That's just an implementation artifact and not a fundamental fact of life.

https://docs.vllm.ai/en/latest/features/batch_invariance/