- added more useful fields to types.LLMRequest:

1. cleaner API declaration
2. data fields are preserved, after-read transformations are done in plugins
3. prefix-cache scorer does not need naive templating
- minor bugfixes and improvements

Signed-off-by: Maroon Ayoub <[email protected]>

Author: vMaroon

pkg/epp/requestcontrol/director.go

@@ -100,10 +100,11 @@ func (d *Director) HandleRequest(ctx context.Context, reqCtx *handlers.RequestCo
 	}
 	reqCtx.Request.Body["model"] = reqCtx.TargetModelName
 
-	prompt, err := requtil.ExtractPromptFromRequestBody(requestBodyMap)
+	requestData, err := requtil.ExtractRequestData(reqCtx.Request.Body)
 	if err != nil {
-		return reqCtx, err
+		return reqCtx, errutil.Error{Code: errutil.BadRequest, Msg: fmt.Errorf("failed to extract request data: %w", err).Error()}
 	}
+
 	infObjective := d.datastore.ObjectiveGet(reqCtx.ObjectiveKey)
 	if infObjective == nil {
 		logger.V(logutil.VERBOSE).Info("No associated InferenceObjective found, using default", "objectiveKey", reqCtx.ObjectiveKey)
@@ -121,7 +122,7 @@ func (d *Director) HandleRequest(ctx context.Context, reqCtx *handlers.RequestCo
 	reqCtx.SchedulingRequest = &schedulingtypes.LLMRequest{
 		RequestId:   reqCtx.Request.Headers[requtil.RequestIdHeaderKey],
 		TargetModel: reqCtx.TargetModelName,
-		Prompt:      prompt,
+		Data:        requestData,
 		Headers:     reqCtx.Request.Headers,
 	}
 

pkg/epp/scheduling/framework/plugins/multi/prefix/plugin.go

@@ -250,33 +250,39 @@ func (p *Plugin) matchLongestPrefix(ctx context.Context, hashes []BlockHash) map
 // For block i, hash(i) = hash(block i content, hash(i-1)).
 func hashPrompt(ctx context.Context, request *types.LLMRequest, cacheBlockSize int, maxPrefixBlocks int) []BlockHash {
 	loggerDebug := log.FromContext(ctx).V(logutil.DEBUG)
-	prompt := []byte(request.Prompt)
-	if len(prompt) < cacheBlockSize {
-		loggerDebug.Info("Request body too small for prefix cache", "size", len(prompt), "block size", cacheBlockSize)
+	if request == nil || request.Data == nil {
+		loggerDebug.Info("Request or request data is nil, skipping hashing")
 		return nil
 	}
-	if len(prompt) > cacheBlockSize*maxPrefixBlocks {
-		loggerDebug.Info("Truncating input", "size", len(prompt), "max prefix blocks", maxPrefixBlocks, "block size", cacheBlockSize)
-		prompt = prompt[:maxPrefixBlocks*cacheBlockSize]
+
+	userInput, err := getUserInputBytes(request)
+	if err != nil {
+		loggerDebug.Error(err, "Failed to get user input bytes")
+		return nil
+	}
+
+	if len(userInput) < cacheBlockSize {
+		loggerDebug.Info("Request body too small for prefix cache", "size", len(userInput), "block size", cacheBlockSize)
+		return nil
+	}
+	if len(userInput) > cacheBlockSize*maxPrefixBlocks {
+		loggerDebug.Info("Truncating input", "size", len(userInput), "max prefix blocks", maxPrefixBlocks, "block size", cacheBlockSize)
+		userInput = userInput[:maxPrefixBlocks*cacheBlockSize]
 	}
 	// Split the body into blocks of size cacheBlockSize. The +1 is to account for the model.
 	// If the last block is smaller than cacheBlockSize, it will be ignored.
-	res := make([]BlockHash, 0, 1+len(prompt)/cacheBlockSize)
+	res := make([]BlockHash, 0, 1+len(userInput)/cacheBlockSize)
 	// Add the model to the first block hash so that different models have different hashes even with the same body.
-
-	firstBlockSize := cacheBlockSize
-	if len(prompt) < cacheBlockSize {
-		firstBlockSize = len(prompt)
-	}
-	firstBlock := prompt[0:firstBlockSize]
-	firstBlockWithModel := append([]byte(request.TargetModel), firstBlock...)
-	res = append(res, BlockHash(xxhash.Sum64(firstBlockWithModel)))
-
-	for i := cacheBlockSize; i+cacheBlockSize <= len(prompt); i += cacheBlockSize {
-		block := prompt[i : i+cacheBlockSize]
-		prevBlockHash := res[len(res)-1]
-		block = append(block, toBytes(prevBlockHash)...)
-		res = append(res, BlockHash(xxhash.Sum64(block)))
+	h := xxhash.New()
+	_, _ = h.Write([]byte(request.TargetModel))
+	prevBlockHash := BlockHash(h.Sum64())
+	for i := 0; i+cacheBlockSize <= len(userInput); i += cacheBlockSize {
+		h.Reset()
+		_, _ = h.Write(userInput[i : i+cacheBlockSize])
+		_, _ = h.Write(toBytes(prevBlockHash))
+		res = append(res, BlockHash(h.Sum64()))
+
+		prevBlockHash = res[len(res)-1]
 	}
 	return res
 }
@@ -286,3 +292,12 @@ func toBytes(i BlockHash) []byte {
 	binary.LittleEndian.PutUint64(bytes, uint64(i))
 	return bytes
 }
+
+func getUserInputBytes(request *types.LLMRequest) ([]byte, error) {
+	if request.Data.Completions != nil { // assumed to be valid if not nil
+		return []byte(request.Data.Completions.Prompt), nil
+	}
+
+	// must be chat-completions request at this point, return bytes of entire messages
+	return json.Marshal(request.Data.ChatCompletions.Messages)
+}