GliDe with a CaPE: A Low-Hassle Method to Accelerate Speculative Decoding

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.

Colossal-Auto: Unified Automation of Parallelization and Activation Checkpoint for Large-scale Models

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

MAP: Memory-aware Automated Intra-op Parallel Training For Foundation Models

Recently, large models have achieved the state of the art performances in various fields. In order to support large model training, we have to use distributed training techniques. However, finding an efficient distributed execution plan not only requires fine-grained model statistics, such as memory and computing overhead of each operator but also is a labor-intensive task even for an expert in the field of distributed training. In this paper, we introduce MAP, a compiler built upon PyTorch to implement Memory-aware Automated Parallelization. To profiling operator costs, existing training systems and machine learning pipelines either physically execute with respect to each operand or estimate the memory usage with a scaled input tensor, which are often time-consuming and misleading. Compared with existing methods, MAP provides an easy-to-use symbolic profiler to generate memory and computing statistics of an arbitrary PyTorch model with trivial time cost, so it will boost high productivity for ML developers. In addition, MAP can also seamlessly speed up different static planning tasks on computation graphs for PyTorch, and requires only a few lines of modification to user code to generate a new module instance that has a top-performing distributed execution plan. The source code is publicly available at https://github.com/hpcaitech/ColossalAI

Elixir: Train a Large Language Model on a Small GPU Cluster

In recent years, the number of parameters of one deep learning (DL) model has been growing much faster than the growth of GPU memory space. People who are inaccessible to a large number of GPUs resort to heterogeneous training systems for storing model parameters in CPU memory. Existing heterogeneous systems are based on parallelization plans in the scope of the whole model. They apply a consistent parallel training method for all the operators in the computation. Therefore, engineers need to pay a huge effort to incorporate a new type of model parallelism and patch its compatibility with other parallelisms. For example, Mixture-of-Experts (MoE) is still incompatible with ZeRO-3 in Deepspeed. Also, current systems face efficiency problems on small scale, since they are designed and tuned for large-scale training. In this paper, we propose Elixir, a new parallel heterogeneous training system, which is designed for efficiency and flexibility. Elixir utilizes memory resources and computing resources of both GPU and CPU. For flexibility, Elixir generates parallelization plans in the granularity of operators. Any new type of model parallelism can be incorporated by assigning a parallel pattern to the operator. For efficiency, Elixir implements a hierarchical distributed memory management scheme to accelerate inter-GPU communications and CPU-GPU data transmissions. As a result, Elixir can train a 30B OPT model on an A100 with 40GB CUDA memory, meanwhile reaching 84% efficiency of Pytorch GPU training. With its super-linear scalability, the training efficiency becomes the same as Pytorch GPU training on multiple GPUs. Also, large MoE models can be trained 5.3x faster than dense models of the same size. Now Elixir is integrated into ColossalAI and is available on its main branch.

EnergonAI: An Inference System for 10-100 Billion Parameter Transformer Models

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.

A Frequency-aware Software Cache for Large Recommendation System Embeddings

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.

Sky Computing: Accelerating Geo-distributed Computing in Federated Learning

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.

PatrickStar: Parallel Training of Pre-trained Models via a Chunk-based Memory Management

The pre-trained model (PTM) is revolutionizing Artificial intelligence (AI) technology. It learns a model with general language features on the vast text and then fine-tunes the model using a task-specific dataset. Unfortunately, PTM training requires prohibitively expensive computing devices, especially fine-tuning, which is still a game for a small proportion of people in the AI community. Enabling PTMs training on low-quality devices, PatrickStar now makes PTM accessible to everyone. PatrickStar reduces memory requirements of computing platforms by using the CPU-GPU heterogeneous memory space to store model data, consisting of parameters, gradients, and optimizer states. We observe that the GPU memory available for model data changes regularly, in a tide-like pattern, decreasing and increasing iteratively. However, the existing heterogeneous training works do not take advantage of this pattern. Instead, they statically partition the model data among CPU and GPU, leading to both memory waste and memory abuse. In contrast, PatrickStar manages model data in chunks, which are dynamically distributed in heterogeneous memory spaces. Chunks consist of stateful tensors which run as finite state machines during training. 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. Symbiosis with the Zero Redundancy Optimizer, PatrickStar scales to multiple GPUs using data parallelism, with the lowest communication bandwidth requirements and more efficient bandwidth utilization. Experimental results show PatrickStar trains a 12 billion parameters GPT model, 2x larger than the STOA work, on an 8-V100 and 240GB CPU memory node, and is also more efficient on the same model size.

Online Evolutionary Batch Size Orchestration for Scheduling Deep Learning Workloads in GPU Clusters

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.

Sequence Parallelism: Making 4D Parallelism Possible

Within Transformer, self-attention is the key module to learn powerful context-aware representations. However, self-attention suffers from quadratic memory requirements with respect to the sequence length, which limits us to process longer sequence on GPU. In this work, we propose sequence parallelism, a memory efficient parallelism method to help us break input sequence length limitation and train with longer sequence on GPUs. Compared with existing parallelism, our approach no longer requires a single device to hold the whole 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 communicate attention embeddings among GPUs. Inspired by ring all-reduce, we integrated ring-style communication with self-attention calculation and proposed Ring Self-Attention (RSA). Our implementation is fully based on PyTorch. Without extra compiler or library changes, our approach is compatible with data parallelism and pipeline parallelism. 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. We plan to integrate our sequence parallelism with data, pipeline and tensor parallelism to further train large-scale models with 4D parallelism in our future work.