The training of large language models (LLMs) typically needs costly GPUs, such as NVIDIA A100 or H100. They possess substantial high-bandwidth on-chip memory and rapid interconnects like NVLinks. The exorbitant expenses associated with LLM training pose not just an economic challenge but also a societal one, as it restricts the ability to train LLMs from scratch to a selected few organizations.

We present CogVideoX, a large-scale text-to-video generation model based on diffusion transformer, which can generate 10-second continuous videos that align seamlessly with text prompts, with a frame rate of 16 fps and resolution of 768 x 1360 pixels. Previous video generation models often struggled with limited motion and short durations. It is especially difficult to generate videos with coherent narratives based on text. We propose several designs to address these issues. First, we introduce a 3D Variational Autoencoder (VAE) to compress videos across spatial and temporal dimensions, enhancing both the compression rate and video fidelity. Second, to improve text-video alignment, we propose an expert transformer with expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. Third, by employing progressive training and multi-resolution frame packing, CogVideoX excels at generating coherent, long-duration videos with diverse shapes and dynamic movements. In addition, we develop an effective pipeline that includes various pre-processing strategies for text and video data. Our innovative video captioning model significantly improves generation quality and semantic alignment. Results show that CogVideoX achieves state-of-the-art performance in both automated benchmarks and human evaluation. We publish the code and model checkpoints of CogVideoX along with our VAE model and video captioning model at https://github.com/THUDM/CogVideo.

Large language models (LLMs) have demonstrated powerful capabilities, requiring huge memory with their increasing sizes and sequence lengths, thus demanding larger parallel systems. The broadly adopted pipeline parallelism introduces even heavier and unbalanced memory consumption. Recomputation is a widely employed technique to mitigate the problem but introduces extra computation overhead.

Evolving graphs consisting of slices are large and constantly changing. For example, in Alipay, the graph generates hundreds of millions of new transaction records every day. Analyzing the graph within a temporary window is time-consuming due to the heavy merging of slices. Fortunately, we have discovered that most queries exhibit consistent patterns and possess monotonic properties. As a result, transitional results can be computed within slice generation for reuse. Accordingly, we develop MergeGraph enabling window-based monotonic graph analytics with reusable transitional results for pattern-consistent queries. MergeGraph has three advantages over previous works. First, it is the first system specifically tailored for window-based monotonic graph analytics with pattern-consistent queries. Second, it effectively utilizes transitional results from different slices concurrently. Third, MergeGraph boasts a high degree of expressiveness, supporting a broad spectrum of monotonic graph queries. Experimental results demonstrate that MergeGraph delivers significant performance benefits. In evaluating four typical graph applications, MergeGraph achieves an average speedup of 11.30× compared to state-of-the-art methods.

GeaFlow is a distributed dataflow system optimized for streaming graph processing, and has been widely adopted at Ant Group, serving various scenarios ranging from risk control of financial activities to analytics on social networks and knowledge graphs. It is built on top of a base with full-fledged stateful stream processing capabilities, extended with a series of graph-aware optimizations to address the space explosion and programming complexity issues of conventional join-based approaches. We propose new state backends and streaming operators that facilitate processing on dynamic graph-structured datasets, reducing space consumed by states. We develop a hybrid domain-specific language that embeds Gremlin into SQL, supporting both table and graph abstractions over streaming data. In addition to streaming workloads, GeaFlow is also extensively used for some batch processing jobs. In the largest deployments to date, GeaFlow is able to process tens of millions of events per second and manage hundreds of terabytes of states.

Out-of-core systems rely on high-performance cache sub-systems to reduce the number of I/O operations. Although the page cache in modern operating systems enables transparent access to memory and storage devices, it suffers from efficiency and scalability issues on cache misses, forcing out-of-core systems to design and implement their own cache components, which is a non-trivial task. This study proposes TriCache, a cache mechanism that enables in-memory programs to efficiently process out-of-core datasets without requiring any code rewrite. It provides a virtual memory interface on top of the conventional block interface to simultaneously achieve user transparency and sufficient out-of-core performance. A multi-level block cache design is proposed to address the challenge of per-access address translations required by a memory interface. It can exploit spatial and temporal localities in memory or storage accesses to render storage-to-memory address translation and page-level concurrency control adequately efficient for the virtual memory interface. Our evaluation shows that in-memory systems operating on top of TriCache can outperform Linux OS page cache by more than one order of magnitude, and can deliver performance comparable to or even better than that of corresponding counterparts designed specifically for out-of-core scenarios.

Domain-Specific Accelerators (DSAs) are being rapidly developed to support high-performance domain-specific computation. Although DSAs provide massive computation capability, they often only support limited native data types. To mitigate this problem, previous works have explored software emulation for certain data types, which provides some compensation for hardware limitations. However, how to efficiently design more emulated data types and choose a high-performance one without hurting correctness or precision for a given application still remains an open problem.

Large-scale pretrained AI models have shown state-of-the-art accuracy in a series of important applications. As the size of pretrained AI models grows dramatically each year in an effort to achieve higher accuracy, training such models requires massive computing and memory capabilities, which accelerates the convergence of AI and HPC. However, there are still gaps in deploying AI applications on HPC systems, which need application and system co-design based on specific hardware features.

Evolving graphs in the real world are large-scale and constantly changing, as hundreds of thousands of updates may come every second. Monotonic algorithms such as Reachability and Shortest Path are widely used in real-time analytics to gain both static and temporal insights and can be accelerated by incremental computing. Existing streaming systems adopt the incremental computing model and achieve either low latency or high throughput, but not both. However, both high throughput and low latency are required in real scenarios such as financial fraud detection. This paper presents RisGraph, a real-time streaming system that provides low-latency analysis for each update with high throughput. RisGraph addresses the challenge with localized data access and inter-update parallelism. We propose a data structure named Indexed Adjacency Lists and use sparse arrays and Hybrid Parallel Mode to enable localized data access. To achieve inter-update parallelism, we propose a domain-specific concurrency control mechanism based on the classification of safe and unsafe updates. Experiments show that RisGraph can ingest millions of updates per second for graphs with several hundred million vertices and billions of edges, and the P999 processing time latency is within 20 milliseconds. RisGraph achieves orders-of-magnitude improvement on throughput when analyses are executed for each update without batching and performs better than existing systems with batches of up to 20 million updates.

Apache Spark is a widely deployed big data analytics framework that offers such attractive features as resiliency, load-balancing, and a rich ecosystem. However, there is still plenty of room for improvement in its performance. Although a data-parallel system in a native programming language significantly improves performance, it may require re-implementing many functionalities of Spark to become a full-featured system. It is desirable for native big data systems to just write a compute engine in native languages to ensure high efficiency, and reuse other mature features provided by Spark rather than re-implement everything. But the interaction between the JVM and the native world risks becoming a bottleneck. This paper proposes Chukonu, a native big data framework that re-uses critical big data features provided by Spark. Owing to our novel DAG-splitting approach, the potential Spark integration overhead is alleviated, and its even outperforms existing pure native big data frameworks. Chukonu splits DAG programs into run-time parts and compile-time parts: The run-time parts are delegated to Spark to offload the complexities due to feature implementations. The compile-time parts are natively compiled. We propose a series of optimization techniques to be applied to the compile-time parts, such as operator fusion, vectorization, and compaction, to significantly reduce the Spark integration overhead. The results of evaluation show that Chukonu has a speedup of up to 71.58X (geometric mean 6.09X) over Apache Spark, and up to 7.20X (geometric mean 2.30X) over pure-native frameworks on six commonly-used big data applications. By translating the physical plan produced by SparkSQL into Chukonu programs, Chukonu accelerates Spark-SQL's TPC-DS performance by 2.29X.

The specific characteristics of graph workloads make it hard to design a one-size-fits-all graph storage system. Systems that support transactional updates use data structures with poor data locality, which limits the efficiency of analytical workloads or even simple edge scans. Other systems run graph analytics workloads efficiently, but cannot properly support transactions. This paper presents LiveGraph, a graph storage system that outperforms both the best graph transactional systems and the best solutions for real-time graph analytics on fresh data. LiveGraph achieves this by ensuring that adjacency list scans, a key operation in graph workloads, are purely sequential: they never require random accesses even in presence of concurrent transactions. Such pure-sequential operations are enabled by combining a novel graph-aware data structure, the Transactional Edge Log (TEL), with a concurrency control mechanism that leverages TEL's data layout. Our evaluation shows that LiveGraph significantly outperforms state-of-the-art (graph) database solutions on both transactional and real-time analytical workloads.

We present a language and compilation framework for productively generating high-performance systolic arrays for dense tensor kernels on spatial architectures, including FPGAs and CGRAs. It decouples a functional specification from a spatial mapping, allowing programmers to quickly explore various spatial optimizations for the same function. The actual implementation of these optimizations is left to a compiler. Thus, productivity and performance are achieved at the same time. We used this framework to implement several important dense tensor kernels. We implemented dense matrix multiply for an Arria-10 FPGA and a research CGRA, achieving 88% and 92% of the performance of manually written, and highly optimized expert (ninja") implementations in just 3% of their engineering time. Three other tensor kernels, including MTTKRP, TTM and TTMc, were also implemented with high performance and low design effort, and for the first time on spatial architectures."