Gemma-4-E4B-it-DSpark

A DSpark draft model for speculative decoding with google/gemma-4-E4B-it as the verifier, trained with speculators. The drafter proposes 8 tokens at a time and the verifier checks them in one forward pass, so output is identical to running the verifier alone. This makes it a lossless speedup. Mean acceptance length is 3.24 tokens committed per verification step, up to 6.05 on math_reasoning and 4.91 on HumanEval. On an A100 at concurrency 1, that gives a throughput speedup over the verifier alone of 4.66× on math_reasoning and 3.85× on HumanEval.

Training code: rasyosef/train-dspark-draft-models.

Trained on 216,000 samples.

Usage

vLLM

vLLM loads the verifier automatically from the config, so don't pass it separately.

vllm serve rasyosef/gemma-4-E4B-it-dspark --port 8000 --gpu-memory-utilization 0.8

Then query the OpenAI-compatible endpoint at http://localhost:8000/v1.

Details

Architecture

Qwen3-style decoder with 4 draft layers, ~0.55B params, bfloat16
Hidden size 2560, MLP size 8192
32 query heads over 8 KV heads (GQA, head dim 128)
First three layers use sliding-window attention (window 2048); the last uses full attention
Block size 8 (8 draft tokens per step)
Draft vocabulary of 32,000 tokens, reduced from the verifier's 262,144 by token frequency over the training data
Aux hidden states from verifier layers 2/10/20/30/40

Training

3 epochs at lr 5e-4 (Muon optimizer, cosine schedule with 4% warmup)
Loss: {"ce": 0.1, "tv": 0.9}
Data: 216,000 filtered prompts from Open PerfectBlend, with every response regenerated by the verifier itself
96/4 train/validation split
Prompts prepared at 2560 tokens
Training sequence length 4096, up to 512 anchors per sample
Verifier hidden states pulled on demand from a running vLLM server and deleted after use, rather than staged to disk up front

Evaluation

evaluate.py across the nine RedHatAI/speculator_benchmarks subsets.

Setup

Sampling: temperature 1.0, top_p 0.95, top_k 64 (the defaults from gemma-4-E4B-it's generation_config.json, which vLLM applies when a request doesn't set them)
Hardware: single A100 40GB
vLLM version: 0.31.0
Same settings for the baseline, MTP and DSpark runs

acceptance_length is mean tokens committed per verification step, including the bonus token. Its floor is 1.0 and its ceiling is 9.0 at block size 8.

subset	acceptance_length	pos_0	pos_1	pos_2	pos_3	pos_4	pos_5	pos_6	pos_7
math_reasoning	6.048	88.2%	79.1%	72.0%	65.2%	58.0%	52.5%	47.3%	42.5%
HumanEval	4.905	81.8%	67.9%	57.2%	48.9%	42.5%	36.2%	30.6%	25.5%
question	2.904	63.0%	39.9%	26.5%	19.4%	14.7%	11.4%	8.9%	6.6%
writing	2.882	62.5%	39.4%	26.3%	19.1%	14.6%	11.2%	8.5%	6.6%
translation	2.757	59.6%	38.5%	27.6%	18.9%	13.2%	8.9%	5.7%	3.2%
rag	2.753	64.3%	41.9%	27.1%	17.2%	11.2%	7.0%	4.0%	2.6%
tool_call	2.697	60.9%	38.6%	24.7%	16.8%	11.6%	8.1%	5.4%	3.6%
qa	2.281	55.7%	31.1%	17.5%	10.2%	6.1%	3.8%	2.3%	1.4%
summarization	2.189	55.9%	29.0%	15.5%	8.8%	5.1%	2.7%	1.2%	0.7%

Weighted across all subsets: 3.244 over 96,410 verification steps.

Acceptance is highest where the verifier's next token is most predictable. math_reasoning leads at 6.05, more than 1.1 tokens ahead of HumanEval, and still accepts more than half of pos_5 drafts (52.5%). Its last position (pos_7, 42.5%) matches HumanEval's pos_4 (42.5%). HumanEval is clearly second at 4.91 and also makes real use of the full 8-token block, with a quarter of pos_7 drafts accepted. The other seven subsets sit between 2.19 and 2.90.

Throughput (tokens/s)

Mean output throughput in tokens/s, using the setup above at max concurrency 1. The last column is DSpark's speedup over the verifier alone, and the last row is each method's mean per-subset speedup over the baseline. Gemma's MTP drafter (google/gemma-4-E4B-it-assistant, 8 draft tokens) is included for comparison.

subset	baseline (no drafter)	MTP	DSpark (this model)	DSpark speedup
math_reasoning	98.0	394.5	456.4	4.66×
HumanEval	97.7	362.2	376.0	3.85×
question	97.9	231.1	254.0	2.59×
writing	97.8	225.8	257.1	2.63×
translation	98.2	280.3	212.6	2.16×
rag	93.7	208.6	176.5	1.88×
tool_call	94.0	219.0	180.2	1.92×
qa	97.7	174.6	182.5	1.87×
summarization	94.5	168.5	148.4	1.57×
average speedup	1.00×	2.60×	2.57×	2.57×

The throughput ranking follows acceptance length: math_reasoning and HumanEval see the largest gains, summarization the smallest. DSpark is the fastest on 5 of 9 subsets and MTP on 4, and the two are nearly level on average speedup (2.57× vs 2.60×). DSpark's best result is math_reasoning at 456.4 tokens/s, 4.66× the baseline and 16% ahead of MTP. It also leads MTP by 10–14% on question and writing. MTP is ahead on translation, rag, tool_call and summarization.

Comparison with MTP

Acceptance length on the same nine subsets for Gemma's MTP drafter (google/gemma-4-E4B-it-assistant). Both drafters propose 8 tokens per step, so they share the same 9.0 ceiling. Per-step draft cost differs between the two, so acceptance length alone doesn't determine speed. See the throughput section above for that.

subset	MTP (8 draft tokens)	DSpark (8 draft tokens)	best
math_reasoning	5.370	6.048	DSpark
HumanEval	4.851	4.905	DSpark
question	2.820	2.904	DSpark
writing	2.851	2.882	DSpark
translation	3.810	2.757	MTP
rag	3.293	2.753	MTP
tool_call	3.154	2.697	MTP
qa	2.303	2.281	MTP
summarization	2.500	2.189	MTP
mean (unweighted)	3.439	3.269	MTP

DSpark has the higher acceptance length on 4 of 9 subsets and MTP on 5, with MTP ahead on average. DSpark's clearest win is math_reasoning, where it leads by 0.68 tokens. HumanEval, question, writing and qa are close, all within 0.1 tokens either way. MTP's lead comes from translation (by 1.05 tokens), rag, tool_call and summarization, the same four subsets where it is faster. On qa, DSpark is slightly faster despite a slightly lower acceptance length. The mean here is unweighted across subsets, so DSpark's 3.269 differs slightly from its weighted 3.244 above.

Limitations

Works only with gemma-4-E4B-it and is not usable as a standalone model. Acceptance falls off steeply past the first two positions on prose-like traffic, most of all on summarization and qa, so the gains concentrate in math and code. On prose-heavy workloads the tail of the 8-token block rarely gets accepted. All results were measured with sampling at temperature 1.0; greedy decoding typically yields higher acceptance, so these numbers are closer to the conservative case. Real-world speedup depends on your traffic mix. Because verification is lossless, the verifier's own behavior and biases carry through unchanged.