Speed test matrix

Qwen3.8-27B-INT4 on RTX 3090 24GB

vllm · 0.32.0 · INT4 W4A16 g1284 contexts (0–40K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 40K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 30, 2026, 17:48 UTC

Hardware

Type: Discrete GPU
GPU: RTX 3090
VRAM: 24 GB
CPU: AMD Ryzen 9 7900X (12C)
RAM: 32 GB
OS: Ubuntu 24.04.5

Engine

Engine: vllm
Version: 0.32.0
Quantization: INT4 W4A16 g128
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: MTP, 2 tokens lookahead. KV cache: int4_per_token_head
# 1x RTX 3090 24GB (Zagitta/gpu-unlock), 350 W limit (VBIOS max), PCIe Gen4 x16 | AMD Ryzen 9 7900X 12C | 32 GB DDR5 | Ubuntu 24.04.5, NVIDIA driver 595.99.02
vllm serve RedHatAI/Qwen3.8-27B-INT4 --max-model-len 49152 --gpu-memory-utilization 0.94 --max-num-seqs 31 --enable-prefix-caching --trust-remote-code --language-model-only --kv-cache-dtype int4_per_token_head --attention-backend TRITON_ATTN --speculative-config {"method":"mtp","num_speculative_tokens":2} --max-cudagraph-capture-size 16

Prefill

prompt tok/s by context length

8K context1,4835,551 ms TTFT

16K context1,25612,950 ms TTFT

32K context979.033,022 ms TTFT

40K context869.046,438 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	57.789 ms TTFT	120.1340 ms TTFT	6.6Requests queued248 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
16K	33.73,808 ms TTFT	63.37,025 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
32K	22.16,486 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
40K	18.55,675 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	1,483.0	—	5,551 ms	—	—	1	—	—
Prefill	16K	1	Valid	1,256.0	—	12,950 ms	—	—	1	—	—
Prefill	32K	1	Valid	979.0	—	33,022 ms	—	—	1	—	—
Prefill	40K	1	Valid	869.0	—	46,438 ms	—	—	1	—	—
Sustained decode	0	1	Valid	57.7	58.7	89 ms	17.0 ms	1.0	0	1,730	30.0 s
Sustained decode	0	2	Valid	120.1	60.5	340 ms	16.5 ms	2.0	0	3,602	30.0 s
Sustained decode	0	4	Requests queued	6.6	35.6	248 ms	28.1 ms	2.6	0	198	29.8 s
Sustained decode	0	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	33.7	34.0	3,808 ms	29.4 ms	1.0	0	1,011	30.0 s
Sustained decode	16K	2	Valid	63.3	31.9	7,025 ms	31.4 ms	2.0	0	1,898	30.0 s
Sustained decode	16K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	8	Over capacity	—	—	—	—	—	—	—	—
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	22.1	22.9	6,486 ms	43.7 ms	1.0	0	662	29.9 s
Sustained decode	32K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	8	Over capacity	—	—	—	—	—	—	—	—
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	40K	1	Valid	18.5	19.3	5,675 ms	51.9 ms	1.0	0	556	30.0 s
Sustained decode	40K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	40K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	40K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	40K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	40K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	40K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	40K	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.