gemma-4-31B-it-DFlash

Engine
vLLM
Listed
16 Oct 2026

vLLM: until Gemma4 DFlash support is merged, install vLLM from PR #41703:

Serves
Chat, Reasoning
Parameters
4.0B
Context
262,144 tokens
Weights
3.1 GB
License
Apache-2.0

Capabilities

Read from the model's capability tags. Requests are admitted per tag.

Chat chat
  • POST /v1/chat/completions
  • POST /v1/responses

Endpoint task mode generate

Inputs
Text
Reasoning effort reasoning_effort
none, low, medium, high
Values a request may send; the model family sets the vocabulary.
Structured output structured_output
not yet probed
Admitted until the capability probe observes the model.
Forced tool choice forced_tool_choice
not yet probed
Admitted until the capability probe observes the model.

Launch settings

Every engine serving the model starts with these.

Set by the owner. auto: the serving node derives it. default: the platform's value, left in place.

Engine
vLLM
Context window --max-model-len
262,144 tokens
Quantization --quantization
none default
Weights load as stored, bfloat16.
KV cache dtype --kv-cache-dtype
auto
No flag: the checkpoint's KV quantization, else the model dtype.
Scheduling policy --scheduling-policy
priority default
Interactive requests ahead of batch work; same-class traffic in arrival order.
Tool-call parser --tool-call-parser
auto
Resolves to qwen3_coder from the architecture.
Reasoning parser --reasoning-parser
auto
Resolves to qwen3 from the architecture.
Trust remote code --trust-remote-code
on default
Python shipped in the bundle may run at load (the default).

Engine tuning

Every setting auto: the serving node derives each one.

Set by the owner. auto: the serving node derives it.

Mixture of experts

Expert parallel --enable-expert-parallel
auto
Expert load balancing --enable-eplb
auto
All-to-all backend --all2all-backend
auto

Memory and batching

GPU memory utilization --gpu-memory-utilization
auto
Max sequences --max-num-seqs
auto
Max batched tokens --max-num-batched-tokens
auto
Long-prefill threshold --long-prefill-token-threshold
auto

Execution

Enforce eager --enforce-eager
auto
Prefix caching --enable-prefix-caching
auto
Async scheduling --no-async-scheduling
auto
Custom all-reduce --disable-custom-all-reduce
auto

Speculative decoding

Draft tokens checked by the model in one pass.

Speculative decoding
not configured
An endpoint's own setting, or the serving node's default for the model family, applies.
Drafts for
gemma-4-31B-it
This model is the speculative draft of a listed model.

Architecture

Read from the model's files at upload.

Architecture
qwen3
Parameters
4,029,731,328
Layers num_hidden_layers
5
Hidden size hidden_size
5,376
Attention heads num_attention_heads
64 (8 KV)
Head dim head_dim
128
Vocabulary vocab_size
262,144
Native dtype torch_dtype
bfloat16
Weights on disk
3.1 GB, safetensors

Call it

Create an endpoint for the model, then send it a request in the OpenAI shape.

Shell
# Create an endpoint for this model, then provision it. # --tier-id is the service tier to run on; pick one at /endpoints/new. erectl endpoints create \ --name "My endpoint" \ --slug my-endpoint \ --model-id 3AEB3B9F-0B02-42F6-B4C8-6DB551945051 \ --tier-id <tier-id> # provision takes the endpoint UUID returned by create erectl endpoints provision <endpoint-id> # Call it. Requests are routed by /<project-id>/<endpoint-slug>. curl https://api.erebine.ai/proj_ABC123/my-endpoint/v1/chat/completions \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $EREBINE_API_KEY" \ -d '{ "model": "gemma-4-31B-it-DFlash", "messages": [{"role": "user", "content": "Hello, world!"}], "stream": true }'

Model card

Images are left out; its HTML tables are set as tables.

Paper | GitHub | Blog

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-4-31B-it.

Quick Start

Installation

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-4-31B-it \
  --speculative-config '{"method": "dflash", "model": "z-lab/gemma-4-31B-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-4-31B-it \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path z-lab/gemma-4-31B-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-4-31B-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 5.8x speedup at concurrency 1.

Generated tokens/sec (speedup vs. autoregressive baseline)

Block Size = 16

TaskConcurrencyARDFlash
Math500177447 (5.8x)
85112650 (5.2x)
3213084962 (3.8x)
GSM8K178408 (5.3x)
85202321 (4.5x)
3213824447 (3.2x)
HumanEval176420 (5.6x)
84942389 (4.8x)
3211454139 (3.6x)
MBPP179343 (4.4x)
85352036 (3.8x)
3213893636 (2.6x)
MT-Bench179236 (3.0x)
85031334 (2.7x)
3211772257 (1.9x)

Acceptance Length

Taskc1c8c32
Math5008.598.598.62
GSM8K7.537.507.52
HumanEval8.007.897.96
MBPP6.136.136.14
MT-Bench4.234.194.19

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}
}