TokenWeave: Efficient Compute-Communication Overlap for Distributed LLM Inference

• URL: http://arxiv.org/abs/2505.11329v2
• Date: Thu, 10 Jul 2025 08:40:35 GMT
• Title: TokenWeave: Efficient Compute-Communication Overlap for Distributed LLM Inference
• Authors: Raja Gond, Nipun Kwatra, Ramachandran Ramjee,
• Abstract summary: Distributed inference of large language models (LLMs) can introduce overheads of up to 20% even over GPUs connected via high-speed interconnects such as NVLink.<n>We present TokenWeave to address these challenges.<n>Our evaluations demonstrate up to 1.29x speedup in latency and 1.26x higher throughput across multiple models and workloads.
• Score: 10.054508615667071
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Distributed inference of large language models (LLMs) can introduce overheads of up to 20% even over GPUs connected via high-speed interconnects such as NVLink. Multiple techniques have been proposed to mitigate these overheads by decomposing computations into finer-grained tasks and overlapping communication with sub-tasks as they complete. However, fine-grained decomposition of a large computation into many smaller computations on GPUs results in overheads. Furthermore, the communication itself uses many streaming multiprocessors (SMs), adding to the overhead. We present TokenWeave to address these challenges. TokenWeave proposes a Token-Splitting technique that divides the tokens in the inference batch into two approximately equal subsets in a wave-aware manner. The communication of one subset is then overlapped with the computation of the other. In addition, TokenWeave optimizes the order of the layer normalization computation with respect to communication operations and implements a novel fused AllReduce--RMSNorm kernel that carefully leverages Multimem instruction support available on NVIDIA Hopper GPUs. These optimizations allow TokenWeave to perform communication and RMSNorm using only 2-8 SMs. Moreover, our kernel enables the memory-bound RMSNorm to be overlapped with the other batch's computation, providing additional gains. Our evaluations demonstrate up to 1.29x speedup in latency and 1.26x higher throughput across multiple models and workloads. In several settings, TokenWeave results in better performance compared to an equivalent model with all communication removed.

Related papers

• Splitwiser: Efficient LM inference with constrained resources [0.29260385019352086]
  Splitwiser is a methodology that splits the two phases of an LLM inference request onto the same GPU.<n>By eliminating the need to transfer data across devices, Splitwiser aims to minimize network-related overheads.<n>We implement our proposed multiprocessing design on two widely-used and independent LLM architectures: Huggingface and vLLM.
  arXiv  Detail & Related papers  (2025-04-21T00:21:08Z)
• EPS-MoE: Expert Pipeline Scheduler for Cost-Efficient MoE Inference [49.94169109038806]
  This paper introduces EPS-MoE, a novel expert pipeline scheduler for MoE that surpasses the existing parallelism schemes.<n>Our results demonstrate at most 52.4% improvement in prefill throughput compared to existing parallel inference methods.
  arXiv  Detail & Related papers  (2024-10-16T05:17:49Z)
• Enhancing MOTION2NX for Efficient, Scalable and Secure Image Inference using Convolutional Neural Networks [4.407841002228536]
  We use the ABY2.0 SMPC protocol implemented on the C++ based MOTION2NX framework for secure convolutional neural network (CNN) inference application with semi-honest security. We also present a novel splitting algorithm that divides the computations at each CNN layer into multiple chunks.
  arXiv  Detail & Related papers  (2024-08-29T09:50:21Z)
• FLUX: Fast Software-based Communication Overlap On GPUs Through Kernel Fusion [9.5114389643299]
  This paper proposes a novel method, Flux, to significantly hide communication latencies with dependent computations for GPUs. Flux can potentially overlap up to 96% of communication given a fused kernel. Overall, it can achieve up to 1.24x speedups for training over Megatron-LM on a cluster of 128 GPU with various GPU generations and interconnects.
  arXiv  Detail & Related papers  (2024-06-11T00:17:39Z)
• Hardware-Aware Parallel Prompt Decoding for Memory-Efficient Acceleration of LLM Inference [19.167604927651073]
  Auto-regressive decoding of Large Language Models (LLMs) results in significant overheads in their hardware performance. We propose a novel parallel prompt decoding that requires only $0.0002$% trainable parameters, enabling efficient training on a single A100-40GB GPU in just 16 hours. Our approach demonstrates up to 2.49$times$ speedup and maintains a minimal memory overhead of just $0.0004$%.
  arXiv  Detail & Related papers  (2024-05-28T22:19:30Z)
• CORE: Common Random Reconstruction for Distributed Optimization with Provable Low Communication Complexity [110.50364486645852]
  Communication complexity has become a major bottleneck for speeding up training and scaling up machine numbers. We propose Common Om REOm, which can be used to compress information transmitted between machines.
  arXiv  Detail & Related papers  (2023-09-23T08:45:27Z)
• Boosting the effective performance of massively parallel tensor network state algorithms on hybrid CPU-GPU based architectures via non-Abelian symmetries [0.0]
  Non-Abelian symmetry related tensor algebra based on Wigner-Eckhart theorem is fully detached from the conventional tensor network layer. We have achieved an order of magnitude increase in performance with respect to results reported in arXiv:2305.05581 in terms of computational complexity. Our solution has an estimated effective performance of 250-500 TFLOPS.
  arXiv  Detail & Related papers  (2023-09-23T07:49:53Z)
• Tensor Slicing and Optimization for Multicore NPUs [2.670309629218727]
  This paper proposes a compiler optimization pass for Multicore NPUs, called Slicing Optimization (TSO) TSO identifies the best tensor slicing that minimizes execution time for a set of CNN models. Results show that TSO is capable of identifying the best tensor slicing that minimizes execution time for a set of CNN models.
  arXiv  Detail & Related papers  (2023-04-06T12:03:03Z)
• Efficient Dataset Distillation Using Random Feature Approximation [109.07737733329019]
  We propose a novel algorithm that uses a random feature approximation (RFA) of the Neural Network Gaussian Process (NNGP) kernel. Our algorithm provides at least a 100-fold speedup over KIP and can run on a single GPU. Our new method, termed an RFA Distillation (RFAD), performs competitively with KIP and other dataset condensation algorithms in accuracy over a range of large-scale datasets.
  arXiv  Detail & Related papers  (2022-10-21T15:56:13Z)
• Instant Neural Graphics Primitives with a Multiresolution Hash Encoding [67.33850633281803]
  We present a versatile new input encoding that permits the use of a smaller network without sacrificing quality. A small neural network is augmented by a multiresolution hash table of trainable feature vectors whose values are optimized through a gradient descent. We achieve a combined speed of several orders of magnitude, enabling training of high-quality neural graphics primitives in a matter of seconds.
  arXiv  Detail & Related papers  (2022-01-16T07:22:47Z)
• Communication-Efficient Distributed Stochastic AUC Maximization with Deep Neural Networks [50.42141893913188]
  We study a distributed variable for large-scale AUC for a neural network as with a deep neural network. Our model requires a much less number of communication rounds and still a number of communication rounds in theory. Our experiments on several datasets show the effectiveness of our theory and also confirm our theory.
  arXiv  Detail & Related papers  (2020-05-05T18:08:23Z)
• Straggler-aware Distributed Learning: Communication Computation Latency Trade-off [56.08535873173518]
  Straggling workers can be tolerated by assigning redundant computations and coding across data and computations. In most existing schemes, each non-straggling worker transmits one message per iteration to the parameter server (PS) after completing all its computations. Imposing such a limitation results in two main drawbacks; over-computation due to inaccurate prediction of the straggling behaviour, and under-utilization due to treating workers as straggler/non-straggler.
  arXiv  Detail & Related papers  (2020-04-10T08:39:36Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.