Asδ is relatively slow compared with other time consumption, we treat it asaconstant that does not scale with variable size. The second process introduces network overhead (e.g.,latency) and network communication. For vi VCC, we model 5 mostly used collective primitives:AllReduce, ReduceScatter, AllGather, Broadcast and Reduce [12]. I1,I2,I3 are true whenAllReduce, ReduceScatter and AllGather, Broadcast and Reduce are activated, respectively. Note thattheLSTM walks through eachvi V0G,θ strictly following their original forward (backward) order in the computational graph, so as to inject this information intothemodeling. A global load balancer (clb) and group assigner (cam) assign their values using randomized and approximate solutions, illustrated in Algorithm 1 and Algorithm 2, respectively.

Machine Learning jobs, carried out on large number of distributed high performance systems, involve periodic communication using operations like AllReduce, AllGather, and Broadcast. These operations may create high bandwidth and bursty traffic patterns, leading to network congestion and packet loss, thus impacting the performance of these jobs. Hence it is imperative to analyze these patterns, which can be helpful in provisioning network resources depending on the type of machine learning workloads. In this poster we carry out extensive analysis of the collective communication behavior seen in a wide variety of models (ex. DeepSeek, GPT, Llama, etc.) To achieve this we instrument Nvidia Collective Communication Library logging functionality for richer context about the collectives and workloads. We adjust configuration parameters that influence collective communication behavior, such as parallelism, number of nodes, and model type. This overview presents and discusses some of the results on the collective communication behavior for the open source DeepSeek V3 inferencing model, which includes operation type and count, transfer sizes per operation, and request size distribution. Our analysis shows that it makes sense to rethink current collective communication frameworks and network topologies so as to accommodate the effect of network anomalies on the mentioned workloads.

Linear sequence modeling approaches, such as linear attention, provide advantages like linear-time training and constant-memory inference over sequence lengths. However, existing sequence parallelism (SP) methods are either not optimized for the right-product-first feature of linear attention or use a ring-style communication strategy, which results in lower computation parallelism, limits their scalability for longer sequences in distributed systems. In this paper, we introduce LASP-2, a new SP method to enhance both communication and computation parallelism when training linear attention transformer models with very-long input sequences. Compared to previous work LASP, LASP-2 rethinks the minimal communication requirement for SP on linear attention layers, reorganizes the whole communication-computation workflow of LASP. In this way, only one single AllGather collective communication is needed on intermediate memory states, whose sizes are independent of the sequence length, leading to significant improvements of both communication and computation parallelism, as well as their overlap. Additionally, we extend LASP-2 to LASP-2H by applying similar communication redesign to standard attention modules, offering an efficient SP solution for hybrid models that blend linear and standard attention layers. Our evaluation on a Linear-Llama3 model, a variant of Llama3 with linear attention replacing standard attention, demonstrates the effectiveness of LASP-2 and LASP-2H. Specifically, LASP-2 achieves training speed improvements of 15.2% over LASP and 36.6% over Ring Attention, with a sequence length of 2048K across 64 GPUs. The Code is released as a part of: https://github.com/OpenSparseLLMs/Linear-MoE.

The Mixture of Experts (MoE) is an advanced model architecture in the industry that combines multiple specialized expert models from various domains into a single supermodel. This approach enables the model to scale without significantly increasing the computational costs of training and inference, while maximizing model performance. However, current distributed training frameworks do not consider the ultimate optimization of communication, especially for large base models. This paper proposes a network-traffic-aware parallel optimization method that selects the optimal parallel strategy based on the communication volume, and the training cluster's inter-node and intra-node network topologies. Compared to the DeepSpeed, MoNTA achieves an 8x increase in AllToAll communication performance under 8-card tensor parallelism. Compared to the baseline, training a 2x70B model using 16 A800 cards, with an 8K sequence, results in a 13% overall latency performance improvement. Project Page: https://github.com/EnflameTechnology/DeepSpeed.

Deep learning (DL) models based on the transformer architecture have revolutionized many DL applications such as large language models (LLMs), vision transformers, audio generation, and time series prediction. Much of this progress has been fueled by distributed training, yet distributed communication remains a substantial bottleneck to training progress. This paper examines the communication behavior of transformer models - that is, how different parallelism schemes used in multi-node/multi-GPU DL Training communicate data in the context of transformers. We use GPT-based language models as a case study of the transformer architecture due to their ubiquity. We validate the empirical results obtained from our communication logs using analytical models. At a high level, our analysis reveals a need to optimize small message point-to-point communication further, correlations between sequence length, per-GPU throughput, model size, and optimizations used, and where to potentially guide further optimizations in framework and HPC middleware design and optimization.

Sparsely-activated Mixture-of-Expert (MoE) layers have found practical applications in enlarging the model size of large-scale foundation models, with only a sub-linear increase in computation demands. Despite the wide adoption of hybrid parallel paradigms like model parallelism, expert parallelism, and expert-sharding parallelism (i.e., MP+EP+ESP) to support MoE model training on GPU clusters, the training efficiency is hindered by communication costs introduced by these parallel paradigms. To address this limitation, we propose Parm, a system that accelerates MP+EP+ESP training by designing two dedicated schedules for placing communication tasks. The proposed schedules eliminate redundant computations and communications and enable overlaps between intra-node and inter-node communications, ultimately reducing the overall training time. As the two schedules are not mutually exclusive, we provide comprehensive theoretical analyses and derive an automatic and accurate solution to determine which schedule should be applied in different scenarios. Experimental results on an 8-GPU server and a 32-GPU cluster demonstrate that Parm outperforms the state-of-the-art MoE training system, DeepSpeed-MoE, achieving 1.13$\times$ to 5.77$\times$ speedup on 1296 manually configured MoE layers and approximately 3$\times$ improvement on two real-world MoE models based on BERT and GPT-2.

Training extremely large language models with billions of parameters is a computationally intensive task that pushes the limits of current data parallel training systems. While techniques like ZeRO++ have enabled efficient distributed training of such giant models on inexpensive low-bandwidth clusters, they can suffer from convergence issues due to potential race conditions in the hierarchical partitioning (hpZ) scheme employed to reduce cross-machine communication. In this work, we first show how these race conditions cause instability when training models with billions of parameters. We then propose a modification to the partitioning algorithm that addresses these convergence challenges while maintaining competitive training efficiency. Empirical evaluation on training the multi-billion parameters Falcon Models and Llama-2 models demonstrates the updated algorithm's ability to achieve reliable convergence on these massive models, where stock ZeRO++ hpZ fails to converge. The updated algorithm enables robust training of larger models with 98\% throughput and model training speed improvement without sacrificing the quality of convergence.

Large deep learning models have demonstrated strong ability to solve many tasks across a wide range of applications. Those large models typically require training and inference to be distributed. Tensor parallelism is a common technique partitioning computation of an operation or layer across devices to overcome the memory capacity limitation of a single processor, and/or to accelerate computation to meet a certain latency requirement. However, this kind of parallelism introduces additional communication that might contribute a significant portion of overall runtime. Thus limits scalability of this technique within a group of devices with high speed interconnects, such as GPUs with NVLinks in a node. This paper proposes a novel method, Flux, to significantly hide communication latencies with dependent computations for GPUs. Flux over-decomposes communication and computation operations into much finer-grained operations and further fuses them into a larger kernel to effectively hide communication without compromising kernel efficiency. Flux can potentially overlap up to 96% of communication given a fused kernel. Overall, it can achieve up to 1.24x speedups for training over Megatron-LM on a cluster of 128 GPUs with various GPU generations and interconnects, and up to 1.66x and 1.30x speedups for prefill and decoding inference over vLLM on a cluster with 8 GPUs with various GPU generations and interconnects.

As modern DNN models grow ever larger, collective communications between the accelerators (allreduce, etc.) emerge as a significant performance bottleneck. Designing efficient communication schedules is challenging given today's highly diverse and heterogeneous network fabrics. In this paper, we present ForestColl, a tool that generates efficient schedules for any network topology. ForestColl constructs broadcast/aggregation spanning trees as the communication schedule, achieving theoretically minimum network congestion. Its schedule generation runs in strongly polynomial time and is highly scalable. ForestColl supports any network fabrics, including both switching fabrics and direct connections, as well as any network graph structure. We evaluated ForestColl on multi-cluster AMD MI250 and NVIDIA A100 platforms. ForestColl's schedules achieved up to 52\% higher performance compared to the vendors' own optimized communication libraries, RCCL and NCCL. ForestColl also outperforms other state-of-the-art schedule generation techniques with both up to 61\% more efficient generated schedules and orders of magnitude faster schedule generation speed.

Modern large-scale deep learning workloads highlight the need for parallel execution across many devices in order to fit model data into hardware accelerator memories. In these settings, array redistribution may be required during a computation, but can also become a bottleneck if not done efficiently. In this paper we address the problem of redistributing multi-dimensional array data in SPMD computations, the most prevalent form of parallelism in deep learning. We present a type-directed approach to synthesizing array redistributions as sequences of MPI-style collective operations. We prove formally that our synthesized redistributions are memory-efficient and perform no excessive data transfers. Array redistribution for SPMD computations using collective operations has also been implemented in the context of the XLA SPMD partitioner, a production-grade tool for partitioning programs across accelerator systems. We evaluate our approach against the XLA implementation and find that our approach delivers a geometric mean speedup of $1.22\times$, with maximum speedups as a high as $5.7\times$, while offering provable memory guarantees, making our system particularly appealing for large-scale models.