Speed test matrix

Tess-4-9B on RTX 3090 24GB

vllm · 0.32.0 · BF166 contexts (0–248K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 31, 2026, 16:11 UTC

Hardware

Type: Discrete GPU
GPU: RTX 3090
VRAM: 24 GB
CPU: Intel Core i7-9700K (8C)
RAM: 32 GB
OS: Ubuntu 24.04.5

Engine

Engine: vllm
Version: 0.32.0
Quantization: BF16
Backend: cuda
Repository: github.com/vllm-project/vllm
Build: vllm/vllm-openai:v0.32.0-x86_64

Run

Served model: bench
Revision: main
Benchmark tool: llm-inference-bench v0.8.0
Decode test: Fixed duration
Prefill mode: Measured during decode warm-up
Time per measurement: 30 s
Max output: 8,192 tokens

Launch command

# Untuned out-of-the-box baseline. Spec decode: none. KV cache: int4_per_token_head
# 1x RTX 3090 24GB (Zagitta/gpu-unlock), 420 W limit, PCIe Gen3 x16 | Intel Core i7-9700K 8C | 32 GB DDR4 | Ubuntu 24.04.5, NVIDIA driver 580.178.04
vllm serve migtissera/Tess-4-9B --max-model-len 262144 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --language-model-only --kv-cache-dtype int4_per_token_head --attention-backend TRITON_ATTN

Prefill

prompt tok/s by context length

8K context6,4991,260 ms TTFT

16K context5,4632,969 ms TTFT

32K context4,3177,479 ms TTFT

64K context3,02421,305 ms TTFT

128K context1,88568,265 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests

Context	1	2	4	8	16	32	64	128
0	49.140 ms TTFT	96.247 ms TTFT	189.592 ms TTFT	372.5137 ms TTFT	692.8244 ms TTFT	1,184.7349 ms TTFT	Over capacity	Over capacity
16K	46.1473 ms TTFT	89.4704 ms TTFT	165.11,162 ms TTFT	304.22,086 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
32K	43.5608 ms TTFT	83.5888 ms TTFT	145.51,443 ms TTFT	255.72,575 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
64K	39.0688 ms TTFT	73.9944 ms TTFT	119.31,496 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
128K	32.31,945 ms TTFT	60.02,740 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
248K	22.5Requests queued2,995 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity

Requests queued: the server could not run all requests at once, so some waited in a queue. Over capacity: not run because this many requests at this context would exceed the server's KV cache.

TTFT is measured after the benchmark warms the server's prompt cache, so it mostly excludes prefill.

All measurements

Latencies are medians. Total tok/s is output across all simultaneous requests.

Phase	Context	Requests at once	Status	Total tok/s	Per request tok/s	TTFT	Time between tokens	Average active requests	Completed requests	Output tokens	Measured for
Prefill	8K	1	Valid	6,499.0	—	1,260 ms	—	—	1	—	—
Prefill	16K	1	Valid	5,463.0	—	2,969 ms	—	—	1	—	—
Prefill	32K	1	Valid	4,317.0	—	7,479 ms	—	—	1	—	—
Prefill	64K	1	Valid	3,024.0	—	21,305 ms	—	—	1	—	—
Prefill	128K	1	Valid	1,885.0	—	68,265 ms	—	—	1	—	—
Sustained decode	0	1	Valid	49.1	49.1	40 ms	20.4 ms	1.0	0	1,472	30.0 s
Sustained decode	0	2	Valid	96.2	48.1	47 ms	20.8 ms	2.0	0	2,884	30.0 s
Sustained decode	0	4	Valid	189.5	47.4	92 ms	21.1 ms	4.0	0	5,684	30.0 s
Sustained decode	0	8	Valid	372.5	46.7	137 ms	21.4 ms	8.0	0	11,177	30.0 s
Sustained decode	0	16	Valid	692.8	43.5	244 ms	23.0 ms	16.0	0	20,784	30.0 s
Sustained decode	0	32	Valid	1,184.7	37.2	349 ms	26.9 ms	32.0	0	35,552	30.0 s
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	46.1	46.2	473 ms	21.7 ms	1.0	0	1,384	30.0 s
Sustained decode	16K	2	Valid	89.4	44.4	704 ms	22.5 ms	2.0	0	2,680	30.0 s
Sustained decode	16K	4	Valid	165.1	40.6	1,162 ms	24.7 ms	4.0	0	4,952	30.0 s
Sustained decode	16K	8	Valid	304.2	36.5	2,086 ms	27.4 ms	8.0	0	9,120	30.0 s
Sustained decode	16K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	43.5	43.5	608 ms	23.0 ms	1.0	0	1,303	30.0 s
Sustained decode	32K	2	Valid	83.5	41.5	888 ms	24.1 ms	2.0	0	2,505	30.0 s
Sustained decode	32K	4	Valid	145.5	35.6	1,443 ms	28.1 ms	4.0	0	4,364	30.0 s
Sustained decode	32K	8	Valid	255.7	30.4	2,575 ms	32.9 ms	8.0	0	7,664	30.0 s
Sustained decode	32K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	1	Valid	39.0	39.0	688 ms	25.6 ms	1.0	0	1,170	30.0 s
Sustained decode	64K	2	Valid	73.9	36.7	944 ms	27.2 ms	2.0	0	2,218	30.0 s
Sustained decode	64K	4	Valid	119.3	29.2	1,496 ms	34.2 ms	4.0	0	3,576	30.0 s
Sustained decode	64K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	1	Valid	32.3	32.3	1,945 ms	30.9 ms	1.0	0	968	30.0 s
Sustained decode	128K	2	Valid	60.0	29.4	2,740 ms	34.1 ms	2.0	0	1,800	30.0 s
Sustained decode	128K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	1	Requests queued	22.5	24.5	2,995 ms	40.8 ms	0.9	0	675	30.0 s
Sustained decode	248K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	128	Over capacity	—	—	—	—	—	—	—	—

Requests queued: the server could not run all requests at once, so some waited in a queue. Over capacity: not run because this many requests at this context would exceed the server's KV cache.