README.md

---
license: other
library_name: transformers
pipeline_tag: text-generation
tags:
- speculative-decoding
- dspark
- dflash
- specforge
- sglang
---

# Ling3-DSpark

A DSpark speculator for Ling3. DSpark extends [DFlash](https://github.com/z-lab/dflash) with target-model auxiliary features and a confidence head that dynamically chooses the number of draft tokens. The model was trained with [SpecForge](https://github.com/sgl-project/SpecForge) and is served with [SGLang](https://github.com/sgl-project/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

Launch recipes for this draft on every supported hardware/quantization cell — including the required `--linear-replayssm-cache-len` sizing — with measured speed and accuracy, are in the [SGLang Ling-3.0-flash cookbook](https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-flash).

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

```bash
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> \
  ......
```

## Serving with llama.cpp

Use a llama.cpp build with DSpark support. Replace the model paths, quantization type, and GPU layer counts with values appropriate for your deployment. First convert
  and quantize the target model:

```bash
python convert_hf_to_gguf.py path/to/Ling-3.0-flash \
    --outfile path/to/Ling-3.0-flash-bf16.gguf \
    --outtype bf16 --model-name Ling-3.0-flash

llama-quantize \
    path/to/Ling-3.0-flash-bf16.gguf \
    path/to/Ling-3.0-flash-Q4_K_M.gguf \
    Q4_K_M
```

Then generate the DSpark draft GGUF:

```bash
python convert_hf_to_gguf.py \
    path/to/Ling-3.0-flash-dspark \
    --target-model-dir path/to/Ling-3.0-flash \
    --outtype bf16 \
    --outfile path/to/Ling-3.0-flash-DSpark.gguf
```

Finally, launch the server with the DSpark draft as the speculative model:

```bash
llama-server \
    --model path/to/Ling-3.0-flash-Q4_K_M.gguf \
    --spec-draft-model path/to/Ling-3.0-flash-DSpark.gguf \
    --spec-type draft-dspark --spec-draft-n-max 8 \
    -ngl all -ngld all -fa on
```