The development of large language models (LLMs) has been instrumental in advancing state-of-the-art natural language processing applications. Training LLMs with billions of parameters and trillions of tokens requires sophisticated distributed systems that enable composing and comparing several state-of-the-art techniques in order to efficiently scale across thousands of accelerators. However, existing solutions are complex, scattered across multiple libraries/repositories, lack interoperability, and are cumbersome to maintain. Thus, curating and empirically comparing training recipes requires non-trivial engineering effort. This paper introduces **TORCHTITAN**$^1$, a PyTorch-native distributed training system that unifies and advances state-of-the-art techniques, streamlining integration and reducing engineering overhead. TORCHTITAN enables seamless application of 4D parallelism in a modular and composable manner, while featuring elastic scaling to adapt to changing computational requirements. The system provides comprehensive logging, efficient checkpointing, and debugging tools, ensuring production-ready training. Moreover, TORCHTITAN incorporates innovative hardware-software co-designed solutions, leveraging cutting-edge features like Float8 training and SymmetricMemory to maximize hardware utilization. As a flexible experimental test bed, TORCHTITAN facilitates the curation and comparison of custom recipes for diverse training contexts. By leveraging TORCHTITAN, we developed optimized training recipes for the Llama 3.1 family and provide actionable guidance on selecting and combining distributed training techniques to maximize training efficiency, based on our hands-on experiences. We thoroughly assess TORCHTITAN on the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its exceptional performance, modular composability, and elastic scalability. By stacking training optimizations, we demonstrate accelerations ranging from 65.08% on Llama 3.1 8B at 128 GPU scale (1D), 12.59% on Llama 3.1 70B at 256 GPU scale (2D), to 30% on Llama 3.1 405B at 512 GPU scale (3D) on NVIDIA H100 GPUs over optimized baselines. We also demonstrate the effectiveness of 4D parallelism in enabling long context training. $^1$ GitHub: [https://github.com/pytorch/torchtitan](https://github.com/pytorch/torchtitan)

It is widely acknowledged that large models have the potential to deliver superior performance across a broad range of domains. Despite the remarkable progress made in the field of machine learning systems research, which has enabled the development and exploration of large models, such abilities remain confined to a small group of advanced users and industry leaders, resulting in an implicit technical barrier for the wider community to access and leverage these technologies. In this paper, we introduce PyTorch Fully Sharded Data Parallel (FSDP) as an industry-grade solution for large model training. FSDP has been closely co-designed with several key PyTorch core components including Tensor implementation, dispatcher system, and CUDA memory caching allocator, to provide non-intrusive user experiences and high training efficiency. Additionally, FSDP natively incorporates a range of techniques and settings to optimize resource utilization across a variety of hardware configurations. The experimental results demonstrate that FSDP is capable of achieving comparable performance to Distributed Data Parallel while providing support for significantly larger models with near-linear scalability in terms of TFLOPS.

It is known that deeper and wider neural networks can achieve better accuracy. But it is difficult to continue the trend to increase model size due to limited GPU memory. One promising solution is to support swapping between GPU and CPU memory. However, existing work on swapping only handle certain models and do not achieve satisfactory performance. Deep learning computation is commonly expressed as a dataflow graph which can be analyzed to improve swapping. We propose SwapAdvisor, which performs joint optimization along 3 dimensions based on a given dataflow graph: operator scheduling, memory allocation, and swap decisions. SwapAdvisor explores the vast search space using a custom-designed genetic algorithm. Evaluations using a variety of large models show that SwapAdvisor can train models up to 12 times the GPU memory limit while achieving 53-99% of the throughput of a hypothetical baseline with infinite GPU memory.

This paper presents Tofu, a system that partitions very large DNN models across multiple GPU devices to reduce per-GPU memory footprint. Tofu is designed to partition a dataflow graph of fine-grained tensor operators used by platforms like MXNet and TensorFlow. In order to automatically partition each operator, we propose to describe the semantics of an operator in a simple language inspired by Halide. To optimally partition different operators in a dataflow graph, Tofu uses a recursive search algorithm that minimizes the total communication cost. Our experiments on an 8-GPU machine show that Tofu enables the training of very large CNN and RNN models. It also achieves 25% - 400% speedup over alternative approaches to train very large models.

Recently, flash-based embedded databases have gained their momentum in various control and monitoring systems, such as cyber-physical systems (CPSes). To support the functionality to access the historical data, a multiversion index is adopted to simultaneously maintain multiple versions of data items, as well as their index information. However, maintaining a multiversion index on flash memory incurs considerable performance overheads on garbage collection, which is to reclaim the spaces occupied by the outdated/invalid data items and their index information on flash memory. In this work, we propose an efficient garbage collection strategy to solve the garbage collection issues of flash-based multiversion databases. In particular, a version-tracking method is proposed to accelerate the performance on the process on identifying/reclaiming the space of invalid data and their indexes, and a pre-summary method is also designed to solve the cascading update problem that is caused by the write-once nature of flash memory and is worsened when more versions refer to the same data item. The capability of the proposed strategy is then verified by analytical and experimental studies.