mirror of
https://github.com/mudler/LocalAI.git
synced 2026-10-28 09:08:54 -04:00
manunicholasjacob
aca800d003
When --threads / LOCALAI_THREADS is unset, LocalAI uses the host's physical core count. Inside a container with a CPU limit (docker --cpus, a Kubernetes limits.cpu) that starts one thread per host core inside a smaller CFS quota, so the backend is throttled for most of each period. Since Go 1.25 the default GOMAXPROCS already follows the affinity mask and the cgroup CPU limit. Use the smaller of the physical core count and GOMAXPROCS as the default. Hosts without a limit are unchanged. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: manunicholasjacob <manunicholasjacob@gmail.com>
107 lines
2.4 KiB
Go
107 lines
2.4 KiB
Go
package xsysinfo
|
|
|
|
import (
|
|
"fmt"
|
|
"maps"
|
|
"runtime"
|
|
"slices"
|
|
"sort"
|
|
|
|
"github.com/jaypipes/ghw"
|
|
"github.com/klauspost/cpuid/v2"
|
|
"github.com/shirou/gopsutil/v3/cpu"
|
|
"github.com/shirou/gopsutil/v3/load"
|
|
)
|
|
|
|
// CPUInfo describes host-wide CPU capacity and current utilization.
|
|
type CPUInfo struct {
|
|
LogicalCores uint64
|
|
UsagePercent float64
|
|
Load1 float64
|
|
}
|
|
|
|
// GetCPUInfo samples host-wide CPU telemetry.
|
|
func GetCPUInfo() (*CPUInfo, error) {
|
|
logicalCores, err := cpu.Counts(true)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("counting logical CPU cores: %w", err)
|
|
}
|
|
if logicalCores < 1 {
|
|
return nil, fmt.Errorf("counting logical CPU cores: no cores reported")
|
|
}
|
|
|
|
usage, err := cpu.Percent(0, false)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("sampling CPU utilization: %w", err)
|
|
}
|
|
if len(usage) != 1 {
|
|
return nil, fmt.Errorf("sampling CPU utilization: expected one aggregate reading, got %d", len(usage))
|
|
}
|
|
|
|
loadAverage, err := load.Avg()
|
|
if err != nil {
|
|
return nil, fmt.Errorf("sampling CPU load: %w", err)
|
|
}
|
|
|
|
return &CPUInfo{
|
|
LogicalCores: uint64(logicalCores),
|
|
UsagePercent: usage[0],
|
|
Load1: loadAverage.Load1,
|
|
}, nil
|
|
}
|
|
|
|
func CPUCapabilities() ([]string, error) {
|
|
cpu, err := ghw.CPU()
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
|
|
caps := map[string]struct{}{}
|
|
|
|
for _, proc := range cpu.Processors {
|
|
for _, c := range proc.Capabilities {
|
|
|
|
caps[c] = struct{}{}
|
|
}
|
|
|
|
}
|
|
|
|
ret := slices.Collect(maps.Keys(caps))
|
|
|
|
// order
|
|
sort.Strings(ret)
|
|
return ret, nil
|
|
}
|
|
|
|
func HasCPUCaps(ids ...cpuid.FeatureID) bool {
|
|
return cpuid.CPU.Supports(ids...)
|
|
}
|
|
|
|
func CPUPhysicalCores() int {
|
|
if cpuid.CPU.PhysicalCores == 0 {
|
|
return 1
|
|
}
|
|
return cpuid.CPU.PhysicalCores
|
|
}
|
|
|
|
// DefaultThreads returns the default number of inference threads: the number
|
|
// of physical cores, capped by the number of CPUs the Go runtime may use.
|
|
// Since Go 1.25 the default GOMAXPROCS follows the CPU affinity mask and the
|
|
// cgroup CPU limit, so a container started with --cpus=2 or a Kubernetes CPU
|
|
// limit of 2 gets 2 threads instead of one per physical core on the host.
|
|
func DefaultThreads() int {
|
|
return capThreads(CPUPhysicalCores(), runtime.GOMAXPROCS(0))
|
|
}
|
|
|
|
// capThreads returns physical, lowered to usable when usable is smaller and
|
|
// positive, and never less than 1.
|
|
func capThreads(physical, usable int) int {
|
|
if physical < 1 {
|
|
physical = 1
|
|
}
|
|
if usable > 0 && usable < physical {
|
|
return usable
|
|
}
|
|
return physical
|
|
}
|