Hunyuan-TurboS 技术报告解读:Mamba-Transformer 协同与自适应长短思维链

解读腾讯混元 TurboS 的两大创新:Mamba 与 Transformer 协同的混合架构,以及让单一模型自动选择长思考或短思考模式的自适应 CoT 融合方案,并梳理其后训练四大模块。

Hunyuan-TurboS技术报告摘要

论文:Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought
团队:Tencent Hunyuan Team
时间:2025.05/2025.07(arXiv:2505.15431)
页数:26页


1. 预训练 Recipe

训练策略

项目详情
架构Hybrid Transformer-Mamba2MoE,560B总参/56B激活,128层
层组成57Mamba2层+ 7 Attention层 + 64 FFN(MoE)层
Block Pattern“AMF”(Attention→Mamba2→FFN) + “MF”(Mamba2→FFN) 交替
MoE32 specialized experts + 1 shared expert,激活1shared+2specialized/token
AttentionGQA (最小化KV cache)
Mamba2State-Space Model,线性复杂度处理长序列
Token16T high-quality tokens
上下文256K
Tokenizer与Hunyuan-Large相同,128K词表
里程碑首个工业部署的大规模Mamba模型

训练数据

项目详情
总量16T tokens
相比Hunyuan-Large改进三维度增强:quantity, quality, diversity
数据PipelineURL-level dedup/filtering → block-level pruning(去噪内容)→ topic classification → content extraction → domain-specific extraction modules(STEM/code) → heuristic filtering → global semantic-level dedup
质量保障综合critique models + 数据mixture models(不同训练阶段提供不同mixture recipe)
多样性dozens of well-organized domain type labels

Infra设计

项目详情
RL框架Angel-RL:集成training和inference,支持TP/PP/EP/CP全面parallelism
Hybrid优化针对TurboS hybrid架构实现三维度优化,实现相比Hunyuan-Large 1.8x加速

2. 后训练 Recipe(四模块)

Module 1: Supervised Fine-Tuning

项目详情
数据量3M instruction data (natural + synthetic)
组织方式按comprehensive topic taxonomy分类
质量控制multi-dimensional metrics for rigorous filtering

Module 2: Adaptive Long-Short CoT Fusion(核心创新)

目的:单一模型自动选择long/short CoT模式

SFT Training流程:

  1. 先训练short CoT模型(Hunyuan-Base + reasoning data)
  2. 对所有reasoning data做inference + consistency check
  3. 如果short CoT正确→直接作为训练样本
  4. 如果short CoT错误→将问题+错误response喂给Hunyuan-T1继续生成扩展推理→转换为short CoT风格→重复直到获得正确答案
  5. 拼接所有失败尝试+最终正确response作为adaptive teacher的训练数据
  6. 用此数据训练Hunyuan-Base得到adaptive SFT model

RL for Adaptive Long-Short CoT:

① Difficulty-Adaptive Reward

  • GRPO sampling生成不同reasoning深度的responses
  • Online rejection sampling评估prompt difficulty
  • 根据难度分配:complex→long CoT,simple→short CoT

② Long CoT Compression Reward

  • 对长reasoning chain施加length penalty
  • 多条等正确性的reasoning paths中,shorter traces获得higher reward
  • 减少冗余同时保持准确性

Module 3: Multi-round Deliberation Learning

项目详情
方法迭代式改进(iterative improvement)
具体机制原文在此部分简略描述

Module 4: Two-Stage Large-Scale RL

项目详情
Stage 1STEM领域RL
Stage 2General instruction-following RL
框架Angel-RL

3. RL Recipe

训练策略

项目详情
算法GRPO
两阶段STEM RL → General RL
Adaptive RewardDifficulty-adaptive + CoT compression reward
框架Angel-RL (TP/PP/EP/CP)

Infra:Angel-RL

项目详情
特点集成training和inference为一体
并行全面parallelism (TP/PP/EP/CP)
创新针对hybrid Mamba-Transformer架构的专门优化

4. 关键亮点

  1. Hybrid Mamba2-Transformer-MoE:128层(57 Mamba + 7 Attention + 64 MoE FFN),首个工业部署的大规模Mamba模型
  2. AMF/MF Block Pattern:Attention→Mamba2→FFN交替,平衡性能和效率
  3. Adaptive Long-Short CoT Fusion:单模型自动选择reasoning depth,simple问题快速回答,complex问题深度推理
  4. Difficulty-Adaptive + Compression Reward:RL中按难度分配CoT模式+长CoT压缩奖励
  5. 16T高质量tokens:完整数据pipeline from dedup to semantic filtering
  6. 256K上下文:Mamba2线性复杂度天然适合长序列
  7. LMSYS Arena Top7 (1356):超过Gemini-2.0-Flash和o4-mini
  8. Angel-RL:支持TP/PP/EP/CP的integratedRL框架
本文结束 感谢您的阅读