Mixture-of-Experts (MoE) architecture with experts parallelism scales LLMs efficiently by activating only a subset of experts per input, avoiding proportional training costs. However, the intensive and heterogeneous communication substantially hinders the efficiency and scalability of MoE training in the resource-constrained scenario. Existing communication compression techniques fall short in MoE training due to: (i) Intensive training amplifies compression overhead, compromising training efficiency; (ii) Accumulated compression errors propagate through the network, degrading training quality. In this paper, we propose RCMoE, a communication-efficient Random Compression framework for MoE training with two core modules: (1) Local-Stochastic Quantization compresses the all-to-all communication by stochastically quantizing each row of the expert's intermediate computing results in parallel, effectively improving the compression efficiency and reducing compression error; (2) Probabilistic Thresholding Sparsification compresses the all-reduce communication by probabilistically sampling large gradients at high probability, thereby reducing the computational complexity and maintaining the convergence efficiency. Experiments on four typical MoE training tasks prove that RCMoE achieves higher 5.9x-8.1x total communication compression ratios and 1.3x-10.1x training speedup compared with the state-of-the-art compression techniques while maintaining the MoE training accuracy.

Mixture-of-Experts (MoE) model parallelism is prevalent in training Large Language Models (e.g., ChatGPT). However, the intensive all-to-all collective communication of the MoE layer’s intermediate computing results substantially degrades MoE training efficiency. In this paper, we propose BirdMoE, a novel load-aware communication compression technique with Bi-random quantization for MoE training with two core modules. Specifically, BirdMoE employs a lightweight Random Quantization (RQ) with expectation invariance property to efficiently map the floating-point intermediate computing results into integers while maintaining the MoE training quality. Additionally, BirdMoE utilizes a Mixed Precision (MP) strategy to dynamically balance the communication loads among expert nodes, significantly improving all-to-all communication efficiency for the MoE training system. Experiments on four typical MoE training tasks demonstrate that BirdMoE achieves higher $4.06 \times- 10.44 \times$ total communication compression ratios and $1.18 \times-5.27 \times$ training speedup compared with the state-of-the-art compression techniques while maintaining the MoE training quality.

Distributed training often suffers from high communication overhead due to large-scale gradient synchronization. Although gradient compression—particularly at 4-bit or even lower precision—significantly reduces transfer volume, it typically results in sacrifice in precision and degradation of the final model accuracy. In this work, we introduce DUO, a distributed training framework designed to mitigate accuracy degradation incurred by gradient compression without involving additional overhead. DUO achieves this by inserting an additional high-precision gradient synchronization step into a previously computation-only phase, so that its communication is fully hidden by computation. We provide a comprehensive theoretical proof of convergence for DUO and validate its effectiveness through extensive pre-training experiments on GPT models. Our results indicate that DUO effectively restores accuracy when using 4-bit gradient compression, achieving performance comparable to uncompressed training. Remarkably, DUO maintains minimal accuracy degradation even under extreme compression scenarios, including 1-bit gradients or complete omission of the low-precision gradient communication step (0-bit transmission).

Second-order optimization methods have been developed to enhance convergence and generalization in deep neural network (DNN) training compared to first-order methods like Stochastic Gradient Descent (SGD). However, these methods face challenges in distributed settings due to high communication overhead. Gradient compression, a technique commonly used to accelerate communication for first-order approaches, often results in low communication reduction ratios, decreased model accuracy, and/or high compression overhead when applied to second-order methods. To address these limitations, we introduce a novel gradient compression method for second-order optimizers called COMPSO. This method effectively reduces communication costs while preserving the advantages of second-order optimization. COMPSO employs stochastic rounding to maintain accuracy and filters out minor gradients to improve compression ratios. Additionally, we develop GPU optimizations to minimize compression overhead and performance modeling to ensure end-to-end performance gains across various systems. Evaluation of COMPSO on different DNN models shows that it achieves a compression ratio of 22.1×, reduces communication time by 14.2×, and improves overall performance by 1.9×, all without any drop in model accuracy.

Error-bounded lossy compression is one of the most efficient solutions to reduce the volume of scientific data. For lossy compression, progressive decompression and random-access decompression are critical features that enable on-demand data access and flexible analysis workflows. However, these features can severely degrade compression quality and speed. To address these limitations, we propose a novel streaming compression framework that supports both progressive decompression and random-access decompression while maintaining high compression quality and speed. Our contributions are three-fold: (1) we design the first compression framework that simultaneously enables both progressive decompression and random-access decompression; (2) we introduce a hierarchical partitioning strategy to enable both streaming features, along with a hierarchical prediction mechanism that mitigates the impact of partitioning and achieves high compression quality—even comparable to state-of-the-art (SOTA) non-streaming compressor SZ3; and (3) our framework delivers high compression and decompression speed, up to 6.7 × faster than SZ3.

Recent years have witnessed a clear trend towards language models with an ever-increasing number of parameters, as well as the growing training overhead and memory usage. Distributed training, particularly through Sharded Data Parallelism (ShardedDP) which partitions optimizer states among workers, has emerged as a crucial technique to mitigate training time and memory usage. Yet, a major challenge in the scalability of ShardedDP is the intensive communication of weights and gradients. While compression techniques can alleviate this issue, they often result in worse accuracy. Driven by this limitation, we propose SDP4Bit (Toward 4Bit Communication Quantization in Sharded Data Parallelism for LLM Training), which effectively reduces the communication of weights and gradients to nearly 4 bits via two novel techniques: quantization on weight differences, and two-level gradient smooth quantization. Furthermore, SDP4Bit presents an algorithm-system co-design with runtime optimization to minimize the computation overhead of compression. Additional to the theoretical guarantees of convergence, we empirically evaluate the accuracy of SDP4Bit on the pre-training of GPT models with up to 6.7 billion parameters, and the results demonstrate a negligible impact on training loss. Furthermore, speed experiments show that SDP4Bit achieves up to 4.08× speedup in end-to-end throughput on a scale of 128 GPUs.