Efficient video generation models are increasingly vital for multimedia synthetic content generation. Leveraging the Transformer architecture and the diffusion process, video DiT models have emerged as a dominant approach for high-quality video generation. However, their multi-step iterative denoising process incurs high computational cost and inference latency, which limits their practical deployment in large-scale and interactive multimedia applications. Caching, a widely adopted optimization method in DiT models, leverages the redundancy in the diffusion process to skip computations in different granularities (e.g., step, cfg, block). Nevertheless, existing caching methods are limited to single-granularity strategies, struggling to balance generation quality and inference speed in a flexible manner. In this work, we propose MixCache, a training-free caching-based framework for efficient video DiT inference. MixCache first distinguishes the interference and boundary between different caching strategies, and then introduces a context-aware cache triggering strategy to determine when caching should be enabled, along with an adaptive hybrid cache decision strategy for dynamically selecting the optimal caching granularity. Extensive experiments on diverse models demonstrate that MixCache can significantly accelerate video generation (e.g., 1.94× speedup on Wan 14B, 1.97× speedup on HunyuanVideo) while delivering both superior generation quality and inference efficiency compared to baseline methods.

As transformer sequence lengths grow, existing pipeline parallelisms incur suboptimal performance due to the quadratic attention computation and the substantial memory overhead. To relieve these challenges, we propose HelixPipe, a novel pipeline parallelism for long sequence transformer training. First, HelixPipe introduces attention parallel partition, which schedules attention computations of different micro batches across different pipeline stages in parallel, reducing pipeline bubbles. Second, it employs a two-fold first-in-last-out micro batch schedule to balance memory usage and overlap communication with computation. Additionally, HelixPipe utilizes recomputation without attention and chunked MLP to mitigate fragmentation and enable longer sequences. Experiments demonstrate that HelixPipe gains increasing advantages with longer sequence lengths, and outperforms existing methods in throughput and scalability across varying pipeline sizes, model sizes, and cluster configurations. Notably, it achieves a 26% speedup over baseline methods when training a 7B model with 128k sequence length on 64 H20 GPUs. Code is available at https://github.com/zxgx/Megatron-LM.

With the exponential growth of deep learning (DL), there arises an escalating need for scalability. Despite significant advancements in communication hardware capabilities, the time consumed by communication remains a bottleneck during training. The existing various optimizations are coupled within parallel systems to implement specific computation-communication overlap. These approaches pose challenges in terms of performance, programmability, and generality. In this paper, we introduce Concerto, a compiler framework designed to address these challenges by automatically optimizing and scheduling communication. We formulate the scheduling problem as a resource-constrained project scheduling problem and use off-the-shelf solver to get the near-optimal scheduling. And use auto-decomposition to create overlap opportunity for critical (synchronous) communication. Our evaluation shows Concerto can match or outperform state-of-the-art parallel frameworks, including Megatron-LM, JAX/XLA, DeepSpeed, and Alpa, all of which include extensive hand-crafted optimization. Unlike previous works, Concerto decouples the parallel approach and communication optimization, then can generalize to a wide variety of parallelisms without manual optimization.

Despite Low-Rank Adaptation (LoRA)’s popularity for fine-tuning large models, it often exhibits a noticeable performance gap compared to full fine-tuning, particularly in complex tasks such as mathematical reasoning and code generation. Motivated by this discrepancy, we propose a novel fusion approach for LoRA fine-tuned models. Our key insight is that LoRA models trained with different random seeds on the same task often exhibit complementary strengths. In contrast to existing research that typically focuses on fusing models trained on diverse tasks, we explore the potential of combining multiple LoRA models fine-tuned on the same task with different random seeds. This intra-task fusion method aims to leverage the strengths of various fine-tuned models to create a more robust and effective adaptation. To validate our approach, we conducted comprehensive experiments across three key areas: mathematical reasoning, code generation, and general instruction-tuning tasks. The results demonstrate that our fusion method significantly enhances LoRA’s performance, outperforming both standalone LoRA models and current fusion methods. Notably, this advancement substantially narrows the gap between LoRA and full fine-tuning, thus offering a more effective approach to model adaptation without the GPU memory burden of full parameter fine-tuning.

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.

Multimodal large language models (MLLMs) extend LLMs to handle images, videos, and audio by incorporating feature extractors and projection modules. However, these additional components—combined with complex inference pipelines and heterogeneous workloads—introduce significant inference overhead. Therefore, efficiently serving MLLMs remains a major challenge. Current tightly coupled serving architectures struggle to distinguish between mixed request types or adapt parallelism strategies to different inference stages, leading to increased time-to-first-token (TTFT) and poor resource utilization. To address this, we introduce Elastic Multimodal Parallelism (EMP), a new serving paradigm that elastically adapts to resource heterogeneity across request types and inference stages. Building upon EMP, we develop ElasticMM, an MLLM serving system that (1) separates requests into independent modality groups with dynamic resource allocation via a modality-aware load balancer; (2) decouples inference stages and enables parallelism adjustment and adaptive scaling via elastic partition scheduling; and (3) improves inference efficiency through unified multimodal prefix caching and non-blocking encoding. Experiments on diverse real-world datasets show that ElasticMM outperforms state-of-the-art (SOTA) serving systems, reducing TTFT by up to 4.2$\times$ and achieving 3.2–4.5$\times$ higher throughput while meeting service-level objectives (SLOs).

Training Transformer models on long sequences in a distributed setting poses significant challenges in terms of efficiency and scalability. Current methods are either constrained by the number of attention heads or excessive communication overheads. To address this problem, we propose StarTrail, a multi-dimensional concentric distributed training system for long sequences, fostering an efficient communication paradigm and providing additional tuning flexibility for communication arrangements. Specifically, StarTrail introduces an extra parallel dimension and divides the peer-to-peer communication into sub-rings to substantially reduce communication volume and avoid bandwidth bottlenecks. Through comprehensive experiments across diverse hardware environments and on both Natural Language Processing (NLP) and Computer Vision (CV) tasks, we demonstrate that our approach significantly surpasses state-of-the-art methods that support Long sequence lengths, achieving performance improvements of up to 77.12% on GPT-style models and up to 114.33% on DiT (Diffusion Transformer) models without affecting the computations results.

Large deep learning models have achieved impressive performance across a range of applications. However, their large memory requirements, including parameter memory and activation memory, have become a significant challenge for their practical serving. While existing methods mainly address parameter memory, the importance of activation memory has been overlooked. Especially for long input sequences, activation memory is expected to experience a significant exponential growth as the length of sequences increases. In this approach, we propose AutoChunk, an automatic and adaptive compiler system that efficiently reduces activation memory for long sequence inference by chunk strategies. The proposed system generates chunk plans by optimizing through multiple stages. In each stage, the chunk search pass explores all possible chunk candidates and the chunk selection pass identifies the optimal one. At runtime, AutoChunk employs code generation to automatically apply chunk strategies. The experiments demonstrate that AutoChunk can reduce over 80% of activation memory while maintaining speed loss within 10%, extend max sequence length by 3.2x to 11.7x, and outperform state-of-the-art methods by a large margin.

Protein structure prediction helps to understand gene translation and protein function, which is of growing interest and importance in structural biology. The AlphaFold model, which used transformer architecture to achieve atomic-level accuracy in protein structure prediction, was a significant breakthrough. However, training and inference of AlphaFold model are challenging due to its high computation and memory cost. In this work, we present FastFold, an efficient implementation of AlphaFold for both training and inference. We propose Dynamic Axial Parallelism (DAP) as a novel model parallelism method. Additionally, we have implemented a series of low-level optimizations aimed at reducing communication, computation, and memory costs. These optimizations include Duality Async Operations, highly optimized kernels, and AutoChunk (an automated search algorithm finds the best chunk strategy to reduce memory peaks). Experimental results show that FastFold can efficiently scale to more GPUs using DAP and reduces overall training time from 11 days to 67 hours and achieves 7.5 ~ 9.5× speedup for long-sequence inference. Furthermore, AutoChunk can reduce memory cost by over 80% during inference by automatically partitioning the intermediate tensors during the computation.

Distributed large model inference is still in a dilemma where balancing cost and effect. The online scenarios demand intraoperator parallelism to achieve low latency and intensive communications makes it costly. Conversely, the inter-operator parallelism can achieve high throughput with much fewer communications, but it fails to enhance the effectiveness.

Large-scale language models have become increasingly challenging and expensive to train. Among various methods addressing this issue, Pipeline Parallelism has been widely employed to accommodate massive model weights within limited GPU memory. This paper introduces Hanayo, a wave-like pipeline parallelism strategy that boasts a concise structure and practical applicability, alongside a high-performance pipeline execution runtime to tackle the challenges of pipeline strategy implementation. Hanayo mitigates the issues of pipeline bubbles and excessive memory consumption prevalent in existing schemes, without resorting to model duplicates as in Chimera. Our evaluation, conducted on four distinct computing clusters and involving both GPT-like and BERT-like architectures with up to 32 GPUs, demonstrates up to a 30.4 % increase in throughput compared to the state-of-the-art approach.

Mixed-precision computing becomes an inevitable trend for HPC and AI applications due to the increasing using mixed-precision units such as NVIDIA Tensor Cores. Fast Fourier transform (FFT) is one of the most widely-used scientific kernels and hence mixed-precision FFT is highly demanded. However, few existing FFT libraries (or algorithms) can support universal size of FFTs on Tensor Cores. Therefore, we proposed tcFFT, a fast half-precision FFT library on Tensor Cores that can support universal size of 1D and 2D FFTs. Our work consists of two parts: framework design and performance optimizations. We designed the tcFFT library framework to support all power-of-two size and multi-dimension of FFTs; we applied two performance optimizations, one to use Tensor Cores efficiently and the other to ease GPU memory bottlenecks. We evaluated tcFFT with a wide range size of 1D and 2D FFTs on NVIDIA V100 and A100 GPUs. The results show that tcFFT can outperform 1.29X-3.24X and 1.10X-3.03X higher on average than NVIDIA cuFFT v11.0 in FP16 on V100 and A100, respectively.

N-body simulations are essential tools in physical cosmology to understand the large-scale structure (LSS) formation of the universe. Large-scale simulations with high resolution are important for exploring the substructure of universe and for determining fundamental physical parameters like neutrino mass. However, traditional particle-mesh (PM) based algorithms use considerable amounts of memory, which limits the scalability of simulations. Therefore, we designed a two-level PM algorithm CUBE towards optimal performance in memory consumption reduction. By using the fixed-point compression technique, CUBE reduces the memory consumption per N-body particle to only 6 bytes, an order of magnitude lower than the traditional PM-based algorithms. We scaled CUBE to 512 nodes (20,480 cores) on an Intel Cascade Lake based supercomputer with ≃95% weak-scaling efficiency. This scaling test was performed in Cosmo-π - a cosmological LSS simulation using ≃4.4 trillion particles, tracing the evolution of the universe over ≃13.7 billion years. To our best knowledge, Cosmo-π is the largest completed cosmological N-body simulation. We believe CUBE has a huge potential to scale on exascale supercomputers for larger simulations.

Dense depth cues are important and have wide applications in various computer vision tasks. In autonomous driving, LIDAR sensors are adopted to acquire depth measurements around the vehicle to perceive the surrounding environments. However, depth maps obtained by LIDAR are generally sparse because of its hardware limitation. The task of depth completion attracts increasing attention, which aims at generating a dense depth map from an input sparse depth map. To effectively utilize multi-scale features, we propose three novel sparsity-invariant operations, based on which, a sparsity-invariant multi-scale encoder-decoder network (HMS-Net) for handling sparse inputs and sparse feature maps is also proposed. Additional RGB features could be incorporated to further improve the depth completion performance. Our extensive experiments and component analysis on two public benchmarks, KITTI depth completion benchmark and NYU-depth-v2 dataset, demonstrate the effectiveness of the proposed approach. As of Aug. 12th, 2018, on KITTI depth completion leaderboard, our proposed model without RGB guidance ranks 1st among all peer-reviewed methods without using RGB information, and our model with RGB guidance ranks 2nd among all RGB-guided methods.