Merge pull request #540 from docker/responses

Add OpenAI Responses API compatibility layer

Author: ericcurtin

main.go

@@ -23,6 +23,7 @@ import (
 	"github.com/docker/model-runner/pkg/metrics"
 	"github.com/docker/model-runner/pkg/middleware"
 	"github.com/docker/model-runner/pkg/ollama"
+	"github.com/docker/model-runner/pkg/responses"
 	"github.com/docker/model-runner/pkg/routing"
 	"github.com/sirupsen/logrus"
 )
@@ -165,6 +166,16 @@ func main() {
 	router.Handle(inference.ModelsPrefix, modelHandler)
 	router.Handle(inference.ModelsPrefix+"/", modelHandler)
 	router.Handle(inference.InferencePrefix+"/", schedulerHTTP)
+	// Add OpenAI Responses API compatibility layer
+	responsesHandler := responses.NewHTTPHandler(log, schedulerHTTP, nil)
+	router.Handle(responses.APIPrefix+"/", responsesHandler)
+	router.Handle(responses.APIPrefix, responsesHandler) // Also register for exact match without trailing slash
+	router.Handle("/v1"+responses.APIPrefix+"/", responsesHandler)
+	router.Handle("/v1"+responses.APIPrefix, responsesHandler)
+	// Also register Responses API under inference prefix to support all inference engines
+	router.Handle(inference.InferencePrefix+responses.APIPrefix+"/", responsesHandler)
+	router.Handle(inference.InferencePrefix+responses.APIPrefix, responsesHandler)
+
 	// Add path aliases: /v1 -> /engines/v1, /rerank -> /engines/rerank, /score -> /engines/score.
 	aliasHandler := &middleware.AliasHandler{Handler: schedulerHTTP}
 	router.Handle("/v1/", aliasHandler)

pkg/inference/models/adapter.go

@@ -1,6 +1,7 @@
 package models
 
 import (
+	"encoding/json"
 	"fmt"
 
 	"github.com/docker/model-runner/pkg/distribution/types"
@@ -27,12 +28,23 @@ func ToModel(m types.Model) (*Model, error) {
 		created = desc.Created.Unix()
 	}
 
-	return &Model{
+	model := &Model{
 		ID:      id,
 		Tags:    m.Tags(),
 		Created: created,
 		Config:  cfg,
-	}, nil
+	}
+
+	// Marshal the config to populate RawConfig
+	if cfg != nil {
+		configData, err := json.Marshal(cfg)
+		if err != nil {
+			return nil, fmt.Errorf("marshal config: %w", err)
+		}
+		model.RawConfig = configData
+	}
+
+	return model, nil
 }
 
 // ToModelFromArtifact converts a types.ModelArtifact (typically from remote registry)
@@ -58,10 +70,21 @@ func ToModelFromArtifact(artifact types.ModelArtifact) (*Model, error) {
 		created = desc.Created.Unix()
 	}
 
-	return &Model{
+	model := &Model{
 		ID:      id,
 		Tags:    nil, // Remote models don't have local tags
 		Created: created,
 		Config:  cfg,
-	}, nil
+	}
+
+	// Marshal the config to populate RawConfig
+	if cfg != nil {
+		configData, err := json.Marshal(cfg)
+		if err != nil {
+			return nil, fmt.Errorf("marshal config: %w", err)
+		}
+		model.RawConfig = configData
+	}
+
+	return model, nil
 }

pkg/inference/models/api.go

@@ -112,7 +112,35 @@ type Model struct {
 	Created int64 `json:"created"`
 	// Config describes the model. Can be either Docker format (*types.Config)
 	// or ModelPack format (*modelpack.Model).
-	Config types.ModelConfig `json:"config"`
+	Config types.ModelConfig `json:"-"`
+	// RawConfig is used for JSON marshaling/unmarshaling
+	RawConfig json.RawMessage `json:"config"`
+}
+
+// MarshalJSON implements custom marshaling for Model
+func (m Model) MarshalJSON() ([]byte, error) {
+	// Define a temporary struct to avoid recursion
+	type Alias Model
+	aux := struct {
+		*Alias
+		RawConfig json.RawMessage `json:"config"`
+	}{
+		Alias: (*Alias)(&m),
+	}
+
+	// Marshal the config separately
+	if m.Config != nil {
+		configData, err := json.Marshal(m.Config)
+		if err != nil {
+			return nil, err
+		}
+		aux.RawConfig = configData
+	} else {
+		// If Config is nil, use the RawConfig if available
+		aux.RawConfig = m.RawConfig
+	}
+
+	return json.Marshal(aux)
 }
 
 // UnmarshalJSON implements custom JSON unmarshaling for Model.