[CI] Lower GSM8K baselines for B200 nightly after eval unification (#22136)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

test/registered/backends/test_flashinfer_trtllm_gen_moe_backend.py

@@ -59,7 +59,7 @@ class FlashinferTrtllmGenMoeBackendFP8Base:
         )
         metrics = run_eval(args)
         print(f"{metrics=}")
-        self.assertGreater(metrics["score"], 0.93)
+        self.assertGreater(metrics["score"], 0.89)
 
 
 class FlashinferTrtllmGenMoeBackendBF16Base:

test/registered/perf/test_dpsk_v3_fp4_4gpu_perf.py

@@ -63,7 +63,10 @@ class TestDeepseekR1FP4Unified(unittest.TestCase):
             models=variants,
             test_name="DeepSeek-V3-0324-FP4 Unified",
             accuracy_params=AccuracyTestParams(
-                dataset="gsm8k", baseline_accuracy=0.935
+                dataset="gsm8k",
+                baseline_accuracy=0.935,
+                num_examples=200,
+                api="completion",
             ),
             performance_params=PerformanceTestParams(
                 profile_dir="performance_profiles_deepseek_v3_fp4",