fix: update vllm to v0.4.0 hotfix version

pkg/appruntime/agent/streamhandler.go

@@ -35,11 +35,11 @@ type StreamHandler struct {
 var _ callbacks.Handler = StreamHandler{}
 
 func (handler StreamHandler) HandleStreamingFunc(ctx context.Context, chunk []byte) {
-	if _, ok := handler.args[base.OutputAnserStreamChanKeyInArg]; ok {
+	if _, ok := handler.args[base.OutputAnswerStreamChanKeyInArg]; ok {
 		logger := klog.FromContext(ctx)
-		streamChan, ok := handler.args[base.OutputAnserStreamChanKeyInArg].(chan string)
+		streamChan, ok := handler.args[base.OutputAnswerStreamChanKeyInArg].(chan string)
 		if !ok {
-			err := fmt.Errorf("answer_stream is not chan string, but %T", handler.args[base.OutputAnserStreamChanKeyInArg])
+			err := fmt.Errorf("answer_stream is not chan string, but %T", handler.args[base.OutputAnswerStreamChanKeyInArg])
 			logger.Error(err, "answer_stream is not chan string")
 			return
 		}

pkg/appruntime/app_runtime.go

@@ -145,7 +145,7 @@ func (a *Application) Run(ctx context.Context, cli client.Client, respStream cha
 	out := map[string]any{
 		base.InputQuestionKeyInArg:                 input.Question,
 		"files":                                    input.Files,
-		base.OutputAnserStreamChanKeyInArg:         respStream,
+		base.OutputAnswerStreamChanKeyInArg:        respStream,
 		base.InputIsNeedStreamKeyInArg:             input.NeedStream,
 		base.LangchaingoChatMessageHistoryKeyInArg: input.History,
 		// Use an empty context before run
@@ -205,7 +205,7 @@ func (a *Application) Run(ctx context.Context, cli client.Client, respStream cha
 				var er *base.RetrieverGetNullDocError
 				if errors.As(err, &er) {
 					agentReturnNothing := true
-					v, ok := out[base.OutputAnserKeyInArg]
+					v, ok := out[base.OutputAnswerKeyInArg]
 					if ok {
 						if answer, ok := v.(string); ok && len(answer) > 0 {
 							agentReturnNothing = false
@@ -229,7 +229,7 @@ func (a *Application) Run(ctx context.Context, cli client.Client, respStream cha
 			waitRunningNodes.PushBack(n)
 		}
 	}
-	if a, ok := out[base.OutputAnserKeyInArg]; ok {
+	if a, ok := out[base.OutputAnswerKeyInArg]; ok {
 		if answer, ok := a.(string); ok && len(answer) > 0 {
 			output = Output{Answer: answer}
 		}

pkg/appruntime/base/keyword.go

@@ -20,10 +20,10 @@ const (
 	InputQuestionKeyInArg                 = "question"
 	InputIsNeedStreamKeyInArg             = "_need_stream"
 	LangchaingoChatMessageHistoryKeyInArg = "_history"
-	OutputAnserKeyInArg                   = "_answer"
+	OutputAnswerKeyInArg                  = "_answer"
 	AgentOutputInArg                      = "_agent_answer"
 	MapReduceDocumentOutputInArg          = "_mapreduce_document_answer"
-	OutputAnserStreamChanKeyInArg         = "_answer_stream"
+	OutputAnswerStreamChanKeyInArg        = "_answer_stream"
 	RuntimeRetrieverReferencesKeyInArg    = "_references"
 	LangchaingoRetrieverKeyInArg          = "retriever"
 	LangchaingoLLMKeyInArg                = "llm"

pkg/appruntime/chain/apichain.go

@@ -112,7 +112,7 @@ func (l *APIChain) Run(ctx context.Context, _ client.Client, args map[string]any
 	out, err = handleNoErrNoOut(ctx, needStream, out, err, l.APIChain, args, options)
 	klog.FromContext(ctx).V(5).Info("use apichain, blocking out:" + out)
 	if err == nil {
-		args[base.OutputAnserKeyInArg] = out
+		args[base.OutputAnswerKeyInArg] = out
 		return args, nil
 	}
 	return args, fmt.Errorf("apichain run error: %w", err)

pkg/appruntime/chain/common.go

@@ -33,11 +33,11 @@ import (
 
 func stream(res map[string]any) func(ctx context.Context, chunk []byte) error {
 	return func(ctx context.Context, chunk []byte) error {
-		if _, ok := res[base.OutputAnserStreamChanKeyInArg]; ok {
+		if _, ok := res[base.OutputAnswerStreamChanKeyInArg]; ok {
 			logger := klog.FromContext(ctx)
-			streamChan, ok := res[base.OutputAnserStreamChanKeyInArg].(chan string)
+			streamChan, ok := res[base.OutputAnswerStreamChanKeyInArg].(chan string)
 			if !ok {
-				err := fmt.Errorf("answer_stream is not chan string, but %T", res[base.OutputAnserStreamChanKeyInArg])
+				err := fmt.Errorf("answer_stream is not chan string, but %T", res[base.OutputAnswerStreamChanKeyInArg])
 				logger.Error(err, "answer_stream is not chan string")
 				return err
 			}

pkg/appruntime/chain/llmchain.go

@@ -153,7 +153,7 @@ func (l *LLMChain) Run(ctx context.Context, cli client.Client, args map[string]a
 	out, err = handleNoErrNoOut(ctx, needStream, out, err, l.LLMChain, args, options)
 	klog.FromContext(ctx).V(5).Info("use llmchain, blocking out:" + out)
 	if err == nil {
-		args[base.OutputAnserKeyInArg] = out
+		args[base.OutputAnswerKeyInArg] = out
 		return args, nil
 	}
 	return args, fmt.Errorf("llmchain run error: %w", err)

pkg/appruntime/chain/retrievalqachain.go

@@ -186,7 +186,7 @@ func (l *RetrievalQAChain) Run(ctx context.Context, cli client.Client, args map[
 	out, err = handleNoErrNoOut(ctx, needStream, out, err, l.ConversationalRetrievalQA, args, options)
 	klog.FromContext(ctx).V(5).Info("use retrievalqachain, blocking out:" + out)
 	if err == nil {
-		args[base.OutputAnserKeyInArg] = out
+		args[base.OutputAnswerKeyInArg] = out
 		return args, nil
 	}
 	return args, fmt.Errorf("retrievalqachain run error: %w", err)

pkg/worker/runner.go

@@ -34,9 +34,9 @@ import (
 
 const (
 	// tag is the same version as fastchat
-	defaultFastChatImage = "kubeagi/arcadia-fastchat-worker:v0.2.36"
+	defaultFastChatImage = "kubeagi/arcadia-fastchat-worker:vllm-v0.4.0-hotfix"
 	// For ease of maintenance and stability, VLLM module is now included in standard image as a default feature.
-	defaultFastchatVLLMImage = "kubeagi/arcadia-fastchat-worker:v0.2.36"
+	defaultFastchatVLLMImage = "kubeagi/arcadia-fastchat-worker:vllm-v0.4.0-hotfix"
 	// defaultKubeAGIImage for RunnerKubeAGI
 	defaultKubeAGIImage = "kubeagi/core-library-cli:v0.0.1"
 
@@ -249,9 +249,7 @@ func (runner *RunnerFastchatVLLM) Build(ctx context.Context, model *arcadiav1alp
 	additionalEnvs = append(additionalEnvs, corev1.EnvVar{Name: "NUMBER_GPUS", Value: gpus})
 
 	modelFileDir := fmt.Sprintf("%s/%s", defaultModelMountPath, model.Name)
-	// --enforce-eager to disable cupy
-	// TODO: remove --enforce-eager when https://github.com/kubeagi/arcadia/issues/878 is fixed
-	extraAgrs = fmt.Sprintf("%s --trust-remote-code --enforce-eager", extraAgrs)
+	extraAgrs = fmt.Sprintf("%s --trust-remote-code", extraAgrs)
 	if runner.modelFileFromRemote {
 		m := arcadiav1alpha1.Model{}
 		if err := runner.c.Get(ctx, types.NamespacedName{Namespace: *model.Namespace, Name: model.Name}, &m); err != nil {