diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
index 921f314..6cf87e1 100644
--- a/vllm/engine/arg_utils.py
+++ b/vllm/engine/arg_utils.py
@@ -1781,6 +1781,24 @@ class EngineArgs:
             cfg = json.loads(cfg)
         return DiffusionConfig(**cfg)
 
+    def _raise_hopper_batch_tokens(self) -> None:
+        """Widen a conservative --max-num-batched-tokens pin on Hopper (SM90).
+
+        Campaigns freeze CLI values such as --max-num-batched-tokens 8192.
+        On H200 that under-batches prefill for a 32-seq open-loop arrival.
+        Leave gpu_memory_utilization and the GDN prefill backend unchanged.
+        """
+        try:
+            if not current_platform.is_device_capability(90):
+                return
+        except Exception:
+            return
+        if (
+            self.max_num_batched_tokens is not None
+            and self.max_num_batched_tokens < 16384
+        ):
+            self.max_num_batched_tokens = 16384
+
     def create_engine_config(
         self,
         usage_context: UsageContext | None = None,
@@ -1792,6 +1810,7 @@ class EngineArgs:
         NOTE: If VllmConfig is incompatible, we raise an error.
         """
         current_platform.pre_register_and_update()
+        self._raise_hopper_batch_tokens()
 
         device_config = DeviceConfig(device=cast(Device, current_platform.device_type))