A unified training framework for MTP and block-parallel speculative decoding

AngelSpec is a torch-native framework for training speculative-decoding draft models, covering both autoregressive MTP drafting and the block-parallel DFlash family. It is the training framework behind the AngelSpec technical report: all drafters in the report — the TTT-trained MTP drafter and the DFly family — are trained and released with it.

Latest News

• [2026/07/29] We release AngelSpec v0.1.0, supporting MTP and block-parallel speculative decoding training. Check out our technical report and released models.

Key Features

• 6 draft architectures — DFly, DFlash, DFlare, Eagle3, DSpark, MTP — behind one training pipeline; switching is a config change
• MTP training with TTT — on-policy multi-depth rollout with memory close to a single causal pass; long-context training up to 128k via Ulysses sequence parallelism
• Acceptance-aligned objectives — CE, top-k KL, LK losses, D-PACE weighting, and end-to-end TV, composable through configuration
• Document-aware sequence packing — Megatron-style fixed-length packing with strict cross-document isolation, on both the DFlash and MTP paths
• Online evaluation — genuine speculative decoding against the latest checkpoint during training, reporting mean accepted length and per-position acceptance as measured by the serving engine

Architecture

Inference and training run as separate GPU worker groups connected by a Mooncake tensor store, so hidden-state generation and optimization scale independently. This disaggregated foundation comes from TorchSpec; AngelSpec extends it with the architectures, objectives, and training features above.

Draft Architectures

Architecture	Method	Key Idea
DFly	Block-parallel	Hybrid target conditioning + hidden-correction AR head
DFlash	Block-parallel	Anchor sampling + parallel block generation
DFlare	Block-parallel	DFlash + learnable per-layer target fusion
Eagle3	Autoregressive TTT	Test-time training with input fusion
DSpark	Hybrid	DFlash backbone + EAGLE-style autoregressive head
MTP	Single-head TTT	Full MoE decoder layer as draft (Hy3-native)

See the draft-model family docs for details and trade-offs.

Quick Start

# Install AngelSpec + the vLLM backend
pip install -e ".[vllm]"
pip install mooncake-transfer-engine

# Single-node quickstart (8 GPUs: 4 inference + 4 training)
./examples/qwen3-8b-dfly/run.sh

# Override config values from CLI
./examples/qwen3-8b-dfly/run.sh training.learning_rate=5e-5 training.num_train_steps=500

Or set up a conda environment (installs mooncake too):

./tools/build_conda.sh 1 vllm     # or: sglang
micromamba activate angelspec

CUDA 12.x hosts: PyPI's default torch / vllm wheels target CUDA 13 and won't load on a CUDA-12 driver. Install CUDA-matched wheels first — see Installation.

Examples

Multi-node (Hy3 target)

Example	Architecture	Mode	Released Model
hy3-dfly	DFly	From scratch	AngelSlim/Hy3-DFly-Block8
hy3-mtp	MTP	From scratch	AngelSlim/Hy3-MTP-TTT3

Single-node (Qwen3-8B target)

Example	Architecture	Mode	Released Model
qwen3-8b-dspark	DSpark	From scratch	—
qwen3-8b-dfly	DFly	From scratch	AngelSlim/Qwen3-8B-DFly-Block8
qwen3-8b-mtp	MTP	From scratch	AngelSlim/Qwen3-8B-MTP-TTT3
qwen3-8b-dfly-cpt	DFly	CPT (from ckpt)	Continue from AngelSlim/Qwen3-8B-DFly-Block8

Available training configs

Config	Architecture	Features
configs/vllm_qwen3_8b_dfly.yaml	DFly	Block-parallel + hidden correction
configs/vllm_qwen3_8b_dflare.yaml	DFlare	Per-layer target fusion
configs/sglang_qwen3_8b_dflash.yaml	DFlash	SGLang backend
configs/sglang_qwen3_8b_dspark.yaml	DSpark	Markov head + confidence
configs/vllm_qwen3_8b_mtp_pack_usp_40k.yaml	MTP	Packing + USP 40k seq

Released Models

Draft models trained with AngelSpec:

Model	Architecture	Target	Mode
AngelSlim/Hy3-DFly-Block8	DFly	Hy3	No-think
AngelSlim/Hy3-DFly-Block8-Think-High	DFly	Hy3	High-think
AngelSlim/Hy3-MTP-TTT3	MTP	Hy3	No-think
AngelSlim/Qwen3-8B-DFly-Block8	DFly	Qwen3-8B	No-think
AngelSlim/Qwen3-8B-MTP-TTT3	MTP	Qwen3-8B	No-think

Benchmark

Offline Throughput (HY3-295B-A21B, TP=8)

Output-token throughput (Tok/s) and speedup relative to AR at temperature 1 across concurrency levels. Each cell uses 3 × 120 s windows; Avg. is the arithmetic mean across six datasets.

Live Traffic Throughput (D-cut)

D-cut on Hy3 live traffic (Hy3-295B-A21B, TP=8, 8× H20; concurrency 2–64). (a) Aggregate throughput vs. per-user decode speed — points up and to the right are better. (b) Aggregate throughput vs. concurrency — DFly saturates beyond concurrency 48, whereas D-cut continues to convert additional load into throughput.

License

This project is released under LICENSE. AngelSpec is built upon TorchSpec by LightSeek Foundation and uses Mooncake for disaggregated hidden-state transfer.

Citation

If you find AngelSpec useful, please cite:

@article{angelspec2026,
  title   = {AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding},
  author  = {Liu, Hong and Cen, Rui and Shi, Junhan and Qin, Guangshuo and Zhang, Jiebin and Liu, Tianyu and Fan, Runzhi and Zhao, Guoliang and Xie, Ruobing and Zhang, Kai and Liu, Song and Yu, Guanghua and Zhu, Jianchen},
  journal = {arXiv preprint arXiv:2607.25852},
  year    = {2026}
}