Shenggui Li

Yuliang Liu, Shenggui Li, Jiarui Fang et al. · berkeley

In recent years, large-scale models have demonstrated state-of-the-art performance across various domains. However, training such models requires various techniques to address the problem of limited computing power and memory on devices such as GPUs. Some commonly used techniques include pipeline parallelism, tensor parallelism, and activation checkpointing. While existing works have focused on finding efficient distributed execution plans (Zheng et al. 2022) and activation checkpoint scheduling (Herrmann et al. 2019, Beaumont et al. 2021}, there has been no method proposed to optimize these two plans jointly. Moreover, ahead-of-time compilation relies heavily on accurate memory and computing overhead estimation, which is often time-consuming and misleading. Existing training systems and machine learning pipelines either physically execute each operand or estimate memory usage with a scaled input tensor. To address these challenges, we introduce a system that can jointly optimize distributed execution and gradient checkpointing plans. Additionally, we provide an easy-to-use symbolic profiler that generates memory and computing statistics for any PyTorch model with a minimal time cost. Our approach allows users to parallelize their model training on the given hardware with minimum code change based. The source code is publicly available at Colossal-AI GitHub or https://github.com/hpcaitech/ColossalAI

Haichen Huang, Jiarui Fang, Hongxin Liu et al. · berkeley

In recent years, large language models have achieved great success due to their unprecedented size. However, training these models poses a challenge for most researchers as it requires a substantial number of GPUs. To reduce GPU memory usage, memory partitioning, and memory offloading have been proposed. These approaches eliminate memory redundancies and offload memory usage to the CPU and NVMe memory, respectively, enabling training on small GPU clusters. However, directly deploying these solutions often leads to suboptimal efficiency. Only experienced experts can unleash the full potential of hardware by carefully tuning the distributed configuration. Thus, we present a novel solution, Elixir, which automates efficient large-model training based on pre-runtime model profiling. Elixir aims to identify the optimal combination of partitioning and offloading techniques to maximize training throughput. In our experiments, Elixir significantly outperforms the current state-of-the-art baseline. Our optimal configuration achieves up to a 3.4$\times$ speedup on GPT-2 models compared with SOTA solutions. We hope that our work will benefit individuals who lack computing resources and expertise, granting them access to large models. The beta version of Elixir is now available at https://github.com/hpcaitech/ColossalAI/tree/feature/elixir.

Jiangsu Du, Ziming Liu, Jiarui Fang et al. · berkeley

Large transformer models display promising performance on a wide range of natural language processing (NLP) tasks. Although the AI community has expanded the model scale to the trillion parameter level, the practical deployment of 10-100 billion parameter models is still uncertain due to the latency, throughput, and memory constraints. In this paper, we proposed EnergonAI to solve the challenges of the efficient deployment of 10-100 billion parameter transformer models on single- or multi-GPU systems. EnergonAI adopts a hierarchy-controller system architecture to coordinate multiple devices and efficiently support different parallel patterns. It delegates the execution of sub-models to multiple workers in the single-controller style and applies tensor parallelism and pipeline parallelism among the workers in a multi-controller style. Upon the novel architecture, we propose three techniques, i.e. non-blocking pipeline parallelism, distributed redundant computation elimination, and peer memory pooling. EnergonAI enables the users to program complex parallel code the same as a serial one. Compared with the FasterTransformer, we have proven that EnergonAI has superior performance on latency and throughput. In our experiments, EnergonAI can achieve 37% latency reduction in tensor parallelism, 10% scalability improvement in pipeline parallelism, and it improves the model scale inferred on a single GPU by using a larger heterogeneous memory space at cost of limited performance reduction.

Jiarui Fang, Geng Zhang, Jiatong Han et al. · berkeley

Deep learning recommendation models (DLRMs) have been widely applied in Internet companies. The embedding tables of DLRMs are too large to fit on GPU memory entirely. We propose a GPU-based software cache approaches to dynamically manage the embedding table in the CPU and GPU memory space by leveraging the id's frequency statistics of the target dataset. Our proposed software cache is efficient in training entire DLRMs on GPU in a synchronized update manner. It is also scaled to multiple GPUs in combination with the widely used hybrid parallel training approaches. Evaluating our prototype system shows that we can keep only 1.5% of the embedding parameters in the GPU to obtain a decent end-to-end training speed.

Shenggui Li, Chao Wang, Yikai Zhu et al.

Large language models incur high inference latency due to sequential autoregressive decoding. Speculative decoding alleviates this bottleneck by using a lightweight draft model to propose multiple tokens for batched verification. However, its adoption has been limited by the lack of high-quality draft models and scalable training infrastructure. We introduce SpecForge, an open-source, production-oriented framework for training speculative decoding models with full support for EAGLE-3. SpecForge incorporates target-draft decoupling, hybrid parallelism, optimized training kernels, and integration with production-grade inference engines, enabling up to 9.9x faster EAGLE-3 training for Qwen3-235B-A22B. In addition, we release SpecBundle, a suite of production-grade EAGLE-3 draft models trained with SpecForge for mainstream open-source LLMs. Through a systematic study of speculative decoding training recipes, SpecBundle addresses the scarcity of high-quality drafts in the community, and our draft models achieve up to 4.48x end-to-end inference speedup on SGLang, establishing SpecForge as a practical foundation for real-world speculative decoding deployment.

Zangwei Zheng, Xiangyu Peng, Tianji Yang et al.

Vision and language are the two foundational senses for humans, and they build up our cognitive ability and intelligence. While significant breakthroughs have been made in AI language ability, artificial visual intelligence, especially the ability to generate and simulate the world we see, is far lagging behind. To facilitate the development and accessibility of artificial visual intelligence, we created Open-Sora, an open-source video generation model designed to produce high-fidelity video content. Open-Sora supports a wide spectrum of visual generation tasks, including text-to-image generation, text-to-video generation, and image-to-video generation. The model leverages advanced deep learning architectures and training/inference techniques to enable flexible video synthesis, which could generate video content of up to 15 seconds, up to 720p resolution, and arbitrary aspect ratios. Specifically, we introduce Spatial-Temporal Diffusion Transformer (STDiT), an efficient diffusion framework for videos that decouples spatial and temporal attention. We also introduce a highly compressive 3D autoencoder to make representations compact and further accelerate training with an ad hoc training strategy. Through this initiative, we aim to foster innovation, creativity, and inclusivity within the community of AI content creation. By embracing the open-source principle, Open-Sora democratizes full access to all the training/inference/data preparation codes as well as model weights. All resources are publicly available at: https://github.com/hpcaitech/Open-Sora.

Lizhuo Luo, Shenggui Li, Yonggang Wen et al.

Diffusion large language models (dLLMs) have emerged as a promising alternative for text generation, distinguished by their native support for parallel decoding. In practice, block inference is crucial for avoiding order misalignment in global bidirectional decoding and improving output quality. However, the widely-used fixed, predefined block (naive) schedule is agnostic to semantic difficulty, making it a suboptimal strategy for both quality and efficiency: it can force premature commitments to uncertain positions while delaying easy positions near block boundaries. In this work, we analyze the limitations of naive block scheduling and disclose the importance of dynamically adapting the schedule to semantic difficulty for reliable and efficient inference. Motivated by this, we propose Dynamic Sliding Block (DSB), a training-free block scheduling method that uses a sliding block with a dynamic size to overcome the rigidity of the naive block. To further improve efficiency, we introduce DSB Cache, a training-free KV-cache mechanism tailored to DSB. Extensive experiments across multiple models and benchmarks demonstrate that DSB, together with DSB Cache, consistently improves both generation quality and inference efficiency for dLLMs. Code is released at https://github.com/lizhuo-luo/DSB.

Cunxiao Du, Jing Jiang, Xu Yuanchen et al.

Speculative decoding is a relatively new decoding framework that leverages small and efficient draft models to reduce the latency of LLMs. In this study, we introduce GliDe and CaPE, two low-hassle modifications to vanilla speculative decoding to further improve the decoding speed of a frozen LLM. Specifically, GliDe is a modified draft model architecture that reuses the cached keys and values from the target LLM, while CaPE is a proposal expansion method that uses the draft model's confidence scores to help select additional candidate tokens for verification. Extensive experiments on different benchmarks demonstrate that our proposed GliDe draft model significantly reduces the expected decoding latency. Additional evaluation using walltime reveals that GliDe can accelerate Vicuna models up to 2.17x and further extend the improvement to 2.61x with CaPE. We will release our code, data, and the trained draft models.

Qiaoling Chen, Zijun Liu, Peng Sun et al.

Adapting large language models (LLMs) via reinforcement learning (RL) is often bottlenecked by the generation stage, which can consume over 75\% of the training time. Speculative decoding (SD) accelerates autoregressive generation in serving systems, but its behavior under RL training remains largely unexplored. We identify three critical gaps that hinder the naive integration of SD into RL systems: diminishing speedups at large batch sizes, drafter staleness under continual actor updates, and drafter-induced policy degradation. To address these gaps, we present ReSpec, a system that adapts SD to RL through three complementary mechanisms: dynamically tuning SD configurations, evolving the drafter via knowledge distillation, and weighting updates by rollout rewards. On Qwen models (3B--14B), ReSpec achieves up to 4.5x speedup while preserving reward convergence and training stability, providing a practical solution for efficient RL-based LLM adaptation.

Jie Zhu, Shenggui Li, Yang You

Federated learning is proposed by Google to safeguard data privacy through training models locally on users' devices. However, with deep learning models growing in size to achieve better results, it becomes increasingly difficult to accommodate the whole model on one single device. Thus, model parallelism is then used to divide the model weights among several devices. With this logic, the approach currently used evenly allocates weights among devices. However, in reality, a computation bottleneck may occur resulting from variant computing power of different users' devices. To address this problem, load balancing is needed to allocate the model weights based on the computational capability of the device. In this paper, we proposed Sky Computing, a load-balanced model parallelism framework to adaptively allocate the weights to devices. Sky Computing outperforms the baseline method by 55% in training time when training 160-layer BERT with 64 nodes. The source code can be found at https://github.com/hpcaitech/SkyComputing.

Jiarui Fang, Zilin Zhu, Shenggui Li et al.

The pre-trained model (PTM) is revolutionizing Artificial Intelligence (AI) technology. However, the hardware requirement of PTM training is prohibitively high, making it a game for a small proportion of people. Therefore, we proposed PatrickStar system to lower the hardware requirements of PTMs and make them accessible to everyone. PatrickStar uses the CPU-GPU heterogeneous memory space to store the model data. Different from existing works, we organize the model data in memory chunks and dynamically distribute them in the heterogeneous memory. Guided by the runtime memory statistics collected in a warm-up iteration, chunks are orchestrated efficiently in heterogeneous memory and generate lower CPU-GPU data transmission volume and higher bandwidth utilization. Symbiosis with the Zero Redundancy Optimizer, PatrickStar scales to multiple GPUs on multiple nodes. % using data parallelism. The system can train tasks on bigger models and larger batch sizes, which cannot be accomplished by existing works. Experimental results show that PatrickStar extends model scales 2.27 and 2.5 times of DeepSpeed, and consistently exhibits significantly higher execution speed. PatricStar also successfully runs the 175B GPT3 training task on a 32 GPU cluster. Our code is publicly available at https://github.com/Tencent/PatrickStar.

Qifan Xu, Shenggui Li, Chaoyu Gong et al.

Huge neural network models have shown unprecedented performance in real-world applications. However, due to memory constraints, model parallelism must be utilized to host large models that would otherwise not fit into the memory of a single device. Previous methods like Megatron partition the parameters of the entire model among multiple devices, while each device has to accommodate the redundant activations in forward and backward pass. In this work, we propose Optimus, a highly efficient and scalable 2D-partition paradigm of model parallelism that would facilitate the training of infinitely large language models. In Optimus, activations are partitioned and distributed among devices, further reducing redundancy. In terms of isoefficiency, Optimus significantly outperforms Megatron. On 64 GPUs of TACC Frontera, Optimus achieves 1.48X speedup for training, 1.78X speedup for inference, and 8X increase in maximum batch size over Megatron. Optimus surpasses Megatron in scaling efficiency by a great margin. The code is available at https://github.com/xuqifan897/Optimus.

Runyu Lu, Shiqi He, Wenxuan Tan et al.

Diffusion Transformer (DiT) models excel at generating highquality images through iterative denoising steps, but serving them under strict Service Level Objectives (SLOs) is challenging due to their high computational cost, particularly at large resolutions. Existing serving systems use fixed degree sequence parallelism, which is inefficient for heterogeneous workloads with mixed resolutions and deadlines, leading to poor GPU utilization and low SLO attainment. In this paper, we propose step-level sequence parallelism to dynamically adjust the parallel degree of individual requests according to their deadlines. We present TetriServe, a DiT serving system that implements this strategy for highly efficient image generation. Specifically, TetriServe introduces a novel round-based scheduling mechanism that improves SLO attainment: (1) discretizing time into fixed rounds to make deadline-aware scheduling tractable, (2) adapting parallelism at the step level and minimize GPU hour consumption, and (3) jointly packing requests to minimize late completions. Extensive evaluation on state-of-the-art DiT models shows that TetriServe achieves up to 32% higher SLO attainment compared to existing solutions without degrading image quality.

Shenggui Li, Hongxin Liu, Zhengda Bian et al.

The success of Transformer models has pushed the deep learning model scale to billions of parameters. Due to the limited memory resource of a single GPU, However, the best practice for choosing the optimal parallel strategy is still lacking, since it requires domain expertise in both deep learning and parallel computing. The Colossal-AI system addressed the above challenge by introducing a unified interface to scale your sequential code of model training to distributed environments. It supports parallel training methods such as data, pipeline, tensor, and sequence parallelism, as well as heterogeneous training methods integrated with zero redundancy optimizer. Compared to the baseline system, Colossal-AI can achieve up to 2.76 times training speedup on large-scale models.

Zhengda Bian, Shenggui Li, Wei Wang et al.

Efficient GPU resource scheduling is essential to maximize resource utilization and save training costs for the increasing amount of deep learning workloads in shared GPU clusters. Existing GPU schedulers largely rely on static policies to leverage the performance characteristics of deep learning jobs. However, they can hardly reach optimal efficiency due to the lack of elasticity. To address the problem, we propose ONES, an ONline Evolutionary Scheduler for elastic batch size orchestration. ONES automatically manages the elasticity of each job based on the training batch size, so as to maximize GPU utilization and improve scheduling efficiency. It determines the batch size for each job through an online evolutionary search that can continuously optimize the scheduling decisions. We evaluate the effectiveness of ONES with 64 GPUs on TACC's Longhorn supercomputers. The results show that ONES can outperform the prior deep learning schedulers with a significantly shorter average job completion time.

Shenggui Li, Fuzhao Xue, Chaitanya Baranwal et al.

Transformer achieves promising results on various tasks. However, self-attention suffers from quadratic memory requirements with respect to the sequence length. Existing work focuses on reducing time and space complexity from an algorithm perspective. In this work, we propose sequence parallelism, a memory-efficient parallelism method to help us break input sequence length limitation and train with longer sequences on GPUs efficiently. Our approach is compatible with most existing parallelisms (e.g. data parallelism, pipeline parallelism and tensor parallelism), which means our sequence parallelism makes 4D parallelism possible. More importantly, we no longer require a single device to hold the whole sequence. That is, with sparse attention, our sequence parallelism enables us to train transformer with infinite long sequence. Specifically, we split the input sequence into multiple chunks and feed each chunk into its corresponding device (i.e. GPU). To compute the attention output, we integrated ring-style communication with self-attention calculation and proposed Ring Self-Attention (RSA). Experiments show that sequence parallelism performs well when scaling with batch size and sequence length. Compared with tensor parallelism, our approach achieved $13.7\times$ and $3.0\times$ maximum batch size and sequence length respectively when scaling up to 64 NVIDIA P100 GPUs. With sparse attention, sequence can handle sequence with over 114K tokens, which is over $27\times$ longer than existing sparse attention works holding the whole sequence on a single device.