Tencent Open-Sources AngelSpec: A Unified Training Framework for MTP and Block-Parallel Speculative Decoding on Hy3 Models
MarkTechPost Michal Sutter
Tencent released AngelSpec, an open-source training framework for speculative-decoding draft models that uses different architectures optimized for different workloads rather than a single universal drafter. The framework includes MTP drafting for conversation (achieving 66.4% acceptance) and DFly block-diffusion drafting for code and math (reaching 4.79 mean accepted length on Hy3-A21B versus 3.69 for prior DFlash). This heterogeneity-aware approach delivers 1.98–2.40× speedup over standard autoregressive decoding across various concurrency levels.
Why it matters
Tencent has released AngelSpec, an open-source torch-native framework for training speculative-decoding draft models across six architectures. It introduces DFly, a block-diffusion drafter with hybrid target conditioning and a hidden-correction autoregressive head, and integrates D-cut for runtime-adaptive verification budgeting. On HY3-295B-A21B with TP=8, DFly-8 delivers a 1.98–2.40× speedup over autoregressive decoding across concurrency 4 to 64. The post Tencent Open-Sources AngelSpec: A Unified Training Framework for MTP and Block-Parallel Speculative Decoding on Hy3 Models appeared first on MarkTechPost.
