{"id":18524921,"url":"https://github.com/paddlepaddle/paddlefleetx","last_synced_at":"2025-04-13T00:48:28.533Z","repository":{"id":36956985,"uuid":"161509700","full_name":"PaddlePaddle/PaddleFleetX","owner":"PaddlePaddle","description":"飞桨大模型开发套件，提供大语言模型、跨模态大模型、生物计算大模型等领域的全流程开发工具链。","archived":false,"fork":false,"pushed_at":"2024-05-24T01:10:17.000Z","size":667480,"stargazers_count":466,"open_issues_count":41,"forks_count":164,"subscribers_count":22,"default_branch":"develop","last_synced_at":"2025-04-13T00:48:19.376Z","etag":null,"topics":["benchmark","cloud","data-parallelism","distributed-algorithm","elastic","fleet-api","large-scale","lightning","model-parallelism","paddlecloud","paddlepaddle","pipeline-parallelism","pretraining","self-supervised-learning","unsupervised-learning"],"latest_commit_sha":null,"homepage":"","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"apache-2.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/PaddlePaddle.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null}},"created_at":"2018-12-12T15:45:00.000Z","updated_at":"2025-04-10T13:55:27.000Z","dependencies_parsed_at":"2024-06-19T03:03:17.851Z","dependency_job_id":"ab05ca3a-bbb3-435d-86d2-1e383e534228","html_url":"https://github.com/PaddlePaddle/PaddleFleetX","commit_stats":null,"previous_names":["paddlepaddle/fleet"],"tags_count":5,"template":false,"template_full_name":null,"repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PaddlePaddle%2FPaddleFleetX","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PaddlePaddle%2FPaddleFleetX/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PaddlePaddle%2FPaddleFleetX/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PaddlePaddle%2FPaddleFleetX/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/PaddlePaddle","download_url":"https://codeload.github.com/PaddlePaddle/PaddleFleetX/tar.gz/refs/heads/develop","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":248650437,"owners_count":21139672,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["benchmark","cloud","data-parallelism","distributed-algorithm","elastic","fleet-api","large-scale","lightning","model-parallelism","paddlecloud","paddlepaddle","pipeline-parallelism","pretraining","self-supervised-learning","unsupervised-learning"],"created_at":"2024-11-06T17:43:43.588Z","updated_at":"2025-04-13T00:48:28.507Z","avatar_url":"https://github.com/PaddlePaddle.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"\u003cp align=\"center\"\u003e\n  \u003cimg src=\"./paddlefleetx-logo.png\" align=\"middle\"  width=\"350\" /\u003e\n\u003c/p\u003e\n\n------------------------------------------------------------------------------------------\n\n\u003cp align=\"center\"\u003e\n    \u003ca href=\"./LICENSE\"\u003e\u003cimg src=\"https://img.shields.io/badge/license-Apache%202-dfd.svg\"\u003e\u003c/a\u003e\n    \u003ca href=\"https://github.com/PaddlePaddle/PaddleFleetX/releases\"\u003e\u003cimg src=\"https://img.shields.io/github/v/release/PaddlePaddle/PaddleFleetX?color=ffa\"\u003e\u003c/a\u003e\n    \u003ca href=\"\"\u003e\u003cimg src=\"https://img.shields.io/badge/python-3.7+-aff.svg\"\u003e\u003c/a\u003e\n    \u003ca href=\"https://github.com/PaddlePaddle/PaddleFleetX/graphs/contributors\"\u003e\u003cimg src=\"https://img.shields.io/github/contributors/PaddlePaddle/PaddleFleetX?color=9ea\"\u003e\u003c/a\u003e\n    \u003ca href=\"https://github.com/PaddlePaddle/PaddleFleetX/issues\"\u003e\u003cimg src=\"https://img.shields.io/github/issues/PaddlePaddle/PaddleFleetX?color=9cc\"\u003e\u003c/a\u003e\n    \u003ca href=\"https://github.com/PaddlePaddle/PaddleFleetX/stargazers\"\u003e\u003cimg src=\"https://img.shields.io/github/stars/PaddlePaddle/PaddleFleetX?color=ccf\"\u003e\u003c/a\u003e\n\u003c/p\u003e\n\n## 简介\n\nPaddleFleetX是基于飞桨深度学习框架开发的大模型套件，旨在提供高性能、灵活易用的大模型全流程应用能力，在**开发**、**训练**、**精调**、**压推**、**推理**、**部署**六大环节提供端到端全流程优化。\n\n\u003cp align=\"center\"\u003e\n  \u003cimg width=\"1000\" alt=\"飞桨大模型套件\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/ab5e87cc-df52-48cb-9968-8951d3b164ba\"\u003e\n\u003c/p\u003e\n\n## 特色介绍\n\n### 大模型开发：动静统一开发模式，4D混合并行策略灵活配置\n\n\u003cp align=\"center\"\u003e\n  \u003cimg width=\"771\" alt=\"大模型开发\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/95d1c0e8-df92-489b-8472-0a8b438cbfcf\"\u003e\n\u003c/p\u003e\n\n基于飞桨动静统一的开发模式，大模型套件全面使用动态图开发，在Generate API中可自动完成算子融合具备静态图的调试性能。全场景统一训练器Trainer可以轻松完成4D混合并行的配置，在预训练与精调环节皆可使用。\n\n### 大模型训练：发挥基础计算潜能、全面提升分布式效率\n\n飞桨针对大模型训练，对数据读取、混合精度计算策略、高性能算子库、并行策略自动寻优、流水线调度的整个全流程实现优化，助力文心大模型训练速度提升3倍。\n\n\u003cp align=\"center\"\u003e  \n  \u003cimg width=\"1000\" alt=\"飞桨支持大模型训练\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/3874440d-0b0c-4730-bbcb-f9b87900d75f\"\u003e\n\u003c/p\u003e\n\n\n\n### 大模型精调：主流精调算法实现性能全面领先\n\n提供了主流的精调算法，包括SFT、Prefix-Tuning、LoRA三种主流的精调算法，有效降低的大模型训练的资源门槛。统一的训练器Trainer实现了预训练加速技术在精调场景的复用，并通过变长数据流优化大幅提升精调性能。\n\n\u003cp align=\"center\"\u003e\n  \u003cimg width=\"800\" alt=\"大模型精调\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/0dad24ae-0549-4166-8426-b0a471a82450\"\u003e\n\u003c/p\u003e\n\n\n### 大模型压缩：自研量化压缩算法实现无损量化\n\n飞桨自研的Shift-SmoothQuant算法相比SmoothQuant算法可以实现更平滑的激活分布，有效提升量化后模型的精度度和生成结果的稳定性。通过PaddleSlim的大模型压缩工具，我们在 C-Eval 和 NL2SQL 两个数据集上对主流开源大模型可以实现无损量化。更多技术介绍与使用说明可以参考[PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim)。\n\n\u003cp align=\"center\"\u003e\n  \u003cimg width=\"350\" alt=\"模型压缩\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/8b8334d6-dc1a-4ab8-a2f6-dbbece6f0e1e\"\u003e\n\u003c/p\u003e\n\u003cp align=\"center\"\u003e\n  \u003cimg width=\"798\" alt=\"模型压缩\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/badb3f10-314a-4259-8179-08f940197352\"\u003e\n\u003c/p\u003e\n\n### 大模型推理：针对大模型场景特性匹配最优量化推理方案\n\nPaddle Inference针对大模型Prompt阶段与Token Generation阶段的计算特性的不同，在通用场景提供静态量化，在访存受限场景提供混合量化与低比特的推理方案。\n\n\u003cp align=\"center\"\u003e\n  \u003cimg width=\"1000\" alt=\"飞桨支撑大模型推理\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/6bf2a373-a550-4359-9285-6fa4337e550d\"\u003e\n\u003c/p\u003e\n\n\u003cp align=\"center\"\u003e\n  \u003cimg width=\"400\" alt=\"推理引擎\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/8d9ab6f9-fc63-4485-bcf2-f9791b1de273\"\u003e\n\u003c/p\u003e\n\n\n### 大模型部署：实时感知负载动态插入请求，最大化硬件利用率\n\n由于大模型生成场景解码阶段耗时较长，且不同Query下生成长度不一，为了最大化服务吞吐，我们在FastDeploy服务框架结合推理引擎实现了动态插入技术，科实时感知服务负载，动态插入用户请求最大化推理硬件利用率。\n\n\u003cp align=\"center\"\u003e\n  \u003cimg width=\"350\" alt=\"大模型服务部署\" src=\"https://github.com/PaddlePaddle/PaddleFleetX/assets/1371212/d2e38f78-9088-4b1a-a9bd-1018385b5b86\"\u003e\n\u003c/p\u003e\n\n\n## PaddleFleetX 应用案例\n\n### 大语言模型\n\n基于PaddleFleetX的核心能力，我们在PaddleNLP中提供了丰富的大语言模型全流程开发与应用示例，更多详细使用说明可以参考[PaddleNLP大语言模型](https://github.com/PaddlePaddle/PaddleNLP/tree/develop/llm)。\n\n### 跨模态大模型\n\n除了大语言模型外，PaddleFleetX还提供跨模态大模型的开发与训练，如多模态预训练、文生图扩散模型等，覆盖图片、文本、视频和音频等模态，更多详细使用说明可以参考[PaddleMIX](https://github.com/PaddlePaddle/PaddleMIX)。\n\n### 生物计算大模型\n\n在生物计算领域，基于飞桨4D并行策略与高性能优化，我们在PaddleHelix中提供众多业界领先的生物计算预训练模型，更多详细使用说明可以参考[PaddleHelix](https://github.com/PaddlePaddle/PaddleHelix)。\n\n\n## Citation\n\n```\n@misc{paddlefleetx,\n    title={PaddleFleetX: An Easy-to-use and High-Performance One-stop Tool for Deep Learning},\n    author={PaddleFleetX Contributors},\n    howpublished = {\\url{https://github.com/PaddlePaddle/PaddleFleetX}},\n    year={2022}\n}\n```\n\n## License\n\nPaddleFleetX 基于 [Apache 2.0 license](./LICENSE) 许可发布。\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fpaddlepaddle%2Fpaddlefleetx","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fpaddlepaddle%2Fpaddlefleetx","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fpaddlepaddle%2Fpaddlefleetx/lists"}