Abstract--A vertically integrated, end-to-end, research prototype system combines 288 NorthPole neural inference accelerator cards, offline training algorithms, a high-performance runtime stack, and a containerized inference pipeline to deliver a scalable and efficient cloud inference service. The system delivers 115 peta-ops at 4-bit integer precision and 3.7 PB/s of memory bandwidth across 18 2U servers, while consuming only 30 kW of power and weighing 730 kg in a 0.67 m The system can run 3 simultaneous instances of the 8-billion-parameter open-source IBM Granite-3.3-8b-instruct The system is scalable, modular, and reconfigurable, supporting various model sizes and context lengths, and is ideal for deploying agentic workflows for enterprise AI applications in existing data center (cloud, on-prem) environments. For example, the system can support 18 instances of a 3-billion-parameter model or a single instance of a 70-billion-parameter model. Large language models have become a pervasive form of computing, and while the current paradigm has been to push frontier models for all applications, it is becoming evident that "Faith in God-like large language models is waning" [1]. In fact, by continuing along this trajectory, global energy requirements for AI-focused data centers are projected to reach double-digit percentages of total electricity consumption by 2030, with individual facilities requiring up to 1 gigawatt or more of dedicated power--driving both infrastructure and cooling costs toward potentially unsustainable or unprofitable levels [2] [3]. However, for many business applications, frontier models containing trillions of parameters may prove less useful and cost efficient than much smaller language models with only a tenth or even a hundredth as many parameters [4].

The implementation is open-sourced at this url.

Training foundation models, such as GPT -3 and PaLM, can be extremely expensive, often involving tens of thousands of GPUs running continuously for months.

Training foundation models, such as GPT -3 and PaLM, can be extremely expensive, often involving tens of thousands of GPUs running continuously for months.

Training large language models (LLMs) is fundamentally constrained by limited device memory and costly inter-device communication. Although pipeline parallelism alleviates memory pressure by partitioning models across devices, it incurs activation communication overhead that scales linearly with sequence length, limiting efficiency in long-context training. Recent weight-passing approaches (e.g., WeiPipe) mitigate this by transmitting model weights instead of activations, but suffer from redundant peer-to-peer (P2P) transfers and underutilized intra-node bandwidth. We propose TawPipe--topology-aware weight pipeline parallelism, which exploits hierarchical bandwidth in distributed clusters for improved communication efficiency. TawPipe: (i) groups devices based on topology to optimize intra-node collective and inter-node P2P communication; (ii) assigns each device a fixed shard of model weights and gradients, avoiding redundant transfers; and (iii) overlaps communication with computation to hide latency. Unlike global collective operations used in fully sharded data parallelism (FSDP), TawPipe confines most communication within node boundaries, significantly reducing cross-node traffic. Extensive experiments on up to 24 GPUs with LLaMA-style models show that TawPipe achieves superior throughput and scalability compared to state-of-the-art baselines.

Long context training is crucial for LLM's context extension. Existing schemes, such as sequence parallelism, incur substantial communication overhead. Pipeline parallelism (PP) reduces this cost, but its effectiveness hinges on partitioning granularity. Batch-level PP dividing input samples exhibits high memory consumption in long-context scenario, whereas token-level PP splitting sequences into slices alleviates memory overhead but may incur hardware under-utilization. This trade-off motivates adaptively selecting PP granularity to match resource and workload characteristics. Moreover, sequence length distribution of the real-world dataset exhibits skewness, posing a challenge on PP's workload balance and efficient scheduling. Current static PP scheduling methods overlook the variance of sequence length, leading to suboptimal performance. In this paper, we propose Elastic Pipeline Parallelism (EPP) that orchestrates token-level PP and batch-level PP to adapt to resource and workload heterogeneity. We build InfiniPipe, a distributed training system that unleashes the potential of EPP via (1) a resource-aware and workload-balanced sequence processor that splits long sequences and packs short ones; and (2) a co-optimization methodology that jointly optimizes pipeline schedule and gradient checkpointing via a mechanism named stage-aware chunk-level adaptive checkpointing. Comprehensive experiments demonstrate that InfiniPipe achieves a 1.69x speedup over state-of-the-art systems.

The training efficiency and scalability of language models on massive clusters currently remain a critical bottleneck. Mainstream approaches like ND parallelism are often cumbersome and complex, while flexible alternatives such as the Zero Redundancy Optimizer (ZeRO) are frequently hampered by communication overhead. In this paper, we propose Asynchronous Hierarchical Zero Parallelism (AsyncHZP), a novel asynchronous variant of ZeRO designed to achieve superior performance while maintaining simplicity and memory efficiency. Unlike traditional ZeRO, which employs over-fine-grained sharding that can lead to inefficient communication, AsyncHZP adaptively reshards parameters, gradients, and optimizer states across different replica groups. This strategy optimizes device memory utilization and significantly reduces communication overhead. In addition, we also design a multi-stream asynchronous scheduling method that executes parameter all-gather and gradient reduce-scatter operations in dedicated background threads, effectively overlapping communication with computation while incurring negligible memory fragmentation. Empirical evaluations on both Dense and Mixture-of-Experts (MoE) models confirm that AsyncHZP maintains robust stability at scale. It consistently outperforms classic ND parallelism, achieving state-of-the-art performance without complex strategic tuning, thereby simplifying the path to efficient large-scale training.

The implementation is open-sourced at this url.

Pipeline parallelism (PP) has become a standard technique for scaling large language model (LLM) training across multiple devices. However, despite recent progress in reducing memory consumption through activation offloading, existing approaches remain largely heuristic and coarse-grained, often overlooking the fine-grained trade-offs between memory, computation, and scheduling latency. In this work, we revisit the pipeline scheduling problem from a principled optimization perspective. We observe that prevailing strategies either rely on static rules or aggressively offload activations without fully leveraging the interaction between memory constraints and scheduling efficiency. To address this, we formulate scheduling as a constrained optimization problem that jointly accounts for memory capacity, activation reuse, and pipeline bubble minimization. Solving this model yields fine-grained schedules that reduce pipeline bubbles while adhering to strict memory budgets. Our approach complements existing offloading techniques: whereas prior approaches trade memory for time in a fixed pattern, we dynamically optimize the tradeoff with respect to model structure and hardware configuration. Experimental results demonstrate that our method consistently improves both throughput and memory utilization. In particular, we reduce idle pipeline time by up to 50% under the same per-device memory limit, and in some cases, enable the training of larger models within limited memory budgets.

Pipeline parallelism is widely used to train large language models (LLMs). However, increasing heterogeneity in model architectures exacerbates pipeline bubbles, thereby reducing training efficiency. Existing approaches overlook the co-optimization of model partition, model placement, and workload scheduling, resulting in limited efficiency improvement or even performance degradation. To respond, we propose AdaPtis, an LLM training system that supports adaptive pipeline parallelism. First, we develop a pipeline performance model to accurately estimate training throughput. Second, AdaPtis jointly optimizes model partition, model placement, and workload scheduling policies guided by this performance model. Third, we design a unified pipeline executor that efficiently supports the execution of diverse pipeline strategies. Extensive experiments show that AdaPtis achieves an average speedup of 1.42x (up to 2.14x) over Megatron-LM I-1F1B across various LLM architectures and scales.