AI Practices AI实践 4d ago Updated 3d ago 更新于 3天前 49

Enhancing Goodput in Large-Scale LLM Training with Nonuniform Tensor Parallelism 通过非均匀张量并行增强大规模LLM训练的吞吐量

Nonuniform Tensor Parallelism (NTP) dynamically adjusts the tensor parallelism degree in response to transient GPU unavailability, ensuring sustained goodput by keeping active GPUs productive. The framework integrates dynamic power boosting within scale-up domains, allowing active GPUs to temporarily increase clock frequencies to compensate for resource reductions and prevent synchronization slowdowns. NTP utilizes highly efficient, overlapped tensor resharding techniques that limit operational 提出非均匀张量并行(NTP)框架,通过动态调整张量并行度来应对大规模LLM训练中的GPU瞬态不可用问题。 结合动态功率提升技术,使活跃GPU暂时提高频率以补偿资源减少,防止全局同步导致的性能下降。 引入高效的重叠张量重分片技术,将引入的开销限制在1%以内,保持近最优的计算效率。 在NVIDIA Blackwell及Blackwell Ultra系统上验证,支持最多72个GPU的NVLink互联域内的弹性扩展与高Goodput维持。

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Nonuniform Tensor Parallelism (NTP) dynamically adjusts the tensor parallelism degree in response to transient GPU unavailability, ensuring sustained goodput by keeping active GPUs productive.
  • The framework integrates dynamic power boosting within scale-up domains, allowing active GPUs to temporarily increase clock frequencies to compensate for resource reductions and prevent synchronization slowdowns.
  • NTP utilizes highly efficient, overlapped tensor resharding techniques that limit operational overhead to less than 1%, preserving near-optimal compute efficiency.
  • The system demonstrates resilience on NVIDIA Blackwell and Blackwell Ultra systems, supporting scale-up domains of up to 72 GPUs connected via NVIDIA NVLink.

Why It Matters

This approach addresses a critical bottleneck in large-scale LLM training: the disproportionate impact of minor hardware fluctuations on global throughput. By shifting the focus from raw hardware utilization to "goodput" (useful convergence-driving work), NTP offers a practical solution for maintaining training stability and efficiency in massive, long-running jobs.

Technical Details

  • Dynamic TP Adaptation: Automatically reconfigures tensor parallelism groups (e.g., reducing TP degree from 8 to 7) when GPUs drop out, allowing remaining devices to absorb the workload without stalling the data parallel replica.
  • Power Boosting Integration: Employs dynamic power boosting to increase clock speeds and throughput on active GPUs, offsetting the performance loss from missing hardware and preventing global synchronization delays.
  • Overlapped Resharding: Implements efficient tensor resharding that overlaps communication with computation, limiting introduced overhead to under 1% and maintaining high compute efficiency.
  • Hardware Specifics: Validated on NVIDIA Blackwell and Blackwell Ultra architectures, leveraging NVLink interconnects to support all-to-all communication within single-hop domains of up to 72 GPUs.

Industry Insight

  • Rethinking Infrastructure Metrics: The industry should prioritize "goodput" over raw FLOPS or throughput metrics when designing large-scale training clusters, as useful convergence speed is the true determinant of ROI.
  • Resilience by Design: Future AI infrastructure must assume hardware instability; systems should be built with elastic, dynamic parallelism strategies rather than relying on static, rigid partitioning.
  • Cost Efficiency: Minimizing downtime through dynamic adaptation reduces the need for expensive hot spares or frequent full checkpoint-restarts, lowering the total cost of ownership for frontier model training.

TL;DR

  • 提出非均匀张量并行(NTP)框架,通过动态调整张量并行度来应对大规模LLM训练中的GPU瞬态不可用问题。
  • 结合动态功率提升技术,使活跃GPU暂时提高频率以补偿资源减少,防止全局同步导致的性能下降。
  • 引入高效的重叠张量重分片技术,将引入的开销限制在1%以内,保持近最优的计算效率。
  • 在NVIDIA Blackwell及Blackwell Ultra系统上验证,支持最多72个GPU的NVLink互联域内的弹性扩展与高Goodput维持。

为什么值得看

随着大模型训练规模扩大至数千GPU,硬件故障和瞬态中断成为影响训练效率的关键瓶颈,本文提供的NTP方案为最大化集群有效算力(Goodput)提供了切实可行的工程路径。对于追求极致训练效率和降低停机成本的AI基础设施团队而言,这种无需完全重启即可弹性适应硬件波动的机制具有重要的实践指导意义。

技术解析

  • 动态TP度自适应:当Scale-up域内的GPU发生中断时,系统自动重新配置张量并行组,仅利用剩余的功能性GPU继续计算。例如,从TP=8动态降级为TP=7,避免整个数据并行副本停滞。
  • 动态功率提升集成:在Scale-up域内,活跃GPU可临时运行在更高的时钟频率和吞吐量下,以抵消因部分GPU离线造成的资源缩减,确保数据并行副本间的全局同步不会变慢。
  • 低开销重叠重分片:采用高度优化的张量重分片技术,将其执行时间与计算过程重叠,使得因拓扑变化带来的额外开销低于1%,从而维持极高的计算利用率。
  • 硬件架构适配:针对NVIDIA Blackwell和Blackwell Ultra架构优化,充分利用NVLink高达1,800 GB/s的带宽,支持单跳全对全通信,适用于多达72个GPU的大规模互联域。

行业启示

  • 从Raw Throughput转向Goodput优化:在超大规模训练中,衡量系统效率的核心指标应从单纯的硬件吞吐量转变为“有效收敛工作”的Goodput,系统设计需优先保障计算连续性而非绝对峰值性能。
  • 弹性容错成为标配能力:面对日益庞大的集群,具备动态适应硬件波动、无需人工干预或完整检查点重启的弹性容错机制,将成为下一代AI训练框架的关键竞争力。
  • 软硬协同设计的重要性:充分发挥新一代互连技术(如NVLink)和低功耗管理特性(如动态功率提升)的潜力,需要算法框架与底层硬件架构的深度耦合与联合优化。

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。

LLM 大模型 Training 训练 GPU GPU