Tencent Open-Sources AngelSpec: A Unified Training Framework for MTP and Block-Parallel Speculative Decoding on Hy3 Models
Tencent
Tencent has released AngelSpec, an open-source training framework for speculative-decoding draft models, supporting both autoregressive multi-token prediction (MTP) and block-parallel DFlash family. The framework addresses workload heterogeneity by specializing structure, training data, and verification depth for different domains. Early results show DFly achieving up to 4.79 mean accepted length on Hy3-A21B, a 59.7% improvement over MTP.
Tencent has open-sourced AngelSpec, a torch-native training framework for speculative-decoding draft models. It covers both autoregressive multi-token prediction (MTP) and the block-parallel DFlash family, including a new architecture DFly. AngelSpec treats workload heterogeneity as a first-class constraint, offering complementary drafters for different domains: MTP for conversation-oriented data and block-diffusion for code/math. The MTP path uses a shared-parameter, multi-depth scheme with target-model rollout to address train-inference mismatch. DFly introduces hybrid target conditioning and a predecessor-conditioned autoregressive head. On Qwen3-8B, DFly reaches 5.41 average mean accepted length, and on Hy3-A21B, 4.79 versus 3.69 for DFlash and 3.00 for MTP. The framework is built on TorchSpec with features like TTT rollout, long-context training up to 128k tokens, and document-aware sequence packing. Seven checkpoints are available on Hugging Face and ModelScope.
- Сокращения
- MTP = Multi-Token Prediction — Предсказание нескольких токенов
- TTT = Training-Time Test — Тест во время обучения
Source: MarkTechPost —
original