Distributed LLM inference requires careful coordination of parallelization strategies across hundreds to thousands of NPUs to meet production SLOs. Current systems like Megatron-LM rely on static heuristics that separately configure parallelism degrees and per-operator sharding dimensions, leaving significant performance on the table as models scale and hardware topologies diversify. We introduce Learn to Shard, to our knowledge, the first RL-based approach to co-optimize both coarse-grained parallelism degrees and fine-grained per-operator sharding dimensions for distributed LLM inference. Our method employs an attention-based policy over an elite history that learns from high-performing strategies to efficiently navigate the vast combinatorial search space. Evaluated on H100 clusters with MoE models up to 1.6T parameters, Learn to Shard achieves up to 3.5x throughput improvement over metaheuristic baselines and 1.06x over Megatron heuristics.

Reinforcement learning (RL) is a paradigm increasingly used to align large language models. Popular RL algorithms utilize multiple workers and can be modeled as a graph, where each node is the status of a worker and each edge represents dataflow between nodes. Owing to the heavy cross-node dependencies, the RL training system usually suffers from poor cluster scalability and low memory utilization. In this article, we introduce MindSpeed RL, an effective and efficient system for large-scale RL training. Unlike existing centralized methods, MindSpeed RL organizes the essential data dependencies in RL training, i.e., sample flow and resharding flow, from a distributed view. On the one hand, a distributed transfer dock strategy, which sets controllers and warehouses on the basis of the conventional replay buffer, is designed to release the dispatch overhead in the sample flow. A practical allgather--swap strategy is presented to eliminate redundant memory usage in resharding flow. In addition, MindSpeed RL further integrates numerous parallelization strategies and acceleration techniques for systematic optimization. Compared with existing state-of-the-art systems, comprehensive experiments on the RL training of popular Qwen2.5-Dense-7B/32B, Qwen3-MoE-30B, and DeepSeek-R1-MoE-671B show that MindSpeed RL increases the throughput by 1.42 ~ 3.97 times. Finally, we open--source MindSpeed RL and perform all the experiments on a super pod of Ascend with 384 neural processing units (NPUs) to demonstrate the powerful performance and reliability of Ascend.

Galvatron is a distributed system for efficiently training large-scale Foundation Models. It overcomes the complexities of selecting optimal parallelism strategies by automatically identifying the most efficient hybrid strategy, incorporating data, tensor, pipeline, sharded data, and sequence parallelism, along with recomputation. The system's architecture includes a profiler for hardware and model analysis, a search engine for strategy optimization using decision trees and dynamic programming, and a runtime for executing these strategies efficiently. Benchmarking on various clusters demonstrates Galvatron's superior throughput compared to existing frameworks. This open-source system offers user-friendly interfaces and comprehensive documentation, making complex distributed training accessible and efficient.

As the artificial intelligence community advances into the era of large models with billions of parameters, distributed training and inference have become essential. While various parallelism strategies-data, model, sequence, and pipeline-have been successfully implemented for popular neural networks on main-stream hardware, optimizing the distributed deployment schedule requires extensive expertise and manual effort. Further more, while existing frameworks with most simple chain-like structures, they struggle with complex non-linear architectures. Mixture-of-experts and multi-modal models feature intricate MIMO and branch-rich topologies that require fine-grained operator-level parallelization beyond the capabilities of existing frameworks. We propose formulating parallelism planning as a scheduling optimization problem using mixed-integer programming. We propose a bi-level solution framework balancing optimality with computational efficiency, automatically generating effective distributed plans that capture both the heterogeneous structure of modern neural networks and the underlying hardware constraints. In experiments comparing against expert-designed strategies like DeepSeek's DualPipe, our framework achieves comparable or superior performance, reducing computational bubbles by half under the same memory constraints. The framework's versatility extends beyond throughput optimization to incorporate hardware utilization maximization, memory capacity constraints, and other considerations or potential strategies. Such capabilities position our solution as both a valuable research tool for exploring optimal parallelization strategies and a practical industrial solution for large-scale AI deployment.

To improve the efficiency of distributed large language model (LLM) inference, various parallelization strategies, such as tensor and pipeline parallelism, have been proposed. However, the distinct computational characteristics inherent in the two stages of LLM inference-prefilling and decoding-render a single static parallelization strategy insufficient for the effective optimization of both stages. In this work, we present Seesaw, an LLM inference engine optimized for throughput-oriented tasks. The key idea behind Seesaw is dynamic model re-sharding, a technique that facilitates the dynamic reconfiguration of parallelization strategies across stages, thereby maximizing throughput at both phases. To mitigate re-sharding overhead and optimize computational efficiency, we employ tiered KV cache buffering and transition-minimizing scheduling. These approaches work synergistically to reduce the overhead caused by frequent stage transitions while ensuring maximum batching efficiency. Our evaluation demonstrates that Seesaw achieves a throughput increase of up to 1.78x (1.36x on average) compared to vLLM, the most widely used state-of-the-art LLM inference engine.

The communication Our aim is to devise methodologies for the online adaptation patterns induced by these training process exhibit of routing configurations in ML training clusters that high regularity and persistence, giving rise to significant improve global training efficiency and fairness. Our approach opportunities for optimizing the manner in which flows are builds on two characteristics of ML training and modern networking: routed across the network. We present an algorithmic framework for quantifying network-wide efficiency in the context of training LLMs (and other large-scale ML models), and for periodically optimizing routing with respect to this global Traffic patterns induced by ML training tend to exhibit metric.

Strengths: The paper picks a very relevant problem. While the stochastic gradient method is easily parallelizable, parallelization strategies for the otherwise beneficial in terms of convergence rates approximate proximal methods are not known yet. The paper proposes natural approaches to parallelize the proximal methods: to compute the updates by independent workers and perform the averaged update (PIA), to build a model for the averaged function and compute the update which can be decomposed as a sum of independently computed terms (PMA), or to build an average of the models and compute an update, which can be parallelized through solving a dual problem (PAM). The paper clearly explains these strategies. It proves upper bounds for the errors in case of smooth and non-smooth functions.

Dramatic increases in the capabilities of neural network models in recent years are driven by scaling model size, training data, and corresponding computational resources. To develop the exceedingly large networks required in modern applications, such as large language models (LLMs), model training is distributed across tens of thousands of hardware accelerators (e.g. GPUs), requiring orchestration of computation and communication across large computing clusters. In this work, we demonstrate that careful consideration of hardware configuration and parallelization strategy is critical for effective (i.e. compute- and cost-efficient) scaling of model size, training data, and total computation. We conduct an extensive empirical study of the performance of large-scale LLM training workloads across model size, hardware configurations, and distributed parallelization strategies. We demonstrate that: (1) beyond certain scales, overhead incurred from certain distributed communication strategies leads parallelization strategies previously thought to be sub-optimal in fact become preferable; and (2) scaling the total number of accelerators for large model training quickly yields diminishing returns even when hardware and parallelization strategies are properly optimized, implying poor marginal performance per additional unit of power or GPU-hour.

Generative AI, in particular large transformer models, are increasingly driving HPC system design in science and industry. We analyze performance characteristics of such transformer models and discuss their sensitivity to the transformer type, parallelization strategy, and HPC system features (accelerators and interconnects). We utilize a performance model that allows us to explore this complex design space and highlight its key components. We find that different transformer types demand different parallelism and system characteristics at different training regimes. Large Language Models are performant with 3D parallelism and amplify network needs only at pre-training scales with reduced dependence on accelerator capacity and bandwidth. On the other hand, long-sequence transformers, representative of scientific foundation models, place a more uniform dependence on network and capacity with necessary 4D parallelism. Our analysis emphasizes the need for closer performance modeling of different transformer types keeping system features in mind and demonstrates a path towards this. Our code is available as open-source.

Training a deep neural network (DNN) requires substantial computational and memory requirements. It is common to use multiple devices to train a DNN to reduce the overall training time. There are several choices to parallelize each layer in a DNN. Exhaustively searching this list to find an optimal parallelization strategy is prohibitively time consuming and impractical. The standard practice is to use data parallelism because of its simplicity. However, data parallelism is often sub-optimal, and suffers from poor performance and high memory requirement. Expert-designed strategies have been proposed on a case-by-case basis using domain specific knowledge. These expert-designed strategies do not generalize well to DNNs other than the ones for which they were designed, and are not always necessarily the best choice. In this paper, we propose an approach to automatically find efficient parallelization strategies for DNNs from their computation graphs. We present an efficient algorithm to compute these strategies within a reasonable time in practice. We evaluate the effectiveness of our approach on various DNNs. We also compare the performance of the strategies identified by our approach against data parallelism, expert-designed strategies, and the state-of-the-art approaches. Our results show that the strategies found using our approach outperform the baseline data parallelism strategy in all the cases. In addition, our strategies achieve better performance than the expert-designed strategies and the state-of-the-art approaches.