CSPS: A Communication-Efficient Sequence-Parallelism based Serving System for Transformer based Models with Long Prompts

• URL: http://arxiv.org/abs/2409.15104v1
• Date: Mon, 23 Sep 2024 15:16:29 GMT
• Title: CSPS: A Communication-Efficient Sequence-Parallelism based Serving System for Transformer based Models with Long Prompts
• Authors: Zeyu Zhang, Haiying Shen,
• Abstract summary: Long-sequence generative large-language model (LLM) applications have become increasingly popular. We find that the existing method for long sequences results in a high TimeToFirstToken (TTFT) due to sequential chunk processing. We propose two Sequence-Parallelism (SP) architectures for both tensor parallelism (TP) and non-TP.
• Score: 11.194752361478567
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Long-sequence generative large-language model (LLM) applications have become increasingly popular. In this paper, through trace-based experiments, we found that the existing method for long sequences results in a high Time-To-First-Token (TTFT) due to sequential chunk processing, long Time-Between-Tokens (TBT) from batching long-sequence prefills and decodes, and low throughput due to constrained key-value cache (KVC) for long sequences. To address these issues, we propose two Sequence-Parallelism (SP) architectures for both tensor parallelism (TP) and non-TP. However, SP introduces two challenges: 1) network communication and computation become performance bottlenecks; 2) the latter two issues above are mitigated but not resolved, and SP's resultant KV value distribution across GPUs still requires communication for decode, increasing TBT. Hence, we propose a Communication-efficient Sparse Attention (CSA) and communication-computation-communication three-phase pipelining. We also propose SP-based decode that processes decode separately from prefill, distributes KV values of a request across different GPUs, and novelly moves Query (Q) values instead of KV values to reduce communication overhead. These methods constitute a communication-efficient Sequence-Parallelism based LLM Serving System (SPS2). Our trace-driven evaluation demonstrates that SPS2 improves the average TTFT, TBT, and response time by up to 7.5x, 1.92x, and 9.8x and improves the prefill and decode throughput by 8.2x and 5.2x while maintaining the accuracy compared to Sarathi-Serve. We distributed our source code.

Related papers

• Optimizing LLM Inference: Fluid-Guided Online Scheduling with Memory Constraints [14.341123057506827]
  Large Language Models (LLMs) are indispensable in today's applications, but their inference procedure demands significant computational resources. This paper formulates LLM inference optimization as a multi-stage online scheduling problem. We develop a fluid dynamics approximation to provide a tractable benchmark that guides algorithm design.
  arXiv  Detail & Related papers  (2025-04-15T16:00:21Z)
• POD-Attention: Unlocking Full Prefill-Decode Overlap for Faster LLM Inference [9.164093249308419]
  We present POD-Attention -- the first GPU kernel that efficiently computes attention for hybrid batches. POD-Attention aims to maximize the utilization of both compute and memory bandwidth by carefully allocating the GPU's resources.
  arXiv  Detail & Related papers  (2024-10-23T17:06:56Z)
• Modelling Concurrent RTP Flows for End-to-end Predictions of QoS in Real Time Communications [5.159808922904932]
  We propose Packet-to-Prediction (P2P), a novel deep learning framework for predicting Quality of Service (QoS) metrics. We implement a streamlined architecture, capable of handling an unlimited number of RTP flows, and employ a multi-task learning paradigm to forecast four key metrics in a single shot. Our work is based on extensive traffic collected during real video calls, and conclusively, P2P excels comparative models in both prediction performance and temporal efficiency.
  arXiv  Detail & Related papers  (2024-10-21T10:16:56Z)
• AdaLog: Post-Training Quantization for Vision Transformers with Adaptive Logarithm Quantizer [54.713778961605115]
  Vision Transformer (ViT) has become one of the most prevailing fundamental backbone networks in the computer vision community. We propose a novel non-uniform quantizer, dubbed the Adaptive Logarithm AdaLog (AdaLog) quantizer.
  arXiv  Detail & Related papers  (2024-07-17T18:38:48Z)
• Not All Prompts Are Made Equal: Prompt-based Pruning of Text-to-Image Diffusion Models [59.16287352266203]
  We introduce Adaptive Prompt-Tailored Pruning (APTP), a novel prompt-based pruning method for text-to-image (T2I) diffusion models. APTP learns to determine the required capacity for an input text prompt and routes it to an architecture code, given a total desired compute budget for prompts. APTP outperforms the single-model pruning baselines in terms of FID, CLIP, and CMMD scores.
  arXiv  Detail & Related papers  (2024-06-17T19:22:04Z)
• USP: A Unified Sequence Parallelism Approach for Long Context Generative AI [1.973144426163543]
  Sequence parallelism (SP) is becoming key to unlocking the long-context capabilities of generative AI models. This paper investigates the state-of-the-art SP approaches, i.e. DeepSpeed-Ulysses and Ring-Attention, and proposes a unified SP approach. We achieved 47% MFU on two 8xA800 nodes using SP for the LLAMA3-8B model training using sequence length 208K.
  arXiv  Detail & Related papers  (2024-05-13T13:08:02Z)
• LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism [12.521026493432181]
  Existing large language models (LLMs) cannot efficiently serve variable-length requests in different phases. We propose a new parallelism paradigm, elastic sequence parallelism (ESP), to adapt to the variance between different requests and phases. LoongServe improves the maximum throughput by up to 3.85$times$ compared to the chunked prefill and 5.81$times$ compared to the prefill-decoding disaggregation.
  arXiv  Detail & Related papers  (2024-04-15T07:45:04Z)
• TCCT-Net: Two-Stream Network Architecture for Fast and Efficient Engagement Estimation via Behavioral Feature Signals [58.865901821451295]
  We present a novel two-stream feature fusion "Tensor-Convolution and Convolution-Transformer Network" (TCCT-Net) architecture. To better learn the meaningful patterns in the temporal-spatial domain, we design a "CT" stream that integrates a hybrid convolutional-transformer. In parallel, to efficiently extract rich patterns from the temporal-frequency domain, we introduce a "TC" stream that uses Continuous Wavelet Transform (CWT) to represent information in a 2D tensor form.
  arXiv  Detail & Related papers  (2024-04-15T06:01:48Z)
• RelayAttention for Efficient Large Language Model Serving with Long System Prompts [59.50256661158862]
  This paper aims to improve the efficiency of LLM services that involve long system prompts. handling these system prompts requires heavily redundant memory accesses in existing causal attention algorithms. We propose RelayAttention, an attention algorithm that allows reading hidden states from DRAM exactly once for a batch of input tokens.
  arXiv  Detail & Related papers  (2024-02-22T18:58:28Z)
• DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme Long Sequence Transformer Models [34.74093040678323]
  We introduce DeepSpeed-Ulysses, a novel, portable and effective methodology for enabling highly efficient and scalable LLM training. DeepSpeed-Ulysses at its core partitions input data along the sequence dimension and employs an efficient all-to-all collective communication for attention. Experiments show that DeepSpeed-Ulysses trains 2.5x faster with 4x longer sequence length than the existing method SOTA baseline.
  arXiv  Detail & Related papers  (2023-09-25T20:15:57Z)
• Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data [50.84488941336865]
  We propose a novel method called Fully- Spatial-Temporal Graph Neural Network (FC-STGNN) For graph construction, we design a decay graph to connect sensors across all timestamps based on their temporal distances. For graph convolution, we devise FC graph convolution with a moving-pooling GNN layer to effectively capture the ST dependencies for learning effective representations.
  arXiv  Detail & Related papers  (2023-09-11T08:44:07Z)
• Training Recommender Systems at Scale: Communication-Efficient Model and Data Parallelism [56.78673028601739]
  We propose a compression framework called Dynamic Communication Thresholding (DCT) for communication-efficient hybrid training. DCT reduces communication by at least $100times$ and $20times$ during DP and MP, respectively. It improves end-to-end training time for a state-of-the-art industrial recommender model by 37%, without any loss in performance.
  arXiv  Detail & Related papers  (2020-10-18T01:44:42Z)

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.