Tencent open-sources AngelSpec for speculative decoding drafter training
aigclink · x · 2026-07-29
- Tencent open-sourced AngelSpec, a framework for training speculative-sampling drafter models.
- It consolidates six draft architectures into one configurable training pipeline: DFly, DFlash, DFlare, Eagle3, DSpark, and MTP.
- The system separates inference and training workers, uses Mooncake tensor storage for RDMA transfer, and can validate drafts with real serving engines like vLLM or SGLang during training.
- In Tencent’s report on a Hy3-295B-A21B MoE, DFly delivered 1.98x–2.40x average speedups at concurrency 4–64, peaking at 2.86x on code/math tasks, and outperformed DFlash by 10.5%–11.8%.
- The live-traffic evaluation showed D-cut continuing to convert added load into throughput after DFly saturated at concurrency 48.
Related event: Tencent Open-Sources AngelSpec Speculative Decoding Framework(4 posts)→
More from Infra
- Redditor Asks: Why No Video Models Built for 24-32GB Consumer VRAM? — BreadUndPeeTears · 2026-10-16
- Rapid-MLX 0.16.2 brings NVIDIA Nemotron 3.5 up to 120B to Apple Silicon Macs — kuchaev · 2026-10-16
- Adobe ships NVIDIA RTX Spark optimizations: up to 2x faster creative apps and 5x Frame.io gains — shashib · 2026-10-16
- Meta's commitments jumped to $349B in a quarter as server memory prices tripled ahead of Oct 21 earnings — tengyanAI · 2026-10-16
- Meta releases Muse Spark Open, trained with vLLM-powered RL inference from day one — colinmcnamara · 2026-10-16
- a16z: bot 403 blocks double as Amazon cracks down on agentic commerce — a16z Newsletter · 2026-10-16