diff --git a/vllm/model_executor/layers/quantization/utils/fp8_utils.py b/vllm/model_executor/layers/quantization/utils/fp8_utils.py
index 78b1234021af..a974e2c57aa0 100644
--- a/vllm/model_executor/layers/quantization/utils/fp8_utils.py
+++ b/vllm/model_executor/layers/quantization/utils/fp8_utils.py
@@ -305,6 +305,11 @@ def run_deepgemm(
         )
         return output
 
+    from vllm.model_executor.layers.batch_invariant import vllm_is_batch_invariant
+
+    if vllm_is_batch_invariant():
+        return run_deepgemm(input, weight, weight_scale)
+
     condition = input.shape[0] < 32
 
     # PyTorch's torch.compile cannot handle input-dependent control flow in standard