[CI] Split Kimi K2.5 performance batches by config (#34669)

test/registered/gb300/test_kimi_k25_nvfp4.py

@@ -38,6 +38,11 @@ DP_EAGLE_ARGS = [
     "--speculative-num-draft-tokens=2",
 ]
 
+PERFORMANCE_BATCH_SIZES = {
+    "TP4+EAGLE3": [1, 8],
+    "TP4+DP4+DPA+EAGLE3": [16],
+}
+
 
 class TestKimiK25Nvfp4(unittest.TestCase):
     """Kimi-K2.5 NVFP4 + EAGLE3 on GB300 (4x GB300 NVL4, tp=4)."""
@@ -60,13 +65,11 @@ class TestKimiK25Nvfp4(unittest.TestCase):
             ),
         ]
 
-        run_combined_tests(
-            models=variants,
-            test_name="Kimi-K2.5-NVFP4",
+        failures = []
         # Pinned to what `ns eval --benchmarks=mmmu-pro:1` sent implicitly --
         # its `:1` suffix means temperature 0.7, not greedy -- so the baseline
         # carries over unchanged. Do not "simplify" these away.
-            accuracy_params=AccuracyTestParams(
+        accuracy_params = AccuracyTestParams(
             dataset="mmmu_pro_vision",
             baseline_accuracy=0.69,
             repeat=1,
@@ -74,11 +77,23 @@ class TestKimiK25Nvfp4(unittest.TestCase):
             temperature=0.7,
             seed=0,
             sgl_eval_thinking=False,
-            ),
+        )
+        for variant in variants:
+            try:
+                run_combined_tests(
+                    models=[variant],
+                    test_name=f"Kimi-K2.5-NVFP4 ({variant.variant})",
+                    accuracy_params=accuracy_params,
                     performance_params=PerformanceTestParams(
+                        batch_sizes=PERFORMANCE_BATCH_SIZES[variant.variant],
                         result_dir="performance_results_gb300",
                     ),
                 )
+            except AssertionError as e:
+                failures.append(f"{variant.variant}: {e}")
+
+        if failures:
+            raise AssertionError("Kimi-K2.5-NVFP4 failures:\n" + "\n".join(failures))
 
 
 if __name__ == "__main__":