inclusionAI 发布 Ling-3.0-flash 的 DSpark 投机解码模型 Ling3-DSpark

inclusionAI/Ling-3.0-flash-dspark

AI 导读

inclusionAI 推出 Ling3-DSpark，一个为 Ling-3.0-flash 设计的 DSpark 投机解码模型，参数量 1.36B，通过置信度头动态选择草稿 token 数量。在 GSM8K 等九项基准上平均接受长度为 5.29，其中 GSM8K 达 6.40。模型经 SpecForge 训练，可通过 SGLang 部署。

推荐理由

该模型把投机解码的接受长度按工作负载分开报告，数学与代码任务明显高于对话类，部署时可据此对不同场景的加速收益有更实际预期。

正文 · 原文

Ling3-DSpark

A DSpark speculator for Ling3. DSpark extends DFlash with target-model auxiliary features and a confidence head that dynamically chooses the number of draft tokens. The model was trained with SpecForge and is served with SGLang.

Model specifications

Target model: Ling-3.0-flash
Draft parameters: 1,363,707,905 (1.36B)
Draft weight dtype: BF16
Hidden size: 2,560
Transformer layers: 5 full-attention layers
Attention: MHA with 32 query heads and 32 key/value heads
Target auxiliary feature layers: 1, 11, 23, 29, 35
Confidence head: vanilla Markov head, rank 256
DSpark block size: 8 draft tokens (verify width 9, including the target bonus token)
Maximum position embeddings: 262,144

Acceptance length

Acceptance length is the mean number of tokens accepted per speculative verification step, including the target bonus token.

Workload	Acceptance length
GSM8K	6.40
MATH-500	6.29
AIME 2025	5.56
HumanEval	6.57
MBPP	6.34
LiveCodeBench	5.33
MT-Bench	3.92
Alpaca	3.51
Arena-Hard-v2	3.72

The macro mean across the nine workload means is 5.29.

Serving with SGLang

Use an SGLang version with DSPARK support. Replace the model paths and tensor-parallel size with values appropriate for your deployment:

sglang serve \
  --trust-remote-code \
  --model-path <LING3_MODEL_PATH> \
  --tp-size <TP_SIZE> \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path <LING3_DSPARK_MODEL_PATH> \
  ......

1B params