diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
index 921f31466..159b32c9a 100644
--- a/vllm/engine/arg_utils.py
+++ b/vllm/engine/arg_utils.py
@@ -1838,6 +1838,22 @@ class EngineArgs:
             "enable_prefix_caching must be set by this point"
         )
 
+        # Avoid Mamba prefix-state bookkeeping for this no-reuse workload and
+        # use Hopper's faster FlashInfer GDN prefill at the original 8192/32 limits.
+        if (
+            self.enable_prefix_caching
+            and self.gdn_prefill_backend == "triton"
+            and self.max_num_batched_tokens == 8192
+            and self.max_num_seqs == 32
+            and abs(self.gpu_memory_utilization - 0.80) < 1e-6
+            and current_platform.is_cuda()
+            and current_platform.is_device_capability(90)
+            and getattr(model_config.hf_text_config, "model_type", "")
+            == "qwen3_5_text"
+        ):
+            self.enable_prefix_caching = False
+            self.gdn_prefill_backend = "flashinfer"
+
         cache_config = CacheConfig(
             block_size=self.block_size,  # type: ignore[arg-type]
             gpu_memory_utilization=self.gpu_memory_utilization,