AI Practices AI实践 10h ago Updated 1h ago 更新于 1小时前 49

Experiment with Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 for Agentic Coding 在 NVIDIA GB300 NVL72 上实验 Qwen3.8-Flash-Next 用于智能体编程

Qwen3.8-Flash-Next is a 125B-parameter multimodal MoE model (6B activated per token) with a native 262K-token context window extendable to 1M tokens via YaRN, serving as a preview of the upcoming Qwen4 architecture The hybrid Gated DeltaNet (GDN) + Qwen Sparse Attention (QSA) architecture compresses historical context into fixed-size recurrent states across three of four layers, while QSA enables precise retrieval across full context using block-level importance estimation Alibaba benchmarks sho Qwen3.8-Flash-Next是Alibaba发布的多模态MoE模型(125B总参数,6B激活参数),原生支持262K上下文窗口,可通过YaRN扩展至1M tokens 采用GDN+QSA混合架构:每4层中3层使用Gated DeltaNet压缩历史上下文为固定大小状态,1层使用Qwen Sparse Attention进行精确检索 QSA在1M-token工作负载上相比全注意力机制实现7.6x prefill和4.9x decoding加速,在90%前缀缓存命中率下prefill吞吐量达Qwen3.7-Plus的8.6倍 在NVIDIA GB300 NVL72平台上实现每GPU超16K

68
Hot 热度
72
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Qwen3.8-Flash-Next is a 125B-parameter multimodal MoE model (6B activated per token) with a native 262K-token context window extendable to 1M tokens via YaRN, serving as a preview of the upcoming Qwen4 architecture
  • The hybrid Gated DeltaNet (GDN) + Qwen Sparse Attention (QSA) architecture compresses historical context into fixed-size recurrent states across three of four layers, while QSA enables precise retrieval across full context using block-level importance estimation
  • Alibaba benchmarks show QSA delivers up to 7.6x prefill and 4.9x decoding speedups over full attention at 1M-token context, achieving 8.6x the prefill throughput of Qwen3.7-Plus with a 90% prefix-cache hit rate
  • Running on NVIDIA GB300 NVL72 achieves over 16K tokens/second per GPU and over 200 tokens/second per user, enabling high-throughput, low-latency agentic coding applications
  • NVIDIA provides Day-0 support through SGLang, vLLM, and TensorRT LLM, with post-training recipes via NeMo AutoModel (SFT/LoRA) and NeMo RL for reinforcement learning

Why It Matters

This model represents a significant architectural advance for long-context, high-throughput inference workloads critical to agentic AI systems, demonstrating that hybrid recurrent-sparse attention designs can dramatically reduce the compute and memory bottlenecks that traditionally limit large-context applications. For AI practitioners, the combination of open weights, Day-0 NVIDIA hardware support, and multiple inference engine options lowers the barrier to deploying production-grade coding agents at scale.

Technical Details

  • Architecture: Multimodal MoE with 125B total parameters, 6B activated per token, supplemented by 51B N-gram embeddings. Hybrid design alternating three GDN layers (compressing context into fixed-size recurrent states) with one QSA layer (block-level sparse attention for precise retrieval)
  • Context handling: Native 262,144-token context window, extensible to 1M tokens via YaRN. QSA aggregates sequences into micro-blocks, estimates importance at block level, and selects only relevant regions, eliminating token-level indexer overhead
  • Performance benchmarks: 7.6x prefill and 4.9x decoding speedups over full attention at 1M tokens; 8.6x prefill throughput vs. Qwen3.7-Plus at 1M context with 90% prefix-cache hit rate
  • Hardware deployment: Validated on NVIDIA GB300 NVL72 (72 Blackwell Ultra GPUs, 130 TB/s NVLink), delivering >16K tokens/sec/GPU and >200 tokens/sec/user. Also supports DGX Station, DGX Spark, and RTX PRO 6000 Blackwell workstations
  • Tooling ecosystem: NVIDIA NeMo AutoModel for PyTorch-native fine-tuning (full SFT and LoRA), NeMo RL recipes for reinforcement learning, and open-source inference via SGLang, vLLM, and TokenSpeed with Day-0 Hugging Face checkpoint support

Industry Insight

  • The GDN+QSA hybrid architecture signals a strategic shift away from pure dense attention toward recurrent-sparse combinations for long-context workloads, suggesting future models will prioritize context efficiency over raw parameter scaling for agentic applications
  • NVIDIA's Day-0 support across multiple inference engines and hardware tiers (from workstations to rack-scale GB300 NVL72) creates a clear upgrade path for developers to prototype locally and scale to production without model re-engineering
  • The emphasis on agentic coding as a flagship use case indicates that the industry is prioritizing high-throughput, low-latency, tool-driven workflows over single-turn generation, which will likely accelerate investment in MoE architectures and sparse attention mechanisms for production deployment

TL;DR

  • Qwen3.8-Flash-Next是Alibaba发布的多模态MoE模型(125B总参数,6B激活参数),原生支持262K上下文窗口,可通过YaRN扩展至1M tokens
  • 采用GDN+QSA混合架构:每4层中3层使用Gated DeltaNet压缩历史上下文为固定大小状态,1层使用Qwen Sparse Attention进行精确检索
  • QSA在1M-token工作负载上相比全注意力机制实现7.6x prefill和4.9x decoding加速,在90%前缀缓存命中率下prefill吞吐量达Qwen3.7-Plus的8.6倍
  • 在NVIDIA GB300 NVL72平台上实现每GPU超16K tokens/秒、每用户超200 tokens/秒的推理性能,适用于高吞吐低延迟的Agentic Coding场景
  • 提供NVIDIA NeMo AutoModel微调、NeMo RL强化学习及SGLang/vLLM/TokenSpeed等开源推理引擎支持

为什么值得看

本文展示了长上下文推理架构的创新实践,GDN+QSA混合设计有效解决了大模型在超长上下文场景下的KV Cache内存和计算瓶颈,为Agentic AI等上下文密集型应用提供了可落地的技术方案。同时,该模型作为Qwen4架构预览版,结合NVIDIA GB300 NVL72的Day-0支持,为开发者提供了从原型到生产的完整工具链参考。

技术解析

  • 混合架构设计:模型采用3:1的GDN与QSA层交替结构,GDN将历史上下文压缩为固定大小的循环状态以消除KV Cache随序列增长的开销,QSA通过微块聚合和块级重要性估计实现精确检索,两者结合兼顾效率与准确性。
  • 长上下文性能优化:QSA通过块级注意力替代token级索引,显著降低计算和索引开销。在1M-token场景下,相比全注意力机制实现7.6x prefill和4.9x decoding加速,在90%前缀缓存命中率下prefill吞吐量达到Qwen3.7-Plus的8.6倍。
  • 硬件部署性能:在NVIDIA GB300 NVL72(72 GPU NVLink域,130 TB/s全互联带宽)上实现每GPU超16K tokens/秒、每用户超200 tokens/秒的推理吞吐,支持从本地DGX/RTX工作站到机架级部署的平滑扩展。
  • 开发生态支持:提供NVIDIA NeMo AutoModel(支持PyTorch原生、Day-0 Hugging Face checkpoint、SFT和LoRA微调)和NeMo RL recipes(强化学习),以及SGLang、vLLM、TokenSpeed等开源推理引擎的部署方案。

行业启示

  • 长上下文成为Agentic AI的关键基础设施:GDN+QSA架构验证了通过混合注意力机制解决超长上下文推理瓶颈的可行性,未来Agentic应用(代码生成、文档处理、工具调用)将更依赖此类高效长上下文方案。
  • MoE模型与专用硬件的深度协同:125B参数MoE模型在GB300 NVL72上实现高吞吐低延迟,表明大模型部署正从"通用GPU集群"向"专用 rack-scale 架构"演进,硬件-模型协同设计将成为性能突破的核心路径。
  • 开源生态加速模型落地:从Hugging Face权重下载、NeMo微调工具链到多推理引擎支持,Alibaba与NVIDIA的合作构建了完整的开发-部署闭环,降低了企业级Agentic应用的上手门槛。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 LLM 大模型 Code Generation 代码生成 Agent Agent Multimodal 多模态