Linear attention mechanisms deliver significant advantages for Large Language Models (LLMs) by providing linear computational complexity, enabling efficient processing of ultra-long sequences (e.g., 1M context). However, existing Sequence Parallelism (SP) methods, essential for distributing these workloads across devices, become the primary performance bottleneck due to substantial communication overhead. In this paper, we introduce ZeCO (Zero Communication Overhead) sequence parallelism for linear attention models, a new SP method designed to overcome these limitations and achieve practically end-to-end near-linear scalability for long sequence training. For example, training a model with a 1M sequence length across 64 devices using ZeCO takes roughly the same time as training with an 16k sequence on a single device. At the heart of ZeCO lies All-Scan, a novel collective communication primitive.

Optimizing large-language model (LLM) training on distributed domain-specific accelerator systems presents significant challenges due to its complex optimization space. Existing optimization methods, however, rely on time-consuming manual tuning or resource-intensive black-box searches, which struggle to keep pace with the rapidly evolving LLM domain, leading to slow development and underutilized resources. To address this, we introduce ASAP, an Agentic Solution to Auto-optimize Performance of Large-Scale LLM Training. It is a multi-agent system, featuring Coordinator, Analyzer, and Proposal agents, which integrates LLM reasoning with insights from performance profiling tools, roofline analysis, and a knowledge base of best practices and successful past optimizations from human experts. Our proposed design can automate the diagnosis of performance bottlenecks and recommend optimized sharding configurations with reasoning, thus effectively improving the efficiency of distributed LLM training. Experiments have shown that the ASAP-generated sharding configurations can contribute up to 28% training step time reduction and 1.43 times throughput improvement. When combined with additional optimization from human experts, throughput can be further increased to 2.58 times. The proposed ASAP promises to provide a scalable and explainable methodology for AI-assisted performance engineering in large-scale LLM training.

Large Language Models (LLMs) nowadays are expected to process inputs of unprecedented length, with current token limits extending to several million tokens.

Recent advances in large language models (LLMs) have driven impressive progress in omni-modal understanding and generation. However, training omni-modal LLMs remains a significant challenge due to the heterogeneous model architectures required to process diverse modalities, necessitating sophisticated system design for efficient large-scale training. Existing frameworks typically entangle model definition with parallel logic, incurring limited scalability and substantial engineering overhead for end-to-end omni-modal training. We present VeOmni, a modular and efficient training framework to accelerate the development of omni-modal LLMs. VeOmni introduces model-centric distributed recipes that decouples communication from computation, enabling efficient 3D parallelism on omni-modal LLMs. VeOmni also features a flexible configuration interface supporting seamless integration of new modalities with minimal code change. Using VeOmni, a omni-modal mixture-of-experts (MoE) model with 30B parameters can be trained with over 2,800 tokens/sec/GPU throughput and scale to 160K context lengths via 3D parallelism on 128 GPUs, showcasing its superior efficiency and scalability for training large omni-modal LLMs.

Linear attention mechanisms deliver significant advantages for Large Language Models (LLMs) by providing linear computational complexity, enabling efficient processing of ultra-long sequences (e.g., 1M context). However, existing Sequence Parallelism (SP) methods, essential for distributing these workloads across devices, become the primary bottleneck due to substantial communication overhead. In this paper, we introduce ZeCO (Zero Communication Overhead) sequence parallelism for linear attention models, a new SP method designed to overcome these limitations and achieve end-to-end near-linear scalability for long sequence training. For example, training a model with a 1M sequence length across 64 devices using ZeCO takes roughly the same time as training with an 16k sequence on a single device. At the heart of ZeCO lies All-Scan, a new collective communication primitive. All-Scan provides each SP rank with precisely the initial operator state it requires while maintaining a minimal communication footprint, effectively eliminating communication overhead. Theoretically, we prove the optimaity of ZeCO, showing that it introduces only negligible time and space overhead. Empirically, we compare the communication costs of different sequence parallelism strategies and demonstrate that All-Scan achieves the fastest communication in SP scenarios. Specifically, on 256 GPUs with an 8M sequence length, ZeCO achieves a 60\% speedup compared to the current state-of-the-art (SOTA) SP method. We believe ZeCO establishes a clear path toward efficiently training next-generation LLMs on previously intractable sequence lengths.

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.

While long-context inference is crucial for advancing large language model (LLM) applications, its prefill speed remains a significant bottleneck. Current approaches, including sequence parallelism strategies and compute reduction through approximate attention mechanisms, still fall short of delivering optimal inference efficiency. This hinders scaling the inputs to longer sequences and processing long-context queries in a timely manner. To address this, we introduce APB, an efficient long-context inference framework that leverages multi-host approximate attention to enhance prefill speed by reducing compute and enhancing parallelism simultaneously. APB introduces a communication mechanism for essential key-value pairs within a sequence parallelism framework, enabling a faster inference speed while maintaining task performance. We implement APB by incorporating a tailored FlashAttn kernel alongside optimized distribution strategies, supporting diverse models and parallelism configurations. APB achieves speedups of up to 9.2x, 4.2x, and 1.6x compared with FlashAttn, RingAttn, and StarAttn, respectively, without any observable task performance degradation. We provide the implementation and experiment code of APB in https://github.com/thunlp/APB.

Sequence parallelism (SP), which divides the sequence dimension of input tensors across multiple computational devices, is becoming key to unlocking the long-context capabilities of generative AI models. This paper investigates the state-of-the-art SP approaches, i.e. DeepSpeed-Ulysses and Ring-Attention, and proposes a unified SP approach, which is more robust to transformer model architectures and network hardware topology. This paper compares the communication and memory cost of SP and existing parallelism, including data/tensor/zero/pipeline parallelism, and discusses the best practices for designing hybrid 4D parallelism involving SP. We achieved 47% MFU on two 8xA800 nodes using SP for the LLAMA3-8B model training using sequence length 208K. Our code is publicly available at https://github.com/feifeibear/long-context-attention.

Scaling multi-dimensional transformers to long sequences is indispensable across various domains. However, the challenges of large memory requirements and slow speeds of such sequences necessitate sequence parallelism. All existing approaches fall under the category of embedded sequence parallelism, which are limited to shard along a single sequence dimension, thereby introducing significant communication overhead. However, the nature of multi-dimensional transformers involves independent calculations across multiple sequence dimensions. To this end, we propose Dynamic Sequence Parallelism (DSP) as a novel abstraction of sequence parallelism. DSP dynamically switches the parallel dimension among all sequences according to the computation stage with efficient resharding strategy. DSP offers significant reductions in communication costs, adaptability across modules, and ease of implementation with minimal constraints. Experimental evaluations demonstrate DSP's superiority over state-of-the-art embedded sequence parallelism methods by remarkable throughput improvements ranging from 32.2% to 10x, with less than 25% communication volume.

This paper introduces PipeFusion, a novel approach that harnesses multi-GPU parallelism to address the high computational and latency challenges of generating high-resolution images with diffusion transformers (DiT) models. PipeFusion splits images into patches and distributes the network layers across multiple devices. It employs a pipeline parallel manner to orchestrate communication and computations. By leveraging the high similarity between the input from adjacent diffusion steps, PipeFusion eliminates the waiting time in the pipeline by reusing the one-step stale feature maps to provide context for the current step. Our experiments demonstrate that it can generate higher image resolution where existing DiT parallel approaches meet OOM. PipeFusion significantly reduces the required communication bandwidth, enabling DiT inference to be hosted on GPUs connected via PCIe rather than the more costly NVLink infrastructure, which substantially lowers the overall operational expenses for serving DiT models. Our code is publicly available at https://github.com/PipeFusion/PipeFusion.