This study proposes a language-aware multi-bin batching strategy that groups code summarization requests by predicted generation length and programming language, reaching 4.67 requests per second and 44.15 tokens per second with a latency of 2.12 seconds.