Nemotron 3 Super 技术报告解读:LatentMoE 与 NVFP4 预训练

解读 NVIDIA Nemotron 3 Super 的三大技术点:在隐空间做专家路由、同时优化精度与 FLOPs 的 LatentMoE,首次大规模稳定训练到 25T token 的 NVFP4 预训练,以及以 Mamba-2 为主的混合 Mamba-Attention 加 MoE 架构。

Nemotron 3 Super —技术报告摘要

论文:Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
团队:NVIDIA
发布时间:2026.04
页数:51页


1. 预训练 Recipe

训练策略

项目详情
模型架构Hybrid Mamba-Attention MoE,120B总参 / 12B激活
层数88层,周期性交错模式(Mamba-2block +LatentMoE layer为主,少量全局Attention作为anchor)
注意力GQA:32Q-heads, 2KV-heads, head dim128;无位置编码、dropout、bias
Mamba-2State dim 128, 8 groups, 128 heads, head dim 64
LatentMoE512 total experts/layer, top-22 activated, latent size 1024, expert hidden2688, shared expert 5376
LatentMoE核心设计优化accuracy per FLOP accuracy per parameter(考虑memory bandwidth/communication/sharding开销);routed experts在compact latent space操作,shared experts走full-width path
MTP2层(shared weight),用于speculative decoding加速推理;平均acceptance length 3.33(vs Qwen3的2.70)
训练精度NVFP4预训练(E2M1+ 16-element micro-blocks + E4M3 scaling);最后15%层+QKV+Mamba output+MTP+Latent projections保持BF16/MXFP8
上下文1M tokens
训练Token数25T(两阶段)
LR schedule原文未详述(继承Nemotron 3 Nano recipe)

训练数据

项目详情
总量25T tokens,分两阶段
Phase 1(80%,20T)多样性和广泛覆盖
Phase 2(20%,5T)高质量数据+benchmark accuracy
合成数据集(开源)Nemotron-Pretraining-Specialized-v1.1:code concepts/algorithms, formal verification, college-level STEM reasoning, creative/factual long-form writing等
合成方法用Qwen3-235B、DeepSeek-R1等生成+验证

Infra设计

项目详情
训练精度NVFP4 GEMMs(fprop/dgrad/wgrad),Transformer Engine + cuBLAS后端
量化方案2D block scaling(weights),1D block scaling(gradients/activations),Random Hadamard Transforms + stochastic rounding
大规模稳定性25T tokens稳定训练无发散;监控zero-valued gradient元素比例确认训练健康

2. 后训练 Recipe (SFT)

训练策略

项目详情
两阶段SFTStage 1: token-level global average loss(学习reasoning behavior)→ Stage 2: sample-level per-conversation normalized loss(恢复long-input-short-output性能)
重点大幅扩展agentic datasets比例
低effort模式新增low effort reasoning mode
Chat template与Nemotron 3 Nano相同

训练数据

项目详情
SFT数据780M samples,开放源为Nemotron-Post-Training-Data
Agentic CLI数据15K synthesis tasks + 3K SWE tasks + 10K web dev tasks;用Qwen3 Coder/OpenCodex/Agentic LLMs录制交互轨迹
合成数据Economics, formal logic, MCQ (3.5M samples ~1.6B tokens), CUDA kernel (100K), financial reasoning (366K)等

3. RL Recipe

训练策略

项目详情
三部分RLVR(3轮)+RLHF + SWE-RL
RLVR Round125 environment types
RLVR Round 226 environment types with Low Effort
RLVR Round 337 environment types, up to 4000 environment instances per batch, Agentic Focused
RLHF18B tokens
SWE-RL20B tokens,长程SWE任务
MTP HealingRL后修复MTP层(RL可能损坏MTP质量)
总规模21+ diverse RL environments,thousands of GPUs

Infra设计

项目详情
RL基础设施大规模异步训练,支持千GPU级别
关键改进提高RL infra resiliency,支持long-horizon agentic tasks
环境多样性37种环境类型,每batch最多4000实例

4. 关键亮点

  1. LatentMoE:首个同时优化accuracy/FLOP和accuracy/parameter的MoE设计,512experts/layer top-22
  2. NVFP4预训练:首次大规模稳定NVFP4训练至25T tokens
  3. Hybrid Mamba-Attention + MoE:Mamba-2为主+少量Attention anchor,1M上下文
  4. MTP for speculative decoding:acceptance length 3.33,显著加速推理
  5. Agentic RL at scale:37种环境类型×4000实例/batch,SWE-RL 20B tokens
  6. 推理效率:比GPT-OSS-120B快2.2×,比Qwen3.5-122B快7.5×(8k/64k ISL/OSL)
  7. 两阶段SFT loss:token-level→sample-level,解决long-input-short-output退化
本文结束 感谢您的阅读