File size: 3,636 Bytes
a181ec9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | # Algorithm Decision Guide
Speculators currently supports six speculative decoding algorithms: **Eagle-3**, **P-EAGLE**, **DFlash**, **DFlash2**, **DSpark**, and **MTP**. All are lossless -- they produce output from the same distribution as the target model.
## How They Differ
**Eagle-3** predicts draft tokens autoregressively, one at a time.
**P-EAGLE** extends Eagle-3 with parallel multi-token prediction across multiple depths, using COD sampling for memory-efficient training.
**DFlash** predicts all draft tokens in a single forward pass using block-based prediction with anchor points.
**DFlash2** extends DFlash with local dynamic convolutions and a low-rank candidate selector. Its training objective is experimental and follows the upstream DFlash2 candidate-set contract.
**DSpark** extends DFlash with a Markov head, so positions inside a block depend on earlier ones, plus a confidence head that predicts per-position acceptance.
**MTP** finetunes the model's native multi-token prediction head on domain-specific data. Unlike the other algorithms, MTP does not train from scratch -- it starts from pre-existing MTP layers and is only available for models with native MTP support.
Eagle-3, P-EAGLE, DFlash, and DSpark can be paired with any supported verifier model (including quantized variants). DFlash2 requires the verifier's full vocabulary because its selector operates on the complete vocabulary. MTP requires a model with native MTP layers (e.g. Qwen3-Next, Qwen3.5).
## Current Support
| | Eagle-3 | P-EAGLE | DFlash | DFlash2 | DSpark | MTP |
| ------------------- | ------------- | ------------------- | ------------------- | ------------------- | ------------------- | --------------------------- |
| **Draft layers** | Llama-style | Llama-style | Qwen3-style | Qwen3-style | Qwen3-style | Native MTP layers |
| **Verifier models** | Any supported | Any supported | Any supported | Full vocabulary | Any supported | Models with native MTP only |
| **Training mode** | From scratch | From scratch | From scratch | Experimental | From scratch | Finetune existing MTP head |
| **Speculators** | Mature | Newer, growing fast | Newer, growing fast | Experimental | Newer, growing fast | Newer, growing fast |
| **vLLM** | Mature | Newer, growing fast | Newer, growing fast | Native/experimental | Newer, growing fast | Newer, growing fast |
Eagle-3 has been available longer and has broader support in both Speculators and vLLM. The others were added more recently and support is improving rapidly; DFlash2 remains experimental.
## Which Should I Use?
If you're unsure, start with Eagle-3 -- it has the most mature tooling and documentation. If you want parallel multi-token prediction with an Eagle-3-based architecture, try P-EAGLE. If you want to experiment with DFlash's single-forward-pass block prediction approach, the training workflow is the same. DFlash2 adds local convolution and candidate reranking, while DSpark adds intra-block dependencies and confidence scheduling. If your model already has native MTP layers (e.g. Qwen3-Next, Qwen3.5), MTP finetuning lets you improve the existing MTP head on domain-specific data without training a separate draft model.
For more details on each algorithm, see:
- [Eagle-3](eagle3.md)
- [P-EAGLE](peagle.md)
- [DFlash](dflash.md)
- [DFlash2](dflash2.md)
- [DSpark](dspark.md)
- [MTP](mtp.md)
|