Speed test matrix

Fara1.5-4B-FP8 on RTX 3090 24GB

vllm · 0.32.0 · FP86 contexts (0–248K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 30, 2026, 08:35 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: FP8
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
# 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 prithivMLmods/Fara1.5-4B-FP8 --max-model-len 262144 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --language-model-only

Prefill

prompt tok/s by context length

8K context11.8K694 ms TTFT

16K context11K1,469 ms TTFT

32K context9,9303,252 ms TTFT

64K context8,2547,804 ms TTFT

128K context6,19520,773 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	119.535 ms TTFT	229.645 ms TTFT	440.568 ms TTFT	803.898 ms TTFT	1,409.2156 ms TTFT	2,189.2230 ms TTFT	Over capacity	Over capacity
16K	111.2131 ms TTFT	204.6133 ms TTFT	384.9186 ms TTFT	Looped	1,012.9599 ms TTFT	Over capacity	Over capacity	Over capacity
32K	104.4122 ms TTFT	183.9140 ms TTFT	340.5204 ms TTFT	549.6344 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
64K	92.9207 ms TTFT	153.1256 ms TTFT	274.2379 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
128K	76.1461 ms TTFT	114.5602 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
248K	56.8913 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity

Looped: the model kept repeating the same text, so the benchmark discarded this measurement. 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	11.8K	—	694 ms	—	—	1	—	—
Prefill	16K	1	Valid	11K	—	1,469 ms	—	—	1	—	—
Prefill	32K	1	Valid	9,930.0	—	3,252 ms	—	—	1	—	—
Prefill	64K	1	Valid	8,254.0	—	7,804 ms	—	—	1	—	—
Prefill	128K	1	Valid	6,195.0	—	20,773 ms	—	—	1	—	—
Sustained decode	0	1	Valid	119.5	119.7	35 ms	8.4 ms	1.0	0	3,584	30.0 s
Sustained decode	0	2	Valid	229.6	115.1	45 ms	8.7 ms	2.0	0	6,886	30.0 s
Sustained decode	0	4	Valid	440.5	110.7	68 ms	9.0 ms	4.0	0	13,212	30.0 s
Sustained decode	0	8	Valid	803.8	101.3	98 ms	9.9 ms	8.0	0	24,112	30.0 s
Sustained decode	0	16	Valid	1,409.2	89.3	156 ms	11.2 ms	16.0	0	42,272	30.0 s
Sustained decode	0	32	Valid	2,189.2	69.5	230 ms	14.4 ms	32.0	0	65,664	30.0 s
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	111.2	111.4	131 ms	9.0 ms	1.0	0	3,336	30.0 s
Sustained decode	16K	2	Valid	204.6	102.4	133 ms	9.8 ms	2.0	0	6,136	30.0 s
Sustained decode	16K	4	Valid	384.9	96.4	186 ms	10.4 ms	4.0	0	11,545	30.0 s
Sustained decode	16K	8	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	16	Valid	1,012.9	62.8	599 ms	15.9 ms	16.0	0	30,384	30.0 s
Sustained decode	16K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	104.4	104.5	122 ms	9.6 ms	1.0	0	3,131	30.0 s
Sustained decode	32K	2	Valid	183.9	92.0	140 ms	10.9 ms	2.0	0	5,518	30.0 s
Sustained decode	32K	4	Valid	340.5	85.2	204 ms	11.7 ms	4.0	0	10,216	30.0 s
Sustained decode	32K	8	Valid	549.6	68.7	344 ms	14.6 ms	8.0	0	16,488	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	92.9	93.0	207 ms	10.8 ms	1.0	0	2,787	30.0 s
Sustained decode	64K	2	Valid	153.1	76.6	256 ms	13.0 ms	2.0	0	4,592	30.0 s
Sustained decode	64K	4	Valid	274.2	68.6	379 ms	14.6 ms	4.0	0	8,224	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	76.1	76.2	461 ms	13.1 ms	1.0	0	2,283	30.0 s
Sustained decode	128K	2	Valid	114.5	57.3	602 ms	17.4 ms	2.0	0	3,435	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	Valid	56.8	57.1	913 ms	17.5 ms	1.0	0	1,703	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	—	—	—	—	—	—	—	—

Looped: the model kept repeating the same text, so the benchmark discarded this measurement. Over capacity: not run because this many requests at this context would exceed the server's KV cache.