Gemma-4-E2B-it-DFlash2

A DFlash2 draft model for speculative decoding with google/gemma-4-E2B-it as the verifier, trained with speculators. The drafter proposes 7 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 3.18 tokens committed per verification step, up to 5.60 on math_reasoning. That gives a throughput speedup over the verifier alone of 3.74× on math_reasoning and 3.13× on HumanEval.

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

Trained on 100,000 samples.

Usage

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

vllm serve rasyosef/gemma-4-E2B-it-dflash2 --port 8000 --gpu-memory-utilization 0.8

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

Details

Architecture. A Qwen3-style decoder with 5 draft layers, ~0.32B params, bfloat16. It has hidden size 1536, MLP size 6144, and 12 query heads over 4 KV heads (GQA, head dim 128). The first four layers use sliding-window attention (window 2048) and the last uses full attention. Block size is 8. The draft vocabulary is the verifier's full 262,144 tokens, with no reduction. Aux hidden states come from verifier layers 3/13/23/33.

Training. 4 epochs at lr 4e-4 (AdamW, cosine schedule with 4% warmup). The data is 100,000 filtered Open PerfectBlend prompts regenerated by the verifier itself, split 96/4 into train and validation. Prompts were prepared at 2048 tokens. Training sequence length was 8192, with up to 512 anchors per sample. Verifier hidden states were pulled on demand from a running vLLM server during training and deleted after use, rather than staged to disk up front.

Evaluation

evaluate.py throughput across the nine RedHatAI/speculator_benchmarks subsets. acceptance_length is mean tokens committed per verification step, including the bonus token. Its floor is 1.0 and its ceiling is 8.0 at block size 8.

subset	acceptance_length	pos_0	pos_1	pos_2	pos_3	pos_4	pos_5	pos_6
math_reasoning	5.602	90.1%	80.8%	72.6%	64.4%	57.1%	50.6%	44.6%
HumanEval	4.270	79.9%	64.3%	51.8%	42.8%	35.5%	29.2%	23.6%
tool_call	2.943	65.7%	43.5%	30.0%	21.4%	15.7%	10.6%	7.3%
translation	2.926	66.4%	45.1%	31.4%	20.9%	14.0%	9.3%	5.5%
rag	2.821	68.2%	45.6%	28.1%	17.9%	11.8%	6.7%	3.9%
writing	2.726	61.4%	37.5%	25.0%	17.9%	13.3%	10.0%	7.5%
question	2.696	60.7%	37.1%	24.3%	17.4%	13.0%	9.8%	7.4%
qa	2.305	56.1%	31.6%	18.2%	11.4%	6.8%	4.1%	2.2%
summarization	2.126	56.3%	28.8%	14.6%	7.2%	3.4%	1.6%	0.8%

Weighted across all subsets: 3.179 over 105,194 verification steps.

Acceptance is highest where the verifier's next token is most predictable, and math_reasoning is in a class of its own. It leads HumanEval by about 1.3 tokens and still accepts nearly three-quarters of pos_2 drafts (72.6%). Its last position (pos_6, 44.6%) sits above HumanEval's pos_4 (35.5%). math_reasoning is the one subset that makes real use of the full block. HumanEval is clearly second at 4.27, well ahead of everything else.

Throughput (tokens/s)

Output throughput (tokens/s) on a single A100 at max concurrency 1, across the same nine subsets. Baseline is gemma-4-E2B-it with no speculative decoding. Speedup is the mean throughput relative to baseline, and the fastest drafter in each row is bolded.

subset	baseline (no drafter)	MTP	DSpark	DFlash2 (this model)
math_reasoning	167.6	520.2	626.5	626.8
HumanEval	164.9	478.5	399.3	515.5
tool_call	155.3	346.7	275.8	321.8
translation	168.2	417.4	298.5	350.1
rag	153.3	270.8	216.4	243.6
question	167.3	266.8	272.2	276.8
writing	166.7	265.2	268.3	270.2
qa	166.2	257.9	261.6	265.8
summarization	159.9	237.0	194.0	220.4
average speedup	1.00×	2.08×	1.90×	2.10×

DFlash2 has the highest average speedup at 2.10× (unweighted mean of per-subset speedups), followed by MTP at 2.08× and DSpark at 1.90×. DFlash2 is fastest on math_reasoning, HumanEval, question, writing, and qa. On HumanEval it reaches 3.13× versus 2.90× for MTP, even though MTP accepts more tokens per step there. On math_reasoning it is effectively tied with DSpark (626.8 vs 626.5 tokens/s, both 3.74×). MTP is fastest on tool_call, translation, rag, and summarization. Against DSpark alone, DFlash2 is ahead on all nine subsets.

Comparison with MTP and DSpark

Acceptance length on the same nine subsets for Gemma's MTP drafter (google/gemma-4-E2B-it-assistant) and the DSpark drafter. The three propose different numbers of tokens per step, so their ceilings differ (9.0 for MTP, 9.0 for DSpark, 8.0 for DFlash2). A longer draft also costs more per step, so acceptance length alone doesn't determine speed. See the throughput section below for that.

subset	MTP (8 draft tokens)	DSpark (8 draft tokens)	DFlash2 (7 draft tokens)	best
math_reasoning	4.867	5.534	5.602	DFlash2
HumanEval	4.680	3.842	4.270	MTP
tool_call	3.208	2.625	2.943	MTP
translation	3.793	2.427	2.926	MTP
rag	3.164	2.311	2.821	MTP
question	2.780	2.496	2.696	MTP
writing	2.785	2.572	2.726	MTP
qa	2.360	2.194	2.305	MTP
summarization	2.418	1.828	2.126	MTP
mean (unweighted)	3.339	2.870	3.157	MTP

MTP has the highest acceptance length on 8 of 9 subsets and on average. DFlash2 leads on math_reasoning, edging out DSpark despite proposing one fewer token per step, and is second on every other subset. DFlash2 is ahead of DSpark on all nine subsets. The mean here is unweighted across subsets, so it differs slightly from the weighted 3.179 above.

Limitations

Works only with gemma-4-E2B-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, so the gains concentrate in math and code. On prose-heavy workloads the tail of the block rarely gets accepted. Real-world speedup depends on your traffic mix. Because verification is lossless, the verifier's own behavior and biases carry through unchanged.

License

Apache 2.0, matching the verifier's license. The speculators training code is Apache-2.0.