Support cache_config_info metric

pkg/llm-d-inference-sim/metrics.go

@@ -49,6 +49,7 @@ const (
 	reqRunningMetricName             = "vllm:num_requests_running"
 	reqWaitingMetricName             = "vllm:num_requests_waiting"
 	gpuCacheUsageMetricName          = "vllm:gpu_cache_usage_perc"
+	cacheConfigName                  = "vllm:cache_config_info"
 )
 
 // createAndRegisterPrometheus creates and registers prometheus metrics used by vLLM simulator
@@ -85,7 +86,6 @@ func (s *VllmSimulator) createAndRegisterPrometheus() error {
 		return err
 	}
 
-	// not supported for now, reports constant value
 	s.metrics.waitingRequests = prometheus.NewGaugeVec(
 		prometheus.GaugeOpts{
 			Subsystem: "",
@@ -288,14 +288,27 @@ func (s *VllmSimulator) createAndRegisterPrometheus() error {
 		return err
 	}
 
-	s.setInitialPrometheusMetrics()
+	cacheConfig := prometheus.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Subsystem: "",
+			Name:      cacheConfigName,
+			Help:      "Information of the LLMEngine CacheConfig.",
+		},
+		[]string{vllmapi.PromLabelCacheBlockSize, vllmapi.PromLabelCacheNumGPUBlocks},
+	)
+	if err := s.metrics.registry.Register(cacheConfig); err != nil {
+		s.logger.Error(err, "prometheus cache config register failed")
+		return err
+	}
+
+	s.setInitialPrometheusMetrics(cacheConfig)
 
 	return nil
 }
 
 // setInitialPrometheusMetrics sends the default values to prometheus or
 // the fake metrics if set
-func (s *VllmSimulator) setInitialPrometheusMetrics() {
+func (s *VllmSimulator) setInitialPrometheusMetrics(cacheConfig *prometheus.GaugeVec) {
 	var nRunningReqs, nWaitingReqs, kvCacheUsage float64
 	modelName := s.getDisplayedModelName(s.config.Model)
 	if s.config.FakeMetrics != nil {
@@ -352,6 +365,8 @@ func (s *VllmSimulator) setInitialPrometheusMetrics() {
 	s.metrics.waitingRequests.WithLabelValues(modelName).Set(nWaitingReqs)
 	s.metrics.kvCacheUsagePercentage.WithLabelValues(modelName).Set(kvCacheUsage)
 
+	cacheConfig.WithLabelValues(strconv.Itoa(s.config.TokenBlockSize), strconv.Itoa(s.config.KVCacheSize)).Set(1)
+
 	if s.config.FakeMetrics != nil && len(s.config.FakeMetrics.LoraMetrics) != 0 {
 		for _, metrics := range s.config.FakeMetrics.LoraMetrics {
 			s.metrics.loraInfo.WithLabelValues(

pkg/llm-d-inference-sim/metrics_test.go

@@ -598,6 +598,24 @@ var _ = Describe("Simulator metrics", Ordered, func() {
 			}()
 			wg.Wait()
 		})
+
+		It("Should send correct kv cache config metrics", func() {
+			ctx := context.TODO()
+			args := []string{"cmd", "--model", qwenModelName, "--mode", common.ModeRandom,
+				"--kv-cache-size", "16", "--block-size", "8"}
+
+			client, err := startServerWithArgs(ctx, args)
+			Expect(err).NotTo(HaveOccurred())
+
+			metricsResp, err := client.Get(metricsUrl)
+			Expect(err).NotTo(HaveOccurred())
+			Expect(metricsResp.StatusCode).To(Equal(http.StatusOK))
+
+			data, err := io.ReadAll(metricsResp.Body)
+			Expect(err).NotTo(HaveOccurred())
+			metrics := string(data)
+			Expect(metrics).To(ContainSubstring("vllm:cache_config_info{block_size=\"8\",num_gpu_blocks=\"16\"} 1"))
+		})
 	})
 
 	Context("fake metrics", func() {

pkg/vllm-api/vllm-models.go

@@ -26,6 +26,8 @@ const (
 	PromLabelMaxLora             = "max_lora"
 	PromLabelModelName           = "model_name"
 	PromLabelFinishReason        = "finish_reason"
+	PromLabelCacheBlockSize      = "block_size"
+	PromLabelCacheNumGPUBlocks   = "num_gpu_blocks"
 )
 
 // modelInfo defines data about model returned by /models API