Gemma 4 26B A4B IT DFlash GGUF
GGUF quantizations of z-lab DFlash draft model for Gemma 4 26B A4B.
Use with BeeLlama.cpp, a llama.cpp fork with advanced quantization features.
gemma-4-26B-A4B-it-DFlash
DFlash is a speculative decoding method that uses a lightweight block diffusion model to draft multiple tokens in parallel. This is the drafter model, which must be paired with google/gemma-4-26B-A4B-it.
Quick Start
Installation
vLLM: until Gemma4 DFlash support is merged, install vLLM from PR #41703:
uv pip install -U --torch-backend=auto \
"vllm @ git+https://github.com/vllm-project/vllm.git@refs/pull/41703/head"
SGLang:
uv pip install "git+https://github.com/sgl-project/sglang.git@refs/pull/23000/head#subdirectory=python"
Launch Server
vLLM:
vllm serve google/gemma-4-26B-A4B-it \
--speculative-config '{"method": "dflash", "model": "z-lab/gemma-4-26B-A4B-it-DFlash", "num_speculative_tokens": 15, "attention_backend": "flash_attn"}' \
--attention-backend triton_attn \
--max-num-batched-tokens 32768 \
--trust-remote-code
SGLang:
# Optional: enable schedule overlapping (experimental, may not be stable)
# export SGLANG_ENABLE_SPEC_V2=1
# export SGLANG_ENABLE_DFLASH_SPEC_V2=1
# export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
python -m sglang.launch_server \
--model-path google/gemma-4-26B-A4B-it \
--speculative-algorithm DFLASH \
--speculative-draft-model-path z-lab/gemma-4-26B-A4B-it-DFlash \
--speculative-num-draft-tokens 16 \
--tp-size 1 \
--attention-backend triton \
--speculative-draft-attention-backend fa4 \
--trust-remote-code
Usage
For vLLM, use port 8000. For SGLang, use port 30000.
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="google/gemma-4-26B-A4B-it",
messages=[{"role": "user", "content": "Write a quicksort in Python."}],
max_tokens=4096,
temperature=0.0,
extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)
print(response.choices[0].message.content)
Benchmark Results
Setup: Single NVIDIA B300 GPU per server/run, vLLM, thinking enabled, max output length 4096, greedy decoding.
Throughput and Speedup
DFlash achieves up to 3.7x speedup at concurrency 8.
Generated tokens/sec (speedup vs. autoregressive baseline)
Block Size = 16
| Task | Concurrency | AR | DFlash |
|---|---|---|---|
| Math500 | 1 | 259 | 925 (3.6x) |
| 8 | 1296 | 4837 (3.7x) | |
| 32 | 3233 | 11435 (3.5x) | |
| GSM8K | 1 | 256 | 825 (3.2x) |
| 8 | 1217 | 4241 (3.5x) | |
| 32 | 3174 | 10306 (3.2x) | |
| HumanEval | 1 | 246 | 818 (3.3x) |
| 8 | 1182 | 4240 (3.6x) | |
| 32 | 2881 | 9150 (3.2x) | |
| MBPP | 1 | 272 | 698 (2.6x) |
| 8 | 1288 | 3387 (2.6x) | |
| 32 | 2950 | 7898 (2.7x) | |
| MT-Bench | 1 | 272 | 492 (1.8x) |
| 8 | 1146 | 2259 (2.0x) | |
| 32 | 2164 | 4829 (2.2x) |
Acceptance Length
| Task | c1 | c8 | c32 |
|---|---|---|---|
| Math500 | 8.61 | 8.55 | 8.60 |
| GSM8K | 7.71 | 7.76 | 7.72 |
| HumanEval | 7.80 | 7.87 | 7.83 |
| MBPP | 6.09 | 5.99 | 6.03 |
| MT-Bench | 4.33 | 4.33 | 4.24 |
Acknowledgements
Special thanks to David Wang for his outstanding engineering support on this project. We are also grateful to Modal, InnoMatrix, and Yotta Labs for providing the compute resources used to train this draft model.
Citation
If you find DFlash useful, please cite our work. To share feedback on DFlash or request new model support, please fill out this form: DFlash Feedback.
@article{chen2026dflash,
title = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
author = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
journal = {arXiv preprint arXiv:2602.06036},
year = {2026}
}