Long-context large language models (LLMs) have seen widespread adoption in recent years. However, during inference, the key-value (KV) cache—which stores intermediate activations—consumes significant memory, particularly as sequence lengths grow. Quantization offers a promising path to compress KV cache, but existing 2-bit approaches fall short of achieving optimal inference efficiency due to hardware-unfriendly algorithms and system implementations.

The Mixture of Experts (MoE) architecture enhances model quality by scaling up model parameters. However, its development is hindered in distributed training scenarios due to significant communication overhead and expert load imbalance. Existing methods, which only allow for coarse-grained overlapping of communication and computation, slightly alleviate communication costs but at the same time, they introduce a notable impairment of computational efficiency. Furthermore, current approaches to addressing load imbalance often compromise model quality.

General matrix-matrix multiplication (GEMM) is a fundamental operation in both deep learning and scientific computing. To accelerate these workloads, GPUs with a large number of streaming multiprocessors (SMs) are widely used. However, as modern GPUs scale in core count and adopt larger tile sizes, the wave quantization problem induced by partially filled waves results in growing hardware underutilization and substantially degraded performance. Existing solutions to this problem often suffer from low execution efficiency or introduce additional synchronization overhead.

Transformer-based models have made significant advancements across various domains, largely due to the self-attention mechanism’s ability to capture contextual relationships in input sequences. However, processing long sequences remains computationally expensive for Transformer models, primarily due to theO(n2) complexity associated with self-attention. To address this, sparse attention has been proposed to reduce the quadratic dependency to linear. Nevertheless, deploying the sparse transformer efficiently encounters two major obstacles: 1) Existing system optimizations are less effective for the sparse transformer due to the algorithm’s approximation properties leading to fragmented attention, and 2) the variability of input sequences results in computation and memory access inefficiencies. We present Raptor-T, a cutting-edge transformer framework designed for handling long and variable-length sequences. Raptor-T harnesses the power of the sparse transformer to reduce resource requirements for processing long sequences while also implementing system-level optimizations to accelerate inference performance. To address the fragmented attention issue, Raptor-T employs fused and memory-efficient Multi-Head Attention. Additionally, we introduce an asynchronous data processing method to mitigate GPU-blocking operations caused by sparse attention. Furthermore, Raptor-T minimizes padding for variable-length inputs, effectively reducing the overhead associated with padding and achieving balanced computation on GPUs. In evaluation, we compare Raptor-T’s performance against state-of-the-art frameworks on an NVIDIA A100 GPU. The experimental results demonstrate that Raptor-T outperforms FlashAttention-2 and FasterTransformer, achieving an impressive average end-to-end performance improvement of 3.41X and 3.71X, respectively.

In recent years, the Mixture-of-Experts (MoE) technique has gained widespread popularity as a means to scale pretrained models to exceptionally large sizes. Dynamic activation of experts allows for conditional computation, increasing the number of parameters of neural networks, which is critical for absorbing the vast amounts of knowledge available in many deep learning areas. However, despite the existing system and algorithm optimizations, there are significant challenges to be tackled when it comes to the inefficiencies of communication and memory consumption. In this paper, we present the design and implementation of MPMoE, a high-performance library that accelerates MoE training with adaptive and memory-efficient pipeline parallelism. Inspired by that the MoE training procedure can be divided into multiple independent sub-stages. We design a pipeline parallelism method for reducing communication latency by overlapping with computation operations. Further, we analyze the memory footprint breakdown of MoE training and identify that activations and temporary buffers are the primary contributors to the overall memory footprint. Toward memory efficiency, we propose memory reuse strategies to reduce memory requirements by eliminating memory redundancies. Finally, to optimize pipeline granularity and memory reuse strategies jointly, we propose a profile-based algorithm and a performance model to determine the configurations of MPMoE at runtime. We implement MPMoE upon PyTorch and evaluate it with common MoE models in two physical clusters, including 64 NVIDIA A100 GPU cards and 16 NVIDIA V100 GPU cards. Compared with the state-of-art approach, MPMoE achieves up to 2.3× speedup while reducing more than 30% memory footprint for training large models.

Temporal Graph Neural Networks(TGNNs) extend the success of Graph Neural Networks to dynamic graphs. Distributed TGNN training requires efficiently tackling temporal dependency, which often leads to excessive cross-device communication that generates significant redundant data. However, existing systems are unable to remove the redundancy in data reuse and transfer, and suffer from severe communication overhead in a distributed setting. This paper presents Sven, an algorithm and system co-designed TGNN training library for the end-to-end performance optimization on multi-node multi-GPU systems. Exploiting dependency patterns of TGNN models and characteristics of dynamic graph datasets, we design redundancy-free data organization and load-balancing partitioning strategies that mitigate the redundant data communication and evenly partition dynamic graphs at the vertex level. Furthermore, we develop a hierarchical pipeline mechanism integrating data prefetching, micro-batch pipelining, and asynchronous pipelining to mitigate the communication overhead. As the first scaling study on the memory-based TGNNs training, experiments conducted on an HPC cluster of 64 GPUs show that Sven can achieve up to 1.7x-3.3x speedup over the state-of-art approaches and a factor of up to 5.26x communication efficiency improvement.