Speed test matrix

LFM2.5-2.6B on H200 NVL 141GB

sglang · 0.5.6.post3.dev9265+g05c584c44 · BF165 contexts (8K–123K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 123K, 128K

by @keennaylmx dev importMeasured Sep 19, 2026, 00:47 UTC

Hardware

Type: Discrete GPU
GPU: H200 NVL
VRAM: 141 GB
OS: linux

Engine

Engine: sglang
Version: 0.5.6.post3.dev9265+g05c584c44
Quantization: BF16
Backend: cuda
Repository: github.com/sgl-project/sglang
Build: cuda-13.2
Commit: 05c584c44fb0

Run

Served model: lfm
Revision: main
Benchmark tool: llm-inference-bench v0.4.29
Decode test: Fixed duration
Prefill mode: Measured during decode warm-up
Time per measurement: 30 s
Max output: 4,096 tokens

Launch command

"$HOME/env_h200-liquidai-sglang/bin/sglang" serve --model-path LiquidAI/LFM2.5-2.6B --served-model-name lfm --trust-remote-code --tp 1 --reasoning-parser qwen3 --tool-call-parser lfm2 --context-length 131072 --mem-fraction-static 0.864737 --enable-metrics --host 0.0.0.0 --port 8000 --api-key <redacted>

Prefill

prompt tok/s by context length

8K context62.2K131 ms TTFT

16K context66.4K244 ms TTFT

32K context67.3K479 ms TTFT

64K context58.9K1,091 ms TTFT

123K context46.6K2,647 ms TTFT

128K context45.8K2,804 ms TTFT

Sustained decode

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

Context	1	2	4	8	16	32	64	128
8K	293.3114 ms TTFT	526.3153 ms TTFT	1,043.4262 ms TTFT	2,001.6134 ms TTFT	3,489.9197 ms TTFT	6,184.5365 ms TTFT	8,904.1698 ms TTFT	9,578.71,281 ms TTFT
16K	289.2141 ms TTFT	388.7204 ms TTFT	1,022.8259 ms TTFT	1,845.7222 ms TTFT	3,291.1361 ms TTFT	5,227.9667 ms TTFT	6,240.11,309 ms TTFT	9,164.42,546 ms TTFT
32K	282.0182 ms TTFT	288.5254 ms TTFT	911.9356 ms TTFT	1,709.1357 ms TTFT	2,747.8702 ms TTFT	3,578.51,486 ms TTFT	5,273.72,783 ms TTFT	Over capacity
64K	270.4257 ms TTFT	374.5318 ms TTFT	815.2419 ms TTFT	1,384.8772 ms TTFT	1,675.41,464 ms TTFT	2,783.52,776 ms TTFT	Over capacity	Over capacity
123K	252.4219 ms TTFT	311.9377 ms TTFT	567.5624 ms TTFT	755.81,317 ms TTFT	1,449.82,679 ms TTFT	Over capacity	Over capacity	Over capacity

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	62.2K	—	131 ms	—	—	1	—	—
Prefill	16K	1	Valid	66.4K	—	244 ms	—	—	1	—	—
Prefill	32K	1	Valid	67.3K	—	479 ms	—	—	1	—	—
Prefill	64K	1	Valid	58.9K	—	1,091 ms	—	—	1	—	—
Prefill	123K	1	Valid	46.6K	—	2,647 ms	—	—	1	—	—
Prefill	128K	1	Valid	45.8K	—	2,804 ms	—	—	1	—	—
Sustained decode	8K	1	Valid	293.3	295.5	114 ms	3.4 ms	1.0	2	8,799	30.0 s
Sustained decode	8K	2	Valid	526.3	276.2	153 ms	3.6 ms	2.3	4	15,788	30.0 s
Sustained decode	8K	4	Valid	1,043.4	262.7	262 ms	3.8 ms	4.5	8	31,300	30.0 s
Sustained decode	8K	8	Valid	2,001.6	257.8	134 ms	3.9 ms	8.8	16	60,049	30.0 s
Sustained decode	8K	16	Valid	3,489.9	219.7	197 ms	4.6 ms	18.9	16	104,696	30.0 s
Sustained decode	8K	32	Valid	6,184.5	196.7	365 ms	5.1 ms	31.7	32	185,541	30.0 s
Sustained decode	8K	64	Valid	8,904.1	141.6	698 ms	7.1 ms	68.9	64	267,125	30.0 s
Sustained decode	8K	128	Valid	9,578.7	75.3	1,281 ms	13.3 ms	192.0	0	287,488	30.0 s
Sustained decode	16K	1	Valid	289.2	291.9	141 ms	3.4 ms	1.0	2	8,674	30.0 s
Sustained decode	16K	2	Valid	388.7	212.7	204 ms	5.1 ms	61.6	2	11,660	30.0 s
Sustained decode	16K	4	Valid	1,022.8	258.2	259 ms	3.9 ms	4.0	8	30,684	30.0 s
Sustained decode	16K	8	Valid	1,845.7	231.4	222 ms	4.3 ms	9.1	8	55,368	30.0 s
Sustained decode	16K	16	Valid	3,291.1	208.9	361 ms	4.8 ms	16.0	16	98,736	30.0 s
Sustained decode	16K	32	Valid	5,227.9	166.4	667 ms	6.0 ms	32.5	32	156,820	30.0 s
Sustained decode	16K	64	Valid	6,240.1	97.9	1,309 ms	10.2 ms	86.4	0	187,200	30.0 s
Sustained decode	16K	128	Valid	9,164.4	70.7	2,546 ms	14.1 ms	140.0	0	274,944	30.0 s
Sustained decode	32K	1	Valid	282.0	285.8	182 ms	3.5 ms	1.0	2	8,460	30.0 s
Sustained decode	32K	2	Valid	288.5	190.8	254 ms	5.9 ms	78.7	2	8,656	30.0 s
Sustained decode	32K	4	Valid	911.9	228.7	356 ms	4.4 ms	4.6	4	27,356	30.0 s
Sustained decode	32K	8	Valid	1,709.1	215.9	357 ms	4.6 ms	7.8	8	51,272	30.0 s
Sustained decode	32K	16	Valid	2,747.8	174.3	702 ms	5.7 ms	15.9	16	82,436	30.0 s
Sustained decode	32K	32	Valid	3,578.5	111.2	1,486 ms	9.0 ms	40.5	0	107,360	30.0 s
Sustained decode	32K	64	Valid	5,273.7	83.4	2,783 ms	12.0 ms	64.0	0	158,197	30.0 s
Sustained decode	32K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	1	Valid	270.4	275.1	257 ms	3.6 ms	1.0	2	8,112	30.0 s
Sustained decode	64K	2	Valid	374.5	199.8	318 ms	5.3 ms	23.3	2	11,236	30.0 s
Sustained decode	64K	4	Valid	815.2	204.5	419 ms	4.9 ms	4.5	4	24,452	30.0 s
Sustained decode	64K	8	Valid	1,384.8	175.4	772 ms	5.7 ms	8.3	8	41,545	30.0 s
Sustained decode	64K	16	Valid	1,675.4	103.2	1,464 ms	9.7 ms	21.1	0	50,256	30.0 s
Sustained decode	64K	32	Valid	2,783.5	84.9	2,776 ms	11.8 ms	32.5	0	83,488	30.0 s
Sustained decode	64K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	123K	1	Valid	252.4	256.7	219 ms	3.9 ms	1.0	2	7,572	30.0 s
Sustained decode	123K	2	Valid	311.9	182.5	377 ms	5.8 ms	15.8	2	9,356	30.0 s
Sustained decode	123K	4	Valid	567.5	161.7	624 ms	6.3 ms	5.2	4	17,024	30.0 s
Sustained decode	123K	8	Valid	755.8	95.7	1,317 ms	10.4 ms	11.6	0	22,671	30.0 s
Sustained decode	123K	16	Valid	1,449.8	88.1	2,679 ms	11.3 ms	16.2	0	43,488	30.0 s
Sustained decode	123K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	123K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	123K	128	Over capacity	—	—	—	—	—	—	—	—

Over capacity: not run because this many requests at this context would exceed the server's KV cache.