Feat/genai Add GenAI Cost Visibility Plugin with Token- and MIG-Aware Efficiency Metrics (Sidecar Architecture)

pkg/plugins/genai/cmd/main/main.go

@@ -0,0 +1,78 @@
+package main
+
+import (
+	"fmt"
+	"log"
+	"os"
+
+	"github.com/hashicorp/go-plugin"
+	"github.com/opencost/opencost-plugins/pkg/plugins/genai/genaiprovider"
+	"github.com/opencost/opencost-plugins/pkg/plugins/genai/internal"
+)
+
+type GenAISource struct {
+	Config *genaiprovider.Config
+}
+
+// GetCustomCosts is called by OpenCost to retrieve the enriched GenAI costs.
+func (s *GenAISource) GetCustomCosts(req *genaiprovider.CustomCostRequest) ([]*genaiprovider.GenAIWorkloadData, error) {
+	log.Printf("GenAI Plugin: Fetching costs for window %s - %s", req.Start, req.End)
+
+	// 1. Call your internal math logic (from internal/join.go)
+	// This performs the Prometheus fetch and the MIG-to-Node join.
+	workloads, err := internal.CalculateGenAIWorkloads(req.Start, req.End, s.Config)
+	if err != nil {
+		return nil, fmt.Errorf("failed to calculate GenAI workloads: %w", err)
+	}
+
+	var responses []*genaiprovider.GenAIWorkloadData
+	for _, w := range workloads {
+		// Map internal results to the genaiprovider format
+		responses = append(responses, &genaiprovider.GenAIWorkloadData{
+			PodName:       w.PodName,
+			ModelName:     w.ModelName,
+			TotalTokens:   w.TotalTokens,
+			TotalCost:     w.TotalCost,
+			TenantID:      w.TenantID,
+			WorkflowPhase: w.WorkflowPhase,
+			MIGProfile:    w.MIGProfile,
+			GPUEfficiency: w.GPUEfficiency,
+		})
+	}
+
+	return responses, nil
+}
+
+func main() {
+	// 1. Load plugin configuration (e.g., Prometheus URL, Cluster ID)
+	configPath := os.Getenv("PLUGIN_CONFIG_PATH")
+	if configPath == "" {
+		configPath = "config/genai-config.json"
+	}
+
+	cfg, err := genaiprovider.LoadConfig(configPath)
+	if err != nil {
+		log.Fatalf("GenAI Plugin: Failed to load config: %v", err)
+	}
+
+	// 2. Create the source instance
+	source := &GenAISource{Config: cfg}
+
+	// 3. Define the Handshake (Security "Cookie" required by go-plugin)
+	var handshakeConfig = plugin.HandshakeConfig{
+		ProtocolVersion:  1,
+		MagicCookieKey:   "OPENCOST_PLUGIN_MAGIC_COOKIE",
+		MagicCookieValue: "genai-visibility",
+	}
+
+	// 4. Start the gRPC Server
+	plugin.Serve(&plugin.ServeConfig{
+		HandshakeConfig: handshakeConfig,
+		Plugins: map[string]plugin.Plugin{
+			"customcost": &genaiprovider.CustomCostPlugin{
+				Impl: source,
+			},
+		},
+		GRPCServer: plugin.DefaultGRPCServer,
+	})
+}

pkg/plugins/genai/config.json

@@ -0,0 +1,4 @@
+{
+  "prometheusUrl": "http://your-prometheus-endpoint:9090",
+  "metrics": { "inputTokensMetric": "genai_input_tokens_total" }
+}

pkg/plugins/genai/config/genai-config.json

@@ -0,0 +1,5 @@
+{
+  "prometheus_url": "http://prometheus:9090",
+  "cluster_id": "test-cluster",
+  "log_level": "info"
+}

pkg/plugins/genai/dockerfile

@@ -0,0 +1,5 @@
+FROM debian:bookworm-slim
+WORKDIR /app
+COPY genai-plugin /app/genai-plugin
+RUN chmod +x /app/genai-plugin
+ENTRYPOINT ["/app/genai-plugin"]

pkg/plugins/genai/genaiplugin/config.go

@@ -0,0 +1,44 @@
+package genaiplugin
+
+type MetricMapping struct {
+	InputTokens  string
+	OutputTokens string
+	GPUActiveSec string
+	GPUUtil      string
+}
+
+func DefaultMetricMapping() MetricMapping {
+	return MetricMapping{
+		InputTokens:  "vllm:prompt_tokens_total",
+		OutputTokens: "vllm:generation_tokens_total",
+		GPUUtil:      "DCGM_FI_DEV_GPU_UTIL",
+		GPUActiveSec: "DCGM_FI_DEV_GPU_UTIL",
+	}
+}
+
+func GetMapping(annotations map[string]string) MetricMapping {
+
+	// "fallback" if no annotations are present
+	m := MetricMapping{
+		InputTokens:  "llm_tokens_input_total",
+		OutputTokens: "llm_tokens_output_total",
+		GPUActiveSec: "llm_gpu_active_seconds_total",
+		GPUUtil:      "llm_gpu_utilization_percent",
+	}
+
+	// Override only if specific annotations exist
+	if val, ok := annotations["opencost.io/metric-input"]; ok {
+		m.InputTokens = val
+	}
+	if val, ok := annotations["opencost.io/metric-output"]; ok {
+		m.OutputTokens = val
+	}
+	if val, ok := annotations["opencost.io/metric-gpu-sec"]; ok {
+		m.GPUActiveSec = val
+	}
+	if val, ok := annotations["opencost.io/metric-gpu-util"]; ok {
+		m.GPUUtil = val
+	}
+
+	return m
+}