AngelSpec

AngelSpec is a unified, torch-native training framework designed for speculative decoding, supporting both autoregressive Multi-Token Prediction (MTP) and block-parallel drafting architectures. It enables independent scaling of inference and training through a disaggregated architecture, facilitating high-performance model deployment.

AngelSpec is a unified, torch-native training framework designed for speculative decoding, supporting both autoregressive Multi-Token Prediction (MTP) and block-parallel drafting architectures. It enables independent scaling of inference and training through a disaggregated architecture, facilitating high-performance model deployment.

• 01Unified training pipeline for 6 draft architectures (DFly, DFlash, DFlare, Eagle3, DSpark, MTP)
• 02Disaggregated training architecture with independent scaling for inference and optimization
• 03Support for long-context training up to 128k tokens via Ulysses sequence parallelism
• 04Online evaluation of speculative decoding performance during training
• 05Document-aware sequence packing with strict cross-document isolation
• 06Acceptance-aligned training objectives including CE, top-k KL, and D-PACE weighting