router=http://127.0.0.1:8080 tokens=128 streams=1 4 8 16 32 record=no
#	at	model	streams	ok	tokens	wall_s	agg_tok_s	per_stream_tok_s	ttft_s

-- deepseek-v4-flash-vllm is cold; asking the router to start it --
   ready after ~370s
BENCH	2026-09-05T05:59Z	deepseek-v4-flash-vllm	1	1/1	128	8.31	15.40	15.40	0.44
BENCH	2026-09-05T05:59Z	deepseek-v4-flash-vllm	4	4/4	512	17.25	29.67	10.88	0.44
BENCH	2026-09-05T05:59Z	deepseek-v4-flash-vllm	8	8/8	1024	37.73	27.14	7.53	0.44
BENCH	2026-09-05T05:59Z	deepseek-v4-flash-vllm	16	16/16	2048	65.58	31.23	5.15	0.44
BENCH	2026-09-05T05:59Z	deepseek-v4-flash-vllm	32	32/32	4096	136.09	30.10	3.12	0.44