feat: add AI Remote Worker (#3168)

This commit adds a new AI remote worker node which can be used to split worker and orchestrator machines similar to how it is done on the transcoding side.

Co-authored-by: Rafał Leszko <[email protected]>
Co-authored-by: Rick Staa <[email protected]>

Author: 3 people

cmd/livepeer/starter/starter.go

@@ -89,7 +89,7 @@ func IgnoreRoutines() []goleak.Option {
 		"github.com/livepeer/go-livepeer/server.(*LivepeerServer).StartMediaServer", "github.com/livepeer/go-livepeer/core.(*RemoteTranscoderManager).Manage.func1",
 		"github.com/livepeer/go-livepeer/server.(*LivepeerServer).HandlePush.func1", "github.com/rjeczalik/notify.(*nonrecursiveTree).dispatch",
 		"github.com/rjeczalik/notify.(*nonrecursiveTree).internal", "github.com/livepeer/lpms/stream.NewBasicRTMPVideoStream.func1", "github.com/patrickmn/go-cache.(*janitor).Run",
-		"github.com/golang/glog.(*fileSink).flushDaemon",
+		"github.com/golang/glog.(*fileSink).flushDaemon", "github.com/livepeer/go-livepeer/core.(*LivepeerNode).transcodeFrames.func2",
 	}
 
 	res := make([]goleak.Option, 0, len(funcs2ignore))

common/testutil.go

@@ -77,6 +77,7 @@ var (
 	ErrProfName     = fmt.Errorf("unknown VideoProfile profile name")
 
 	ErrAudioDurationCalculation = fmt.Errorf("audio duration calculation failed")
+	ErrNoExtensionsForType      = fmt.Errorf("no extensions exist for mime type")
 
 	ext2mime = map[string]string{
 		".ts":  "video/mp2t",
@@ -571,3 +572,17 @@ func CalculateAudioDuration(audio types.File) (int64, error) {
 func ValidateServiceURI(serviceURI *url.URL) bool {
 	return !strings.Contains(serviceURI.Host, "0.0.0.0")
 }
+
+func ExtensionByType(contentType string) (string, error) {
+	contentType = strings.ToLower(contentType)
+	switch contentType {
+	case "video/mp2t":
+		return ".ts", nil
+	case "video/mp4":
+		return ".mp4", nil
+	case "image/png":
+		return ".png", nil
+	}
+
+	return "", ErrNoExtensionsForType
+}

common/util.go

@@ -519,3 +519,21 @@ func TestValidateServiceURI(t *testing.T) {
 		}
 	}
 }
+func TestExtensionByType(t *testing.T) {
+	assert := assert.New(t)
+
+	// Test valid content types
+	contentTypes := []string{"image/png", "video/mp4", "video/mp2t"}
+	expectedExtensions := []string{".png", ".mp4", ".ts"}
+
+	for i, contentType := range contentTypes {
+		ext, err := ExtensionByType(contentType)
+		assert.Nil(err)
+		assert.Equal(expectedExtensions[i], ext)
+	}
+
+	// Test invalid content type
+	invalidContentType := "invalid/type"
+	_, err := ExtensionByType(invalidContentType)
+	assert.Equal(ErrNoExtensionsForType, err)
+}

common/util_test.go

@@ -11,6 +11,7 @@ import (
 	"strconv"
 	"strings"
 
+	"github.com/golang/glog"
 	"github.com/livepeer/ai-worker/worker"
 )
 
@@ -64,15 +65,19 @@ func PipelineToCapability(pipeline string) (Capability, error) {
 }
 
 type AIModelConfig struct {
-	Pipeline          string                   `json:"pipeline"`
-	ModelID           string                   `json:"model_id"`
+	Pipeline string `json:"pipeline"`
+	ModelID  string `json:"model_id"`
+	// used by worker
 	URL               string                   `json:"url,omitempty"`
 	Token             string                   `json:"token,omitempty"`
 	Warm              bool                     `json:"warm,omitempty"`
-	PricePerUnit      JSONRat                  `json:"price_per_unit,omitempty"`
-	PixelsPerUnit     JSONRat                  `json:"pixels_per_unit,omitempty"`
-	Currency          string                   `json:"currency,omitempty"`
+	Capacity          int                      `json:"capacity,omitempty"`
 	OptimizationFlags worker.OptimizationFlags `json:"optimization_flags,omitempty"`
+	// used by orchestrator
+	Gateway       string  `json:"gateway"`
+	PricePerUnit  JSONRat `json:"price_per_unit,omitempty"`
+	PixelsPerUnit JSONRat `json:"pixels_per_unit,omitempty"`
+	Currency      string  `json:"currency,omitempty"`
 }
 
 func ParseAIModelConfigs(config string) ([]AIModelConfig, error) {
@@ -112,7 +117,7 @@ func ParseAIModelConfigs(config string) ([]AIModelConfig, error) {
 	return configs, nil
 }
 
-// parseStepsFromModelID parses the number of inference steps from the model ID suffix.
+// ParseStepsFromModelID parses the number of inference steps from the model ID suffix.
 func ParseStepsFromModelID(modelID *string, defaultSteps float64) float64 {
 	numInferenceSteps := defaultSteps
 
@@ -127,3 +132,102 @@ func ParseStepsFromModelID(modelID *string, defaultSteps float64) float64 {
 
 	return numInferenceSteps
 }
+
+// AddAICapabilities adds AI capabilities to the node.
+func (n *LivepeerNode) AddAICapabilities(caps *Capabilities) {
+	aiConstraints := caps.PerCapability()
+	if aiConstraints == nil {
+		return
+	}
+
+	n.Capabilities.mutex.Lock()
+	defer n.Capabilities.mutex.Unlock()
+	for aiCapability, aiConstraint := range aiConstraints {
+		_, capExists := n.Capabilities.constraints.perCapability[aiCapability]
+		if !capExists {
+			n.Capabilities.constraints.perCapability[aiCapability] = &CapabilityConstraints{
+				Models: make(ModelConstraints),
+			}
+		}
+
+		for modelId, modelConstraint := range aiConstraint.Models {
+			_, modelExists := n.Capabilities.constraints.perCapability[aiCapability].Models[modelId]
+			if modelExists {
+				n.Capabilities.constraints.perCapability[aiCapability].Models[modelId].Capacity += modelConstraint.Capacity
+			} else {
+				n.Capabilities.constraints.perCapability[aiCapability].Models[modelId] = &ModelConstraint{Warm: modelConstraint.Warm, Capacity: modelConstraint.Capacity}
+			}
+		}
+	}
+}
+
+// RemoveAICapabilities removes AI capabilities from the node.
+func (n *LivepeerNode) RemoveAICapabilities(caps *Capabilities) {
+	aiConstraints := caps.PerCapability()
+	if aiConstraints == nil {
+		return
+	}
+
+	n.Capabilities.mutex.Lock()
+	defer n.Capabilities.mutex.Unlock()
+	for capability, constraint := range aiConstraints {
+		_, ok := n.Capabilities.constraints.perCapability[capability]
+		if ok {
+			for modelId, modelConstraint := range constraint.Models {
+				_, modelExists := n.Capabilities.constraints.perCapability[capability].Models[modelId]
+				if modelExists {
+					n.Capabilities.constraints.perCapability[capability].Models[modelId].Capacity -= modelConstraint.Capacity
+					if n.Capabilities.constraints.perCapability[capability].Models[modelId].Capacity <= 0 {
+						delete(n.Capabilities.constraints.perCapability[capability].Models, modelId)
+					}
+				} else {
+					glog.Errorf("failed to remove AI capability capacity, model does not exist pipeline=%v modelID=%v", capability, modelId)
+				}
+			}
+		}
+	}
+}
+
+func (n *LivepeerNode) ReserveAICapability(pipeline string, modelID string) error {
+	cap, err := PipelineToCapability(pipeline)
+	if err != nil {
+		return err
+	}
+
+	_, hasCap := n.Capabilities.constraints.perCapability[cap]
+	if hasCap {
+		_, hasModel := n.Capabilities.constraints.perCapability[cap].Models[modelID]
+		if hasModel {
+			n.Capabilities.mutex.Lock()
+			defer n.Capabilities.mutex.Unlock()
+			if n.Capabilities.constraints.perCapability[cap].Models[modelID].Capacity > 0 {
+				n.Capabilities.constraints.perCapability[cap].Models[modelID].Capacity -= 1
+			} else {
+				return fmt.Errorf("failed to reserve AI capability capacity, model capacity is 0 pipeline=%v modelID=%v", pipeline, modelID)
+			}
+			return nil
+		}
+		return fmt.Errorf("failed to reserve AI capability capacity, model does not exist pipeline=%v modelID=%v", pipeline, modelID)
+	}
+	return fmt.Errorf("failed to reserve AI capability capacity, pipeline does not exist pipeline=%v modelID=%v", pipeline, modelID)
+}
+
+func (n *LivepeerNode) ReleaseAICapability(pipeline string, modelID string) error {
+	cap, err := PipelineToCapability(pipeline)
+	if err != nil {
+		return err
+	}
+	_, hasCap := n.Capabilities.constraints.perCapability[cap]
+	if hasCap {
+		_, hasModel := n.Capabilities.constraints.perCapability[cap].Models[modelID]
+		if hasModel {
+			n.Capabilities.mutex.Lock()
+			defer n.Capabilities.mutex.Unlock()
+			n.Capabilities.constraints.perCapability[cap].Models[modelID].Capacity += 1
+
+			return nil
+		}
+		return fmt.Errorf("failed to release AI capability capacity, model does not exist pipeline=%v modelID=%v", pipeline, modelID)
+	}
+	return fmt.Errorf("failed to release AI capability capacity, pipeline does not exist pipeline=%v modelID=%v", pipeline, modelID)
+}