LLaMA3.1-8B-Instruct-DFlash-UltraChat-GGUF

LLaMA 3.1 8B Instruct DFlash UltraChat GGUF

GGUF quantizations of z-lab DFlash draft model for LLaMA 3.1 8B Instruct DFlash UltraChat.

Use with BeeLlama.cpp, a llama.cpp fork with advanced quantization features.

---

LLaMA3.1-8B-Instruct-DFlash-UltraChat

Paper | GitHub | Blog

DFlash is a novel speculative decoding method that utilizes a lightweight block diffusion model for drafting. It enables efficient, high-quality parallel drafting that pushes the limits of inference speed.

This model is the drafter component. It must be used in conjunction with the target model meta-llama/Llama-3.1-8B-Instruct.

<div align="center"> <img src="assets/dflash_system.png" alt="DFlash Architecture" width="100%"> </div>

📊 Training Data

LLaMA3.1-8B-Instruct-DFlash-UltraChat is trained on Ultrachat-200K and ShareGPT datasets, aiming to align with EAGLE-3 training data. The assistant reponses in the datasets are regenerated by meta-llama/Llama-3.1-8B-Instruct.

🚀 Quick Start

SGLang

Installation

uv pip install "git+https://github.com/sgl-project/sglang.git@refs/pull/20547/head#subdirectory=python"

Launch Server

# 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 meta-llama/Llama-3.1-8B-Instruct \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path z-lab/LLaMA3.1-8B-Instruct-DFlash-UltraChat \
    --tp-size 1 \
    --dtype bfloat16 \
    --attention-backend fa3 \
    --mem-fraction-static 0.75 \
    --trust-remote-code

Usage

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="meta-llama/Llama-3.1-8B-Instruct",
    messages=[{"role": "user", "content": "Write a quicksort in Python."}],
    max_tokens=2048,
    temperature=0.0,
)
print(response.choices[0].message.content)

vLLM

Installation

uv pip install vllm
uv pip install -U vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly

Launch Server

vllm serve meta-llama/Llama-3.1-8B-Instruct \
  --speculative-config '{"method": "dflash", "model": "z-lab/LLaMA3.1-8B-Instruct-DFlash-UltraChat", "num_speculative_tokens": 9}' \
  --attention-backend flash_attn \
  --max-num-batched-tokens 32768

Usage

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="meta-llama/Llama-3.1-8B-Instruct",
    messages=[{"role": "user", "content": "Write a quicksort in Python."}],
    max_tokens=2048,
    temperature=0.0,
)
print(response.choices[0].message.content)

Transformers

Installation

pip install transformers==4.57.3 torch==2.9.0 accelerate

Inference

from transformers import AutoModel, AutoModelForCausalLM, AutoTokenizer

model = AutoModel.from_pretrained(
    "z-lab/LLaMA3.1-8B-Instruct-DFlash-UltraChat", 
    trust_remote_code=True, 
    dtype="auto", 
    device_map="cuda:0"
).eval()

target = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct", 
    dtype="auto", 
    device_map="cuda:0"
).eval()

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
prompt = "How many positive whole-number divisors does 196 have?"
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generate_ids = model.spec_generate(
    input_ids=model_inputs["input_ids"], 
    max_new_tokens=2048, 
    temperature=0.0, 
    target=target, 
    stop_token_ids=[tokenizer.eos_token_id]
)

print(tokenizer.decode(generate_ids[0], skip_special_tokens=True))

Evaluation

DFlash consistently achieves higher speedups than the state-of-the-art speculative decoding method EAGLE-3. All experiments are conducted using SGLang on a single B200 GPU.

For EAGLE-3, we evaluate two speculative decoding configurations:

• --speculative-num-steps 7, --speculative-eagle-topk 10, --speculative-num-draft-tokens 10
• --speculative-num-steps 7, --speculative-eagle-topk 10, --speculative-num-draft-tokens 60, which is the official setting used in the EAGLE-3 paper.

For DFlash, we use a block size of 10 during speculation.

We compare against the EAGLE-3 checkpoint lmsys/sglang-EAGLE3-LLaMA3.1-Instruct-8B, which is the official EAGLE-3 checkpoint adapted for SGLang inference.

Both the DFlash and EAGLE-3 draft models are trained on the UltraChat-200K and ShareGPT datasets.

GSM8K

---

HumanEval

---

Alpaca

Acknowledgement

We are grateful to Yotta Labs for their compute support in training this draft model.

Citation

If you find DFlash useful for your research or applications, please cite our project.

@misc{chen2026dflash,
  title         = {DFlash: Block Diffusion for Flash Speculative Decoding},
  author        = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  year          = {2026},
  eprint        = {2602.06036},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2602.06036}
}