Skip to main content

A lightweight, effective and easy-to-extend inference runtime

Project description

xlite logo

Xlite轻量化推理运行时

介绍

xlite (GVirt前端):轻量级Transformer模型运行时,支持多样性算力协同,当前支持在昇腾硬件上高效运行。 xlite公开了Transformer运行所需的模型构图以及算子,所有算子基于昇腾AscendC/CCE开发。

背景与动机

在大模型推理场景中,传统的单流串行执行模式存在以下问题:

  • 核间负载不均:不同AICORE之间的任务分配不均衡,部分核心闲置
  • 资源浪费多:计算资源和传输资源利用率低,存在显著浪费
  • 执行时间长:Host CPU下发算子开销大,造成严重的host bond问题

xlite通过以下技术手段解决上述问题:

  • 多流并行:充分利用卡内资源,将单流串行改为多流并行执行
  • 核间负载均衡:核间负载均衡,提升资源利用率
  • CPU NPU协同:C++侧完全消除Python的GC、线程等干扰,简化Host tiling计算,去除小块内存申请释放及拷贝,消除Host bond

性能效果显著:

在GLM-4.7双机推理场景(40K输入、1K输出、prefix cache命中率约90%)下:

  • TPOT时延降低17%~30%
  • 吞吐提升13%~41%

详细性能数据参考 PR #7935

软件架构

image

xlite已适配vllm_ascend,可通过xlite_graph_config配置快速使能xlite加速效果,使用方法参考官方指导文档

xlite支持模型见模型列表

快速开始

安装

# 安装vllm_ascend, 可参考https://github.com/vllm-project/vllm-ascend/blob/main/README.md
# 安装xlite
pip install xlite

离线推理示例

import os
from vllm import LLM

# xlite默认支持decode-only模式, 可通过设置 "full_mode": True 使能full模式
model = LLM(model="path/to/Qwen3-32B", tensor_parallel_size=8, additional_config={"xlite_graph_config": {"enabled": True, "full_mode": True}})
outputs = model.generate("Hello, how are you?")

在线服务示例

vllm serve path/to/Qwen3-32B --tensor-parallel-size 8 --additional-config='{"xlite_graph_config": {"enabled": true, "full_mode": true}}'

开发指南

编译构建、容器镜像、源码安装等开发相关内容,请参考 开发指南

环境变量

xlite使用的环境变量及其配置方法,请参考 环境变量

目录结构

.
├── csrc/    - 轻量化运行时的核心代码
├── xlite/   - python代码,包括tools等
├── doc/     - 相关文档介绍
├── docker/  - 容器镜像Dockerfile
└── tests/   - 测试用例

致谢

本项目的核心算子初始版本和思路来自于华为终端小艺AI Infra团队的贡献,相关优化实现可参考论文:

《XY-Serve: End-to-End Versatile Production Serving for Dynamic LLM Workloads》 [ASPLOS 2026]

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distributions

If you're not sure about the file name format, learn more about wheel file names.

xlite-0.1.0rc9-cp311-cp311-manylinux2014_x86_64.whl (768.1 kB view details)

Uploaded CPython 3.11

xlite-0.1.0rc9-cp311-cp311-manylinux2014_aarch64.whl (734.1 kB view details)

Uploaded CPython 3.11

File details

Details for the file xlite-0.1.0rc9-cp311-cp311-manylinux2014_x86_64.whl.

File metadata

File hashes

Hashes for xlite-0.1.0rc9-cp311-cp311-manylinux2014_x86_64.whl
Algorithm Hash digest
SHA256 f9737bf75c0a31865f293eab628d03570e8e11d3881a6d3105ad29b975262fb1
MD5 539a76e20914a6f5b1895322206a245a
BLAKE2b-256 10d8346bf57e8f6f85892a3b864e25361acbc9453b08d5244b70170050cbf2d2

See more details on using hashes here.

File details

Details for the file xlite-0.1.0rc9-cp311-cp311-manylinux2014_aarch64.whl.

File metadata

File hashes

Hashes for xlite-0.1.0rc9-cp311-cp311-manylinux2014_aarch64.whl
Algorithm Hash digest
SHA256 a66448c7d5dccf9eacf8d7813606a585bb5b5bbca1b5c20467f56c8742286c3c
MD5 f9ab783d4cefdb9c9db40ca8a4f3952d
BLAKE2b-256 308034a898b82da1509de29cde60d3cfbb9c2d05a41d01eb82afdad5c64a11f2

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page