benchmark: fix TTFT unit label and achieved RPS extraction

- TTFT mean in PR comment showed "ms" but value was already divided
  by 1000 (should be "s")
- Achieved RPS was always 0.00 because GuideLLM may not expose
  rate.completed_rate; add fallback: completed_requests / duration

Made-with: Cursor

Author: kahilam

.github/workflows/ci-benchmark.yaml

@@ -1074,7 +1074,7 @@ jobs:
             | **Avg vLLM Queue Depth** | ${hpa.avg_queue_depth.toFixed(1)} | **${wva.avg_queue_depth.toFixed(1)}** | ${delta(hpa.avg_queue_depth, wva.avg_queue_depth)} |
             | **Avg EPP Queue Depth** | ${(hpa.avg_epp_queue_depth||0).toFixed(1)} | **${(wva.avg_epp_queue_depth||0).toFixed(1)}** | ${delta(hpa.avg_epp_queue_depth||0, wva.avg_epp_queue_depth||0)} |
             | **Avg KV Cache** | ${hpa.avg_kv_cache.toFixed(3)} | ${wva.avg_kv_cache.toFixed(3)} | ${delta(hpa.avg_kv_cache, wva.avg_kv_cache)} |
-            | **TTFT mean** | ${fmtM(hpa.ttft, 1000)}ms | **${fmtM(wva.ttft, 1000)}ms** | ${hpa.ttft && wva.ttft ? delta(hpa.ttft.mean, wva.ttft.mean) : '—'} |
+            | **TTFT mean** | ${fmtM(hpa.ttft, 1000)}s | **${fmtM(wva.ttft, 1000)}s** | ${hpa.ttft && wva.ttft ? delta(hpa.ttft.mean, wva.ttft.mean) : '—'} |
             | **TTFT p50** | ${fmtP(hpa.ttft, 'p50', 1000)}s | **${fmtP(wva.ttft, 'p50', 1000)}s** | — |
             | **TTFT p99** | ${fmtP(hpa.ttft, 'p99', 1000)}s | **${fmtP(wva.ttft, 'p99', 1000)}s** | — |
             | **ITL mean** | ${fmtM(hpa.itl, 1, 2)}ms | **${fmtM(wva.itl, 1, 2)}ms** | ${hpa.itl && wva.itl ? delta(hpa.itl.mean, wva.itl.mean) : '—'} |

test/benchmark/prefill_heavy_benchmark_test.go

@@ -651,7 +651,7 @@ var _ = Describe("Prefill Heavy Workload Benchmark", Label("benchmark", "phase4"
 		}
 
 		// Extract error counts and achieved RPS from GuideLLM output
-		var errorCount, incompleteCount int
+		var errorCount, incompleteCount, completedCount int
 		var achievedRPS float64
 		if guidellmRaw != nil {
 			var parsed map[string]interface{}
@@ -666,6 +666,9 @@ var _ = Describe("Prefill Heavy Workload Benchmark", Label("benchmark", "phase4"
 								if f, ok := rt["incomplete"].(float64); ok {
 									incompleteCount = int(f)
 								}
+								if f, ok := rt["successful"].(float64); ok {
+									completedCount = int(f)
+								}
 							}
 						}
 						if rateObj, ok := bm["rate"].(map[string]interface{}); ok {
@@ -677,6 +680,9 @@ var _ = Describe("Prefill Heavy Workload Benchmark", Label("benchmark", "phase4"
 				}
 			}
 		}
+		if achievedRPS == 0 && completedCount > 0 && loadDuration > 0 {
+			achievedRPS = float64(completedCount) / loadDuration
+		}
 
 		result := PrefillResult{
 			AutoscalerType:   autoscalerType,