Stabilize hybrid nanov3 gb200 perf

tests/performance_tests/test_cases/hybrid/hybrid_nanov3_3b_perf_gb200_4gpu/baseline_values.json

@@ -3,50 +3,50 @@
     "batch_1": {
       "batch_size": 1,
       "dataset": "gsm8k",
-      "num_input_tokens_avg": 60.2,
+      "num_input_tokens_avg": 66.2,
       "num_output_tokens": 128,
-      "num_iters": 5,
-      "throughput_tok_per_sec": 205.28963339383708,
-      "avg_latency_ms": 623.4671195968986,
-      "p50_latency_ms": 608.3062621764839,
-      "p99_latency_ms": 664.121120236814,
-      "tpot_ms_per_tok": 4.8711665731389076
+      "num_iters": 10,
+      "throughput_tok_per_sec": 211.27146353361152,
+      "avg_latency_ms": 605.8221792030963,
+      "p50_latency_ms": 604.2382380110212,
+      "p99_latency_ms": 614.8939310005517,
+      "tpot_ms_per_tok": 4.733246900809718
     },
     "batch_8": {
       "batch_size": 8,
       "dataset": "gsm8k",
-      "num_input_tokens_avg": 59.625,
+      "num_input_tokens_avg": 58.925,
       "num_output_tokens": 128,
-      "num_iters": 5,
-      "throughput_tok_per_sec": 1479.8669689494195,
-      "avg_latency_ms": 689.8111295537092,
-      "p50_latency_ms": 690.9317341633141,
-      "p99_latency_ms": 697.1075707115233,
-      "tpot_ms_per_tok": 5.4058913185144775
+      "num_iters": 10,
+      "throughput_tok_per_sec": 1479.7306591133317,
+      "avg_latency_ms": 689.6378291000474,
+      "p50_latency_ms": 689.3478269921616,
+      "p99_latency_ms": 702.6289499917766,
+      "tpot_ms_per_tok": 5.406389298437375
     },
     "batch_32": {
       "batch_size": 32,
       "dataset": "gsm8k",
-      "num_input_tokens_avg": 62.475,
+      "num_input_tokens_avg": 61.79375,
       "num_output_tokens": 128,
-      "num_iters": 5,
-      "throughput_tok_per_sec": 4169.359459179444,
-      "avg_latency_ms": 976.2204706174089,
-      "p50_latency_ms": 951.845585834235,
-      "p99_latency_ms": 1092.4915480427444,
-      "tpot_ms_per_tok": 7.675039850437315
+      "num_iters": 10,
+      "throughput_tok_per_sec": 4150.460740435596,
+      "avg_latency_ms": 980.5532394029342,
+      "p50_latency_ms": 969.1448689991375,
+      "p99_latency_ms": 1130.6974129984155,
+      "tpot_ms_per_tok": 7.709987396879114
     },
     "batch_128": {
       "batch_size": 128,
       "dataset": "gsm8k",
-      "num_input_tokens_avg": 61.75,
+      "num_input_tokens_avg": 61.88671875,
       "num_output_tokens": 128,
-      "num_iters": 5,
-      "throughput_tok_per_sec": 12046.630331615843,
-      "avg_latency_ms": 1346.5355291846208,
-      "p50_latency_ms": 1354.1678641922772,
-      "p99_latency_ms": 1424.3129258975387,
-      "tpot_ms_per_tok": 10.62537792531657
+      "num_iters": 10,
+      "throughput_tok_per_sec": 11512.760737650928,
+      "avg_latency_ms": 1409.1983075107692,
+      "p50_latency_ms": 1402.4255399999674,
+      "p99_latency_ms": 1589.2388540014508,
+      "tpot_ms_per_tok": 11.118097814835437
     }
   }
 }

tests/performance_tests/test_cases/hybrid/hybrid_nanov3_3b_perf_gb200_4gpu/model_config.yaml

@@ -13,10 +13,12 @@ EP: 4
 DATASET: gsm8k
 NUM_OUTPUT_TOKENS: 128
 # 2 warmup iters left the first timed iteration cold at batch 128 (~2.5x slower
-# than steady state), poisoning the mean/tail metrics (throughput, avg/p99
-# latency, tpot). Warm up more so timing starts at steady state.
+# than steady state), poisoning the mean/tail metrics. Warm up more so timing
+# starts at steady state.
 NUM_WARMUP_ITERS: 5
-NUM_TIMED_ITERS: 5
+# 5 timed iters left batch-128 p99/throughput sensitive to a single outlier
+# iteration (~10–14% run-to-run swing on GB200). 10 iters stabilizes the mean.
+NUM_TIMED_ITERS: 10
 BATCH_SIZES:
   - 1
   - 8