Speed test matrix

gemma-4-26B-A4B-it on RTX PRO 6000 Blackwell Server Edition 96GB

sglang · 0.5.22.dev236+g1093c501d · BF166 contexts (8K–251K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 256K

by @keennaylmx dev importMeasured Oct 3, 2026, 19:47 UTC

Hardware

Type: Discrete GPU
GPU: RTX PRO 6000 Blackwell Server Edition
VRAM: 96 GB
OS: linux

Engine

Engine: sglang
Version: 0.5.22.dev236+g1093c501d
Quantization: BF16
Backend: cuda
Repository: github.com/sgl-project/sglang
Build: cuda-13.2
Commit: 1093c501dfef

Run

Served model: gemma
Revision: main
Benchmark tool: llm-inference-bench v0.7.6
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_rtxpro6k-google-sglang-1093c501dfef/bin/sglang" serve --model-path google/gemma-4-26B-A4B-it --served-model-name gemma --trust-remote-code --tp 1 --reasoning-parser gemma4 --tool-call-parser gemma4 --context-length 262144 --mem-fraction-static 0.76 --enable-metrics --host 0.0.0.0 --port 8000 --api-key <redacted> --enable-unified-memory --mm-process-config '{"image":{"max_soft_tokens":1120},"video":{"max_soft_tokens":1120}}' --attention-backend triton --bf16-gemm-backend torch --moe-runner-backend triton

Prefill

prompt tok/s by context length

8K context16.7K490 ms TTFT

16K context15.4K1,055 ms TTFT

32K context12.1K2,677 ms TTFT

64K context7,2708,874 ms TTFT

256K context2,258114,077 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	119.4425 ms TTFT	Looped	Looped	Looped	Over capacity	Over capacity	Over capacity	Over capacity
16K	Looped	Looped	Looped	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
32K	Looped	Looped	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	Looped	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
128K	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
251K	Over capacity	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	16.7K	—	490 ms	—	—	1	—	—
Prefill	16K	1	Valid	15.4K	—	1,055 ms	—	—	1	—	—
Prefill	32K	1	Valid	12.1K	—	2,677 ms	—	—	1	—	—
Prefill	64K	1	Valid	7,270.0	—	8,874 ms	—	—	1	—	—
Prefill	256K	1	Valid	2,258.0	—	114,077 ms	—	—	1	—	—
Sustained decode	8K	1	Valid	119.4	119.8	425 ms	8.3 ms	1.0	0	3,582	30.0 s
Sustained decode	8K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	8K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	8K	8	Looped	—	—	—	—	—	0	0	—
Sustained decode	8K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	8K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	8K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	8K	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	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	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	2	Looped	—	—	—	—	—	0	0	—
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	64K	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	64K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	4	Over capacity	—	—	—	—	—	—	—	—
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	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	2	Over capacity	—	—	—	—	—	—	—	—
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	Over capacity	—	—	—	—	—	—	—	—
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	—	—	—	—	—	—	—	—

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.