From 81cd20f7162c2b7837bb184c35a18fc951c37587 Mon Sep 17 00:00:00 2001 From: Xianyan Jia Date: Wed, 11 Oct 2023 15:20:26 +0800 Subject: [PATCH] update readme (#18) Co-authored-by: xianyan.xianyanjia --- README.md | 11 +++++++++-- README_CN.md | 11 +++++++++-- 2 files changed, 18 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index 75ac0cd2..cb65466b 100644 --- a/README.md +++ b/README.md @@ -20,11 +20,15 @@ A flexible and efficient training framework for large-scale RLHF --- *Latest News* 🔥 -- [2023/08] We officially released ChatLearn! Check out our [documentation](docs/en/chatlearn.md)。 +- [2023/10] We officially released ChatLearn! Check out our [documentation](docs/en/chatlearn.md). --- -ChatLearn is a flexible and efficient training framework for large-scale RLHF. Chatlearn has the following advantages: +ChatLearn is a flexible and efficient training framework for large-scale RLHF. + +![RLHF Flow](docs/images/rlhf.png) + +Chatlearn has the following advantages: 1. **User-friendly programming interface**: Users can focus on programming individual models by wrapping a few functions, while the system takes care of resource scheduling, data and control flow transmission, and distributed execution. 2. **Multiple distributed acceleration backends**: Users can use different computation backends for model development, such as Megatron-LM and DeepSpeed. 3. **Hybrid parallel strategies**: Various parallel strategies can be employed, including Data Parallel, Tensor Parallel, Sequence Parallel, Pipeline Parallel, ZeRO, and the combination thereof. @@ -39,6 +43,7 @@ Please refer to the [documentation](https://chatlearn.readthedocs.io/en/latest/) 1. [Environment and Code Setup](docs/en/installation.md) 2. [End-to-End Training Tutorial with LLaMA Model](docs/en/tutorial.md) +3. [End-to-End Training Tutorial with BLOOM Model](docs/en/tutorial_bloom.md) # Supported Models @@ -49,6 +54,8 @@ The current ChatLearn framework supports RLHF training for GPT/LLaMA models of a | GPT (various scales of GPT models) | | LLaMA (`lmsys/vicuna-13b-v1.3`, `decapoda-research/llama-7b-hf`, `decapoda-research/llama-13b-hf`, `decapoda-research/llama-30b-hf`, `decapoda-research/llama-65b-hf`, etc.) | | LLaMA2 (`meta-llama/Llama-2-7b-hf`, `meta-llama/Llama-2-13b-hf`) | +| Baichuan (`baichuan-inc/Baichuan-7B`, `baichuan-inc/Baichuan-13B-Base`) | +| BLOOM (`bigscience/bloom-1b1`, `bigscience/bloom-7b1`, `bigscience/bloom`) | Note: The current performance benchmark is based on the GPT series models. diff --git a/README_CN.md b/README_CN.md index 594b063f..e938fc4e 100644 --- a/README_CN.md +++ b/README_CN.md @@ -18,11 +18,15 @@ --- *最新进展* 🔥 -- [2023/08] 正式开源 ChatLearn，更多介绍请参考我们的 [文档](docs/zh/chatlearn.md)。 +- [2023/10] 正式开源 ChatLearn，更多介绍请参考我们的 [文档](docs/zh/chatlearn.md)。 --- -ChatLearn 是一个灵活易用、大规模 RLHF 高效训练框架。ChatLearn 通过对模型计算逻辑的抽象，解耦了模型和计算 backend、分布式策略的绑定，提供灵活的资源调度机制，可以支持灵活的资源分配和并行调度策略。ChatLearn的优点总结如下: +ChatLearn 是一个灵活易用、大规模 RLHF 高效训练框架。ChatLearn 通过对模型计算逻辑的抽象，解耦了模型和计算 backend、分布式策略的绑定，提供灵活的资源调度机制，可以支持灵活的资源分配和并行调度策略。 + +![RLHF Flow](docs/images/rlhf.png) + +ChatLearn的特点如下: 1. **易用的编程接口**: ChatLearn提供通用的编程抽象，用户只需要封装几个函数即可完成模型构造。用户只需要专注于单模型的编程，系统负责资源调度、数据流传输、控制流传输、分布式执行等。 2. **多种分布式加速引擎**: 用户可以使用不同的计算 backend 进行模型建模，如 Megatron-LM、DeepSpeed 等。 3. **Hybrid 并行策略**: ChatLearn 支持各种并行策略组合：Data Parallel/Tensor Parallel/Sequence Parallel/Pipeline Parallel/ZeRO 及其组合。 @@ -36,6 +40,7 @@ ChatLearn 是一个灵活易用、大规模 RLHF 高效训练框架。ChatLearn 1. [环境和代码准备](docs/zh/installation.md) 2. [基于 LLaMA 模型的端到端训练教程](docs/zh/tutorial.md) +3. [基于 BLOOM 模型的端到端训练教程](docs/zh/tutorial_bloom.md) # 支持的模型 @@ -46,6 +51,8 @@ ChatLearn 是一个灵活易用、大规模 RLHF 高效训练框架。ChatLearn | GPT (GPT 系列各种规模的模型) | | LLaMA (`lmsys/vicuna-13b-v1.3`, `decapoda-research/llama-7b-hf`, `decapoda-research/llama-13b-hf`, `decapoda-research/llama-30b-hf`, `decapoda-research/llama-65b-hf`, etc.) | | LLaMA2 (`meta-llama/Llama-2-7b-hf`, `meta-llama/Llama-2-13b-hf`) | +| Baichuan (`baichuan-inc/Baichuan-7B`, `baichuan-inc/Baichuan-13B-Base`) | +| BLOOM (`bigscience/bloom-1b1`, `bigscience/bloom-7b1`, `bigscience/bloom`) | 注：当前的性能 benchmark 均基于 GPT 系列模型。