Speed test matrix

gemma-4-12B-it on H200 NVL 141GB

sglang · 0.0.0.dev16562+g6283bde44 · BF166 contexts (8K–251K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 251K, 256K

by @keennaylmx dev importMeasured Sep 1, 2026, 20:59 UTC

Hardware

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

Engine

Engine: sglang
Version: 0.0.0.dev16562+g6283bde44
Quantization: BF16
Backend: cuda
Repository: github.com/SubSir/sglang
Build: cuda-13.2
Commit: 6283bde44381

Run

Served model: gemma
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

env SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 "$HOME/env_h200-z-lab-sglang-pr-35209/bin/sglang" serve --model-path google/gemma-4-12B-it --served-model-name gemma --trust-remote-code --tp 1 --reasoning-parser gemma4 --tool-call-parser gemma4 --context-length 262144 --mem-fraction-static 0.838218 --enable-metrics --host 0.0.0.0 --port 8000 --api-key <redacted> --page-size 1 --attention-backend triton --speculative-algorithm DFLASH --speculative-draft-model-path z-lab/gemma4-12B-it-DFlash --speculative-num-draft-tokens 16 --speculative-draft-attention-backend triton

Prefill

prompt tok/s by context length

8K context5,3761,526 ms TTFT

16K context9,6531,684 ms TTFT

32K context6,8234,744 ms TTFT

64K context4,32114,941 ms TTFT

128K context2,49051,803 ms TTFT

251K context858.0294,644 ms TTFT

256K context1,352190,659 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	20.3Requests queued735 ms TTFT	20.02,306 ms TTFT	43.81,062 ms TTFT	81.21,088 ms TTFT	123.7830 ms TTFT	Over capacity	Over capacity	Over capacity
16K	272.8Requests queued130 ms TTFT	199.422,105 ms TTFT	611.1256 ms TTFT	593.9401 ms TTFT	547.9760 ms TTFT	Over capacity	Over capacity	Over capacity
32K	140.8205 ms TTFT	41.7436 ms TTFT	164.8498 ms TTFT	405.1790 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
64K	112.8351 ms TTFT	59.7760 ms TTFT	142.0848 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
128K	13.3813 ms TTFT	50.71,463 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
251K	7.8Requests queued1,909 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	5,376.0	—	1,526 ms	—	—	1	—	—
Prefill	16K	1	Valid	9,653.0	—	1,684 ms	—	—	1	—	—
Prefill	32K	1	Valid	6,823.0	—	4,744 ms	—	—	1	—	—
Prefill	64K	1	Valid	4,321.0	—	14,941 ms	—	—	1	—	—
Prefill	128K	1	Valid	2,490.0	—	51,803 ms	—	—	1	—	—
Prefill	251K	1	Valid	858.0	—	294,644 ms	—	—	1	—	—
Prefill	256K	1	Valid	1,352.0	—	190,659 ms	—	—	1	—	—
Sustained decode	8K	1	Requests queued	20.3	20.4	735 ms	49.0 ms	1.9	0	602	29.7 s
Sustained decode	8K	2	Valid	20.0	9.2	2,306 ms	109.6 ms	3.0	0	593	29.7 s
Sustained decode	8K	4	Valid	43.8	10.1	1,062 ms	100.5 ms	7.0	0	1,306	29.8 s
Sustained decode	8K	8	Valid	81.2	7.7	1,088 ms	129.7 ms	15.0	0	2,422	29.8 s
Sustained decode	8K	16	Valid	123.7	6.6	830 ms	150.8 ms	30.6	0	3,696	29.9 s
Sustained decode	8K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	8K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	8K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Requests queued	272.8	273.6	130 ms	3.7 ms	1.0	2	8,162	29.9 s
Sustained decode	16K	2	Valid	199.4	88.9	22,105 ms	11.3 ms	23.6	0	5,973	30.0 s
Sustained decode	16K	4	Valid	611.1	131.3	256 ms	7.6 ms	9.2	3	18,323	30.0 s
Sustained decode	16K	8	Valid	593.9	39.1	401 ms	25.6 ms	12.5	3	17,777	29.9 s
Sustained decode	16K	16	Valid	547.9	19.3	760 ms	51.8 ms	27.4	0	16,437	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	140.8	96.8	205 ms	14.1 ms	0.9	1	4,220	30.0 s
Sustained decode	32K	2	Valid	41.7	20.3	436 ms	50.6 ms	24.7	0	1,249	30.0 s
Sustained decode	32K	4	Valid	164.8	34.5	498 ms	40.4 ms	21.6	0	4,915	29.8 s
Sustained decode	32K	8	Valid	405.1	54.6	790 ms	18.3 ms	21.9	0	12,127	29.9 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	112.8	99.5	351 ms	10.0 ms	1.0	0	3,374	29.9 s
Sustained decode	64K	2	Valid	59.7	27.7	760 ms	50.6 ms	18.6	0	1,786	29.9 s
Sustained decode	64K	4	Valid	142.0	32.1	848 ms	31.8 ms	12.4	0	4,242	29.9 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	13.3	13.8	813 ms	72.6 ms	1.0	0	399	30.0 s
Sustained decode	128K	2	Valid	50.7	23.6	1,463 ms	56.6 ms	9.7	0	1,518	29.9 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	251K	1	Requests queued	7.8	8.1	1,909 ms	123.1 ms	0.6	0	231	29.7 s
Sustained decode	251K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	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.