[CI] Trim Qwen3.5 FP8 GB300 performance batches (#34882)

test/registered/gb300/test_qwen35_fp8.py

@@ -36,6 +36,11 @@ DP_MTP_ARGS = [
     "--speculative-num-draft-tokens=2",
 ]
 
+PERFORMANCE_BATCH_SIZES = {
+    "TP4+MTP": [1, 4],
+    "TP4+DP4+DPA+MTP": [16],
+}
+
 
 class TestQwen35Fp8(unittest.TestCase):
     """Qwen3.5-397B FP8 on GB300 (4x GB300 NVL4, tp=4)."""
@@ -58,9 +63,7 @@ class TestQwen35Fp8(unittest.TestCase):
             ),
         ]
 
-        run_combined_tests(
-            models=variants,
-            test_name="Qwen3.5-397B-FP8",
+        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.
@@ -72,11 +75,23 @@ class TestQwen35Fp8(unittest.TestCase):
             temperature=0.7,
             seed=0,
             sgl_eval_thinking=False,
-            ),
+        )
+        for variant in variants:
+            try:
+                run_combined_tests(
+                    models=[variant],
+                    test_name=f"Qwen3.5-397B-FP8 ({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("Qwen3.5-397B-FP8 failures:\n" + "\n".join(failures))
 
 
 if __name__ == "__main__":