gemma4-26b-a4b-it-dspark-multidomain

DSpark Multidomain — Gemma-4 26B-A4B Speculative Decoding Draft

Multidomain DSpark draft for google/gemma-4-26B-A4B-it. Trained for broad coverage across domains (coding, math, chat, RAG, and mixed traffic) so acceptance stays more uniform instead of peaking on a single workload. Speeds up Gemma 4 26B-A4B inference with DSpark speculative decoding in vLLM — up to ~2× output throughput vs the same target without a draft.

This checkpoint is a draft head only (not a standalone LLM). Pair it with the Gemma-4 26B-A4B-it target.

<p> <a href="https://arxiv.org/abs/2607.05147"><img alt="DSpark paper" src="https://img.shields.io/badge/Paper-DSpark-b31b1b?logo=arxiv"></a> <a href="https://github.com/deepseek-ai/DeepSpec"><img alt="DeepSpec" src="https://img.shields.io/badge/Code-DeepSpec-111111?logo=github"></a> <a href="https://docs.vllm.ai/"><img alt="vLLM" src="https://img.shields.io/badge/Serve-vLLM-1a7f37"></a> <a href="https://huggingface.co/google/gemma-4-26B-A4B-it"><img alt="Target" src="https://img.shields.io/badge/Target-Gemma--4--26B--A4B--it-4285F4"></a> </p>

Quick start

vllm serve google/gemma-4-26B-A4B-it \
  --speculative-config '{"method": "dspark", "model": "systalyze/gemma4-26b-a4b-it-dspark-multidomain", "num_speculative_tokens": 7}'

Model details

Benchmarks vs base (no draft)

Same GPUs, same prompts: base = target alone; +DSpark = this draft + target.
AL = (accepted + num_drafts) / num_drafts. aiperf 0.12, 512 requests, temperature 1.0, ignore_eos=true.

Coding throughput

ISL ~U[3500,10000], OSL ~U[200,400].

Speedup by domain

ISL/OSL held fixed; only prompt content changes.

Acceptance length by domain

Context length (coding, OSL fixed)

AL stays flat as context grows; speedup falls mainly from target verify cost.

Speedup vs base

Acceptance length

What is DSpark?

DSpark is a speculative-decoding method that combines a parallel draft backbone with a lightweight Markov head and a confidence head for scheduled verification. Training/reference code: DeepSpec.

Citation

@article{dspark2026,
  title={DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation},
  year={2026},
  url={https://arxiv.org/abs/2607.05147}
}