Global Multi-Level KV Cache

Background

In the decoding phase of large language models (LLMs), frequent access to historical KV cache due to autoregressive generation creates a bottleneck in memory bandwidth. As model sizes and context windows expand (e.g., 128K Tokens consuming over 40GB of memory), the pressure on single-device memory increases dramatically. Existing solutions (such as vLLM) exhibit significant limitations in long-context scenarios: prefill time surges, severe memory bandwidth contention during decoding, and the need for excessive resource reservation to meet SLO requirements (TTFT < 2s, TBT < 100ms). This often results in GPU utilization below 40% and difficulties in leveraging cross-server resources. To address this, we propose a distributed global multi-level KV cache management system, adopting a memory-compute integrated architecture to break through single-machine resource constraints.

Feature Introduction

The xLLM Global KV Cache feature is primarily implemented through the following three modules:

• etcd: For cluster service registration, load information synchronization, and global cache state management.
• xLLM Service: For scheduling requests and managing all compute instances.
• xLLM: The compute instances handling requests.

The overall architecture is shown in the diagram below:

Usage Example

Preparation

Install Dependencies

• xLLM: Refer to Quick Start
• xLLM Service: Refer to xLLM Service

Usage Instructions

1. etcd Startup Configuration:

   ./etcd --listen-peer-urls=http://0.0.0.0:10999 --listen-client-urls=http://0.0.0.0:10998

2. xLLM Service Startup Configuration:

   ./xllm_master_serving --etcd_addr="127.0.0.1:10998" --http_server_port 28888 --rpc_server_port 28889 --tokenizer_path=/path/to/tokenizer_config_dir/

3. xLLM Startup Configuration: Add the following gflag parameters when starting xLLM:

   --enable_service_routing=true
   --enable_cache_upload=true
   # PD separation currently does not support Global KVCache Management
   --enable_disagg_pd=false