Tested on one L40S with single-stream greedy decoding and 640-token generations. GSM8K answers were checked numerically. HumanEval solutions were run against their unit tests.

configuration	decode tok/s	speedup	accepted tokens (GSM8K / HumanEval)	GSM8K	HumanEval
Base model	44.8	1.00x	—	91.7%	79.2%
DFlash, block size 16	207.5	4.63x	6.5 / 8.5	90.8%	80.8%

Block size 16 was the fastest tested setting. The task pass rates stayed close to the baseline in this sample. The benchmark does not claim identical text across separate server runs.

Run it

Use lmsysorg/sglang:v0.5.17-cu129 with a GPU that has at least 24 GB of VRAM:

bash

python -m sglang.launch_server \
  --model-path ornith-ai/Ornith-1.5-9B \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path z-lab/Qwen3.5-9B-DFlash \
  --speculative-dflash-block-size 16 \
  --reasoning-parser qwen3 \
  --trust-remote-code

The tested setup used an L40S. An A100 also works. Keep transformers at 5.12.1 with SGLang 0.5.17.

Full reproduction files are in the GitHub repository.

Attribution

Base model: Ornith-1.5-9B
Draft model: Qwen3.5-9B-DFlash
Serving engine: SGLang

This recipe and benchmark were prepared by ultimatechris. The models keep their own licenses.

ultimatechris/Ornith-1.5-9B-DFlash-SGLang

Ornith-1.5-9B with DFlash in SGLang

Results

Run it

Attribution

ultimatechris/
Ornith-1.5-9B-DFlash-SGLang