Speed test matrix

gemma-4-26B-A4B-it-NVFP4 on RTX 3090 24GB

vllm · 0.32.0 · NVFP46 contexts (0–248K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 29, 2026, 17:48 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: NVFP4
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 RedHatAI/gemma-4-26B-A4B-it-NVFP4 --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 context5,8921,360 ms TTFT

16K context5,0733,129 ms TTFT

32K context4,0637,777 ms TTFT

64K context2,91021,665 ms TTFT

128K context1,84668,231 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	122.632 ms TTFT	213.459 ms TTFT	343.057 ms TTFT	532.373 ms TTFT	Looped	1,137.8108 ms TTFT	Over capacity	Over capacity
16K	Looped	Looped	Looped	Looped	Looped	Over capacity	Over capacity	Over capacity
32K	Looped	Looped	Looped	Looped	Over capacity	Over capacity	Over capacity	Over capacity
64K	64.7314 ms TTFT	112.9423 ms TTFT	Looped	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
128K	Looped	Looped	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
248K	Looped	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	5,892.0	—	1,360 ms	—	—	1	—	—
Prefill	16K	1	Valid	5,073.0	—	3,129 ms	—	—	1	—	—
Prefill	32K	1	Valid	4,063.0	—	7,777 ms	—	—	1	—	—
Prefill	64K	1	Valid	2,910.0	—	21,665 ms	—	—	1	—	—
Prefill	128K	1	Valid	1,846.0	—	68,231 ms	—	—	1	—	—
Sustained decode	0	1	Valid	122.6	124.2	32 ms	8.1 ms	1.0	0	3,676	30.0 s
Sustained decode	0	2	Valid	213.4	108.4	59 ms	9.2 ms	2.0	0	6,403	30.0 s
Sustained decode	0	4	Valid	343.0	87.9	57 ms	11.4 ms	4.0	0	10,296	30.0 s
Sustained decode	0	8	Valid	532.3	68.8	73 ms	14.5 ms	8.0	0	15,968	30.0 s
Sustained decode	0	16	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	32	Valid	1,137.8	38.3	108 ms	26.1 ms	32.0	0	34,112	30.0 s
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	8	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	16	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	8	Looped	—	—	—	—	—	0	0	—
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	64.7	64.8	314 ms	15.4 ms	1.0	0	1,940	30.0 s
Sustained decode	64K	2	Valid	112.9	706.6	423 ms	9.2 ms	2.0	0	3,386	30.0 s
Sustained decode	64K	4	Looped	—	—	—	—	—	0	0	—
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	Looped	—	—	—	—	—	0	0	—
Sustained decode	128K	2	Looped	—	—	—	—	—	0	0	—
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	Looped	—	—	—	—	—	0	0	—
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.