the system
func GetSystemInfo() ml.SystemInfo {
memInfo, err := GetCPUMem()
if err != nil {
slog.Warn("error looking up system memory", "error", err)
}
return GpuInfo{
memInfo: mem,
DeviceID: ml.DeviceID{
Library: "cpu",
ID: "0",
},
}
}
func GetGPUInfo(ctx context.Context, runners []FilteredRunnerDiscovery) GpuInfoList {
devs := GPUDevices(ctx, runners)
return devInfoToInfoList(devs)
}
func devInfoToInfoList(devs []ml.DeviceInfo) GpuInfoList {
resp := []GpuInfo{}
// Our current packaging model places ggml-hip in the main directory
// but keeps rocm in an isolated directory. We have to add it to
// the [LD_LIBRARY_]PATH so ggml-hip will load properly
rocmDir := filepath.Join(LibOllamaPath, "rocm")
if _, err := os.Stat(rocmDir); err != nil {
rocmDir = ""
var threadCount int
cpus := GetCPUDetails()
for _, c := range cpus {
threadCount += c.CoreCount - c.EfficiencyCoreCount
}
for _, dev := range devs {
info := GpuInfo{
DeviceID: dev.DeviceID,
filterID: dev.FilteredID,
Name: dev.Description,
memInfo: memInfo{
TotalMemory: dev.TotalMemory,
FreeMemory: dev.FreeMemory,
},
// TODO can we avoid variant
DependencyPath: dev.LibraryPath,
DriverMajor: dev.DriverMajor,
DriverMinor: dev.DriverMinor,
ComputeMajor: dev.ComputeMajor,
ComputeMinor: dev.ComputeMinor,
}
if dev.Library == "CUDA" || dev.Library == "ROCm" {
info.MinimumMemory = 457 * format.MebiByte
}
if dev.Library == "ROCm" && rocmDir != "" {
info.DependencyPath = append(info.DependencyPath, rocmDir)
}
// TODO any special processing of Vulkan devices?
resp = append(resp, info)
}
if len(resp) == 0 {
mem, err := GetCPUMem()
if err != nil {
slog.Warn("error looking up system memory", "error", err)
}
resp = append(resp, GpuInfo{
memInfo: mem,
DeviceID: ml.DeviceID{
Library: "cpu",
ID: "0",
},
})
}
return resp
}
// Given the list of GPUs this instantiation is targeted for,
// figure out the visible devices environment variable
//
// If different libraries are detected, the first one is what we use
func (l GpuInfoList) GetVisibleDevicesEnv() []string {
if len(l) == 0 {
return nil
}
res := []string{}
envVar := rocmGetVisibleDevicesEnv(l)
if envVar != "" {
res = append(res, envVar)
}
envVar = vkGetVisibleDevicesEnv(l)
if envVar != "" {
res = append(res, envVar)
}
return res
}
func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) string {
ids := []string{}
for _, info := range gpuInfo {
if info.Library != "ROCm" {
continue
}
// If the devices requires a numeric ID, for filtering purposes, we use the unfiltered ID number
if info.filterID != "" {
ids = append(ids, info.filterID)
} else {
ids = append(ids, info.ID)
}
}
if len(ids) == 0 {
return ""
}
envVar := "ROCR_VISIBLE_DEVICES="
if runtime.GOOS != "linux" {
envVar = "HIP_VISIBLE_DEVICES="
}
// There are 3 potential env vars to use to select GPUs.
// ROCR_VISIBLE_DEVICES supports UUID or numeric but does not work on Windows
// HIP_VISIBLE_DEVICES supports numeric IDs only
// GPU_DEVICE_ORDINAL supports numeric IDs only
return envVar + strings.Join(ids, ",")
}
func vkGetVisibleDevicesEnv(gpuInfo []GpuInfo) string {
ids := []string{}
for _, info := range gpuInfo {
if info.Library != "Vulkan" {
continue
}
if info.filterID != "" {
ids = append(ids, info.filterID)
} else {
ids = append(ids, info.ID)
}
}
if len(ids) == 0 {
return ""
}
envVar := "GGML_VK_VISIBLE_DEVICES="
return envVar + strings.Join(ids, ",")
}
// GetSystemInfo returns the last cached state of the GPUs on the system
func GetSystemInfo() SystemInfo {
deviceMu.Lock()
defer deviceMu.Unlock()
gpus := devInfoToInfoList(devices)
if len(gpus) == 1 && gpus[0].Library == "cpu" {
gpus = []GpuInfo{}
if threadCount == 0 {
// Fall back to Go's num CPU
threadCount = runtime.NumCPU()
}
return SystemInfo{
System: CPUInfo{
CPUs: GetCPUDetails(),
GpuInfo: GetCPUInfo(),
},
GPUs: gpus,
return ml.SystemInfo{
ThreadCount: threadCount,
TotalMemory: memInfo.TotalMemory,
FreeMemory: memInfo.FreeMemory,
FreeSwap: memInfo.FreeSwap,
}
}

162
discover/runner.go
View File

@@ -4,13 +4,8 @@ package discover
import (
"context"
"encoding/json"
"fmt"
"io"
"log/slog"
"math/rand"
"net"
"net/http"
"os"
"os/exec"
"path/filepath"
@@ -23,6 +18,7 @@ import (
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format"
"github.com/ollama/ollama/llm"
"github.com/ollama/ollama/logutil"
"github.com/ollama/ollama/ml"
)
@@ -36,7 +32,7 @@ var (
bootstrapped bool
)
func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.DeviceInfo {
func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.DeviceInfo {
deviceMu.Lock()
defer deviceMu.Unlock()
startDiscovery := time.Now()
@@ -154,9 +150,9 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
slog.Error("Unknown Library:" + devices[i].Library)
}
extraEnvs := []string{
"GGML_CUDA_INIT=1", // force deep initialization to trigger crash on unsupported GPUs
envVar + "=" + id, // Filter to just this one GPU
extraEnvs := map[string]string{
"GGML_CUDA_INIT": "1", // force deep initialization to trigger crash on unsupported GPUs
envVar: id, // Filter to just this one GPU
}
if len(bootstrapDevices(ctx2ndPass, devices[i].LibraryPath, extraEnvs)) == 0 {
needsDelete[i] = true
@@ -449,100 +445,35 @@ func (r *bootstrapRunner) HasExited() bool {
return false
}
func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []string) []ml.DeviceInfo {
// TODO DRY out with llm/server.go
slog.Debug("spawning runner with", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs)
func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs map[string]string) []ml.DeviceInfo {
var out io.Writer
if envconfig.LogLevel() == logutil.LevelTrace {
out = os.Stderr
}
start := time.Now()
defer func() {
slog.Debug("bootstrap discovery took", "duration", time.Since(start), "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs)
}()
port := 0
if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
var l *net.TCPListener
if l, err = net.ListenTCP("tcp", a); err == nil {
port = l.Addr().(*net.TCPAddr).Port
l.Close()
}
}
if port == 0 {
slog.Debug("ResolveTCPAddr failed, using random port")
port = rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
}
params := []string{"runner", "--ollama-engine", "--port", strconv.Itoa(port)}
var pathEnv string
switch runtime.GOOS {
case "windows":
pathEnv = "PATH"
case "darwin":
pathEnv = "DYLD_LIBRARY_PATH"
default:
pathEnv = "LD_LIBRARY_PATH"
}
libraryPaths := append([]string{LibOllamaPath}, ollamaLibDirs...)
if rocmDir != "" {
libraryPaths = append(libraryPaths, rocmDir)
}
// Note: we always put our dependency paths first
// since these are the exact version we compiled/linked against
if libraryPath, ok := os.LookupEnv(pathEnv); ok {
libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
}
cmd := exec.Command(exe, params...)
cmd.Env = os.Environ()
if envconfig.LogLevel() == logutil.LevelTrace {
cmd.Stdout = os.Stdout
cmd.Stderr = os.Stderr
}
// cmd.SysProcAttr = llm.LlamaServerSysProcAttr // circular dependency - bring back once refactored
pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))
pathNeeded := true
ollamaPathNeeded := true
extraDone := make([]bool, len(extraEnvs))
for i := range cmd.Env {
cmp := strings.SplitN(cmd.Env[i], "=", 2)
if strings.EqualFold(cmp[0], pathEnv) {
cmd.Env[i] = pathEnv + "=" + pathEnvVal
pathNeeded = false
} else if strings.EqualFold(cmp[0], "OLLAMA_LIBRARY_PATH") {
cmd.Env[i] = "OLLAMA_LIBRARY_PATH=" + strings.Join(ollamaLibDirs, string(filepath.ListSeparator))
ollamaPathNeeded = false
} else {
for j := range extraEnvs {
if extraDone[j] {
continue
}
extra := strings.SplitN(extraEnvs[j], "=", 2)
if cmp[0] == extra[0] {
cmd.Env[i] = extraEnvs[j]
extraDone[j] = true
}
}
}
}
if pathNeeded {
cmd.Env = append(cmd.Env, pathEnv+"="+pathEnvVal)
}
if ollamaPathNeeded {
cmd.Env = append(cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ollamaLibDirs, string(filepath.ListSeparator)))
}
for i := range extraDone {
if !extraDone[i] {
cmd.Env = append(cmd.Env, extraEnvs[i])
}
}
logutil.Trace("starting runner for device discovery", "env", cmd.Env, "cmd", cmd)
if err := cmd.Start(); err != nil {
slog.Warn("unable to start discovery subprocess", "cmd", cmd, "error", err)
logutil.Trace("starting runner for device discovery", "libDirs", ollamaLibDirs, "extraEnvs", extraEnvs)
cmd, port, err := llm.StartRunner(
true, // ollama engine
"", // no model
ollamaLibDirs,
out,
extraEnvs,
)
if err != nil {
slog.Debug("failed to start runner to discovery GPUs", "error", err)
return nil
}
go func() {
cmd.Wait() // exit status ignored
}()
defer cmd.Process.Kill()
devices, err := GetDevicesFromRunner(ctx, &bootstrapRunner{port: port, cmd: cmd})
devices, err := ml.GetDevicesFromRunner(ctx, &bootstrapRunner{port: port, cmd: cmd})
if err != nil {
if cmd.ProcessState != nil && cmd.ProcessState.ExitCode() >= 0 {
// Expected during bootstrapping while we filter out unsupported AMD GPUs
@@ -555,52 +486,3 @@ func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []s
return devices
}
func GetDevicesFromRunner(ctx context.Context, runner BaseRunner) ([]ml.DeviceInfo, error) {
var moreDevices []ml.DeviceInfo
port := runner.GetPort()
tick := time.Tick(10 * time.Millisecond)
for {
select {
case <-ctx.Done():
return nil, fmt.Errorf("failed to finish discovery before timeout")
case <-tick:
r, err := http.NewRequestWithContext(ctx, http.MethodGet, fmt.Sprintf("http://127.0.0.1:%d/info", port), nil)
if err != nil {
return nil, fmt.Errorf("failed to create request: %w", err)
}
r.Header.Set("Content-Type", "application/json")
resp, err := http.DefaultClient.Do(r)
if err != nil {
// slog.Warn("failed to send request", "error", err)
if runner.HasExited() {
return nil, fmt.Errorf("runner crashed")
}
continue
}
defer resp.Body.Close()
if resp.StatusCode == http.StatusNotFound {
// old runner, fall back to bootstrapping model
return nil, fmt.Errorf("llamarunner free vram reporting not supported")
}
body, err := io.ReadAll(resp.Body)
if err != nil {
slog.Warn("failed to read response", "error", err)
continue
}
if resp.StatusCode != 200 {
logutil.Trace("runner failed to discover free VRAM", "status", resp.StatusCode, "response", body)
return nil, fmt.Errorf("runner error: %s", string(body))
}
if err := json.Unmarshal(body, &moreDevices); err != nil {
slog.Warn("unmarshal encode response", "error", err)
continue
}
return moreDevices, nil
}
}
}

143
discover/types.go
View File

@@ -1,10 +1,8 @@
package discover
import (
"context"
"log/slog"
"path/filepath"
"runtime"
"strings"
"github.com/ollama/ollama/format"
@@ -17,50 +15,6 @@ type memInfo struct {
FreeSwap uint64 `json:"free_swap,omitempty"` // TODO split this out for system only
}
// Beginning of an `ollama info` command
type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
ml.DeviceID
memInfo
// Optional variant to select (e.g. versions, cpu feature flags)
Variant string `json:"variant"`
// MinimumMemory represents the minimum memory required to use the GPU
MinimumMemory uint64 `json:"-"`
// Any extra PATH/LD_LIBRARY_PATH dependencies required for the Library to operate properly
DependencyPath []string `json:"lib_path,omitempty"`
// Set to true if we can NOT reliably discover FreeMemory. A value of true indicates
// the FreeMemory is best effort, and may over or under report actual memory usage
// False indicates FreeMemory can generally be trusted on this GPU
UnreliableFreeMemory bool
// GPU information
filterID string // AMD/Vulkan Workaround: The numeric ID of the device used to filter out other devices
Name string `json:"name"` // user friendly name if available
ComputeMajor int `json:"compute_major"` // Compute Capability or gfx
ComputeMinor int `json:"compute_minor"`
// Driver Information - TODO no need to put this on each GPU
DriverMajor int `json:"driver_major,omitempty"`
DriverMinor int `json:"driver_minor,omitempty"`
// TODO other performance capability info to help in scheduling decisions
}
func (gpu GpuInfo) RunnerName() string {
if gpu.Variant != "" {
return gpu.Library + "_" + gpu.Variant
}
return gpu.Library
}
type CPUInfo struct {
GpuInfo
CPUs []CPU
}
// CPU type represents a CPU Package occupying a socket
type CPU struct {
ID string `cpuinfo:"processor"`
@@ -71,32 +25,6 @@ type CPU struct {
ThreadCount int
}
type GpuInfoList []GpuInfo
func (l GpuInfoList) ByLibrary() []GpuInfoList {
resp := []GpuInfoList{}
libs := []string{}
for _, info := range l {
found := false
requested := info.Library
if info.Variant != "" {
requested += "_" + info.Variant
}
for i, lib := range libs {
if lib == requested {
resp[i] = append(resp[i], info)
found = true
break
}
}
if !found {
libs = append(libs, requested)
resp = append(resp, []GpuInfo{info})
}
}
return resp
}
func LogDetails(devices []ml.DeviceInfo) {
for _, dev := range devices {
var libs []string
@@ -141,74 +69,3 @@ func LogDetails(devices []ml.DeviceInfo) {
)
}
}
// Sort by Free Space
type ByFreeMemory []GpuInfo
func (a ByFreeMemory) Len() int { return len(a) }
func (a ByFreeMemory) Swap(i, j int) { a[i], a[j] = a[j], a[i] }
func (a ByFreeMemory) Less(i, j int) bool { return a[i].FreeMemory < a[j].FreeMemory }
type SystemInfo struct {
System CPUInfo `json:"system"`
GPUs []GpuInfo `json:"gpus"`
}
// Return the optimal number of threads to use for inference
func (si SystemInfo) GetOptimalThreadCount() int {
if len(si.System.CPUs) == 0 {
// Fall back to Go's num CPU
return runtime.NumCPU()
}
coreCount := 0
for _, c := range si.System.CPUs {
coreCount += c.CoreCount - c.EfficiencyCoreCount
}
return coreCount
}
// For each GPU, check if it does NOT support flash attention
func (l GpuInfoList) FlashAttentionSupported() bool {
for _, gpu := range l {
supportsFA := gpu.Library == "cpu" ||
gpu.Name == "Metal" || gpu.Library == "Metal" ||
(gpu.Library == "CUDA" && gpu.DriverMajor >= 7 && !(gpu.ComputeMajor == 7 && gpu.ComputeMinor == 2)) || // We don't have kernels for Jetson Xavier
gpu.Library == "ROCm" ||
gpu.Library == "Vulkan"
if !supportsFA {
return false
}
}
return true
}
type BaseRunner interface {
// GetPort returns the localhost port number the runner is running on
GetPort() int
// HasExited indicates if the runner is no longer running. This can be used during
// bootstrap to detect if a given filtered device is incompatible and triggered an assert
HasExited() bool
}
type RunnerDiscovery interface {
BaseRunner
// GetDeviceInfos will perform a query of the underlying device libraries
// for device identification and free VRAM information
// During bootstrap scenarios, this routine may take seconds to complete
GetDeviceInfos(ctx context.Context) []ml.DeviceInfo
}
type FilteredRunnerDiscovery interface {
RunnerDiscovery
// GetActiveDeviceIDs returns the filtered set of devices actively in
// use by this runner for running models. If the runner is a bootstrap runner, no devices
// will be active yet so no device IDs are returned.
// This routine will not query the underlying device and will return immediately
GetActiveDeviceIDs() []ml.DeviceID
}