glpi-neural-brain/internal/engine/speed_mode.go

package engine

import (
	"context"
	"runtime"
	"time"

	"github.com/local/glpi-neural-brain/internal/model"
	"github.com/local/glpi-neural-brain/internal/ollama"
)

const speedModeVersion = "throughput-v1"

func (e *Engine) effectiveSpeedCPUWorkers(settings RuntimeSettings) int {
	workers := settings.SpeedCPUWorkers
	if workers < 1 {
		workers = runtime.NumCPU()
	}
	if workers < 1 {
		workers = 1
	}
	if workers > 256 {
		workers = 256
	}
	return workers
}

func (e *Engine) effectiveSpeedGPUInflight(settings RuntimeSettings) int {
	limit := settings.SpeedGPUInflight
	if limit < 1 {
		limit = 1
	}
	if limit > 64 {
		limit = 64
	}
	return limit
}

func (e *Engine) applyRuntimePerformance(settings RuntimeSettings) {
	if e == nil {
		return
	}
	sharedInflight := e.Cfg.ResearchOllamaMaxInflight
	queueSize := e.Cfg.ResearchOllamaQueueSize
	gpuInflight := e.Cfg.OllamaNodeMaxInflight
	if sharedInflight < 1 {
		sharedInflight = 1
	}
	if queueSize < 1 {
		queueSize = 1
	}
	if gpuInflight < 1 {
		gpuInflight = 1
	}
	if settings.SpeedMode {
		cpu := e.effectiveSpeedCPUWorkers(settings)
		gpu := e.effectiveSpeedGPUInflight(settings)
		// The shared limiter covers model calls and outbound research work. Do not
		// let it become the bottleneck below either requested CPU fan-out or total
		// per-node GPU fan-out.
		nodeCount := 1
		if e.Ollama != nil {
			nodeCount = len(e.Ollama.NodeStatuses())
		}
		if nodeCount < 1 {
			nodeCount = 1
		}
		sharedInflight = maxInt(cpu, gpu*nodeCount)
		queueSize = maxInt(queueSize, sharedInflight*16)
		gpuInflight = gpu
	}
	if e.sharedWork != nil {
		e.sharedWork.SetLimits(sharedInflight, queueSize)
	}
	if e.Ollama != nil {
		e.Ollama.SetNodeMaxInflight(gpuInflight)
	}
}

func (e *Engine) signalSpeedWork() {
	if e == nil || !e.SpeedModeEnabled() {
		return
	}
	if e.ThinkingEnabled() {
		e.RequestEnrich("speed-mode")
	}
	if e.AutonomousResearchEnabled() {
		e.RequestAutonomousResearchScan("speed-mode")
		e.signalAutonomousResearch()
	}
	if e.sourceInboxWake != nil {
		select {
		case e.sourceInboxWake <- struct{}{}:
		default:
		}
	}
	if e.Broker != nil {
		settings := e.RuntimeSettings()
		e.Broker.Publish(model.Activity{Type: "performance.speed.pulse", Source: "brain", Phase: "performance", Message: "Speed-Modus treibt vorhandene CPU-/GPU-Arbeit ohne künstliche Pacing-Pausen weiter", Strength: .58, Metadata: map[string]any{"algorithm": speedModeVersion, "cpu_tasks": settings.SpeedCPUWorkers, "gpu_tasks_per_node": settings.SpeedGPUInflight}})
	}
}

func (e *Engine) backgroundOllamaContext(ctx context.Context) context.Context {
	if e != nil && e.SpeedModeEnabled() {
		return ctx
	}
	return ollama.WithLowPriority(ctx)
}

func (e *Engine) effectiveEnrichStepDelay() time.Duration {
	if e != nil && e.SpeedModeEnabled() {
		return 0
	}
	return e.Cfg.EnrichStepDelay
}