Offline Throughput Benchmark Summary

• model: gpt2
• dtype: fp16
• num_prompts: 128
• input_len: 256
• output_len: 64
• sampling: temperature=0.7, top_p=0.9, top_k=50
• wall_s: 18.06826117899618
• run_start_time: 2025-12-30T18:51:52+08:00
• run_end_time: 2026-01-01T08:00:24+08:00
• versions: git_rev=a547c7c, rosellm=0.1.0, vllm=0.7.2, sglang=0.4.6, tensorrt_llm=1.1.0, torch=2.6.0, transformers=4.51.3, python=3.11.11

backend	req/s	output tok/s	total tok/s	total latency (s)
roseinfer	201.13	12872.49	64362.44	0.636
roseinfer (+idle keepalive)	201.53	12897.97	64489.86	0.635
roseinfer (in-proc)	204.11	13062.83	65314.13	0.627
roseinfer (in-proc, +single-chunk fp)	185.10	11846.11	59230.57	0.692
roseinfer (+eager prefill)	201.03	12865.71	64328.54	0.637
roseinfer (+pprio1)	203.70	13036.67	65183.33	0.628
roseinfer (+single-chunk fp)	183.49	11743.51	58717.56	0.698
SGLang	243.20	15564.48	77822.40	0.526
TensorRT-LLM	248.69	15916.24	79581.21	0.515
vLLM	140.44	8988.14	44940.70	0.911