The vllm/vllm-openai version 0.9.1 is nearly 30% faster compared to lmsysorg/sglang:v0.4.7.post, but it stops running every two to three hours

Model Support DeepSeek
1

bug