Fix huggingface download error

Update README.md

Dockerfile

@@ -0,0 +1,28 @@
+# Use the specified CUDA image as the base image
+FROM nvidia/cuda:12.0.0-cudnn8-devel-ubuntu20.04
+
+# Set environment variables to avoid interactive prompts during package installation
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Change the APT package server (fix "Hash Sum Mismatch" problem)
+RUN sed -i 's/archive.ubuntu.com/repo.catswords.com/g' /etc/apt/sources.list
+RUN sed -i 's/security.ubuntu.com/repo.catswords.com/g' /etc/apt/sources.list
+
+# Update the package list and install Python 3 and necessary packages
+RUN apt-get update && \
+    apt-get install -y python3 python3-pip python3-dev && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Set Python 3 as the default python
+RUN update-alternatives --install /usr/bin/python python /usr/bin/python3 1
+
+# Create a working directory
+WORKDIR /app
+
+# Copy the requirements.txt file and install Python packages if available
+COPY requirements.txt ./
+RUN pip3 install --no-cache-dir -r requirements.txt
+
+# Command to run the application (update this as needed)
+CMD ["tail", "-f", "/dev/null"]

README.md

@@ -1,104 +1,173 @@
 <!--
- * @Author: LiangSong(sl12160010@gmail.com)
+ * @Author: s-JoL(sl12160010@gmail.com)
  * @Date: 2023-03-10 21:18:35
- * @LastEditors: LiangSong(sl12160010@gmail.com)
- * @LastEditTime: 2023-04-16 23:49:06
+ * @LastEditors: s-JoL(sl12160010@gmail.com)
+ * @LastEditTime: 2023-05-17 22:44:35
  * @FilePath: /Open-Llama/README.md
  * @Description: 
  * 
- * Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+ * Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 -->
+[**中文**](./README_zh.md) | [**English**](./README.md)
+
+![camel](assets/logo.png)
+
 # Open-Llama
 
-[English](https://github.com/Bayes-Song/Open-Llama/blob/main/README_en.md)
+<p align="center">
+    <img alt="GitHub" src="https://img.shields.io/github/license/s-JoL/Open-Llama.svg?color=blue&style=flat-square">
+    <img alt="GitHub release (latest by date)" src="https://img.shields.io/github/v/release/s-JoL/Open-Llama">
+    <img alt="GitHub top language" src="https://img.shields.io/github/languages/top/s-JoL/Open-Llama">
+    <img alt="GitHub last commit" src="https://img.shields.io/github/last-commit/s-JoL/Open-Llama">
+</p>
 
-Open-Llama是一个开源项目，提供了一整套用于构建大型语言模型的训练流程，从数据集准备到分词、预训练、指令调优，以及强化学习技术 RLHF。
+Open-Llama is an open-source project that offers a complete training pipeline for building large language models, ranging from dataset preparation to tokenization, pre-training, prompt tuning, lora, and the reinforcement learning technique RLHF.
 
-## 进展
+**You can try this model directly from the [Demo](http://home.ustc.edu.cn/~sl9292/).**
 
-**采用FastChat项目相同方法测评Open-Llama的效果和GPT3.5的效果对比，经过测试在中文问题上可以达到GPT3.5 84%的水平，具体测试结果和CheckPoint将在近期放出**
+Join [discord](https://discord.gg/TrKxrTpnab) to discuss the development of large language models.
+
+## **Main contents**
+
+- **Support Transformers/HuggingFace.** The CheckPoint after Instruct-tuning is open-source on [Hugging Face: s-JoL/Open-Llama-V2](https://huggingface.co/s-JoL/Open-Llama-V2).
+
+- **By adopting the same evaluation method as the FastChat project, Open-Llama's performance is compared to GPT3.5’s. After testing, it can reach 89% of GPT3.5's performance on Chinese questions.**
+
+- **The training speed reaches 3620 tokens/s, faster than the 3370 tokens/s reported in the original Llama paper, reaching the current state-of-the-art level.**
+
+``` python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("s-JoL/Open-Llama-V2", use_fast=False)
+model = AutoModelForCausalLM.from_pretrained("s-JoL/Open-Llama-V2", device_map="auto")
+
+inputs = tokenizer('user:implement quick sort in python\nsystem:', return_tensors='pt', return_attention_mask=False, add_special_tokens=False)
+for k, v in inputs.items():
+   inputs[k] = v.cuda()
+pred = model.generate(**inputs, max_new_tokens=512, do_sample=True)
+print(tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))
 
-经过Instruct-tuning的CheckPoint已开源在[HuggingFace](https://huggingface.co/s-JoL/Open-Llama-V1)。使使用ckpt需要先用下面命令安装最新版本Transformers
-``` base
-pip install git+https://github.com/s-JoL/transformers.git@dev
 ```
-模型已提交[PR](https://github.com/huggingface/transformers/pull/22795)合并至Transformers main分支。
+The CheckPoint after pre-training only is also uploaded to [s-JoL/Open-Llama-V2-pretrain](https://huggingface.co/s-JoL/Open-Llama-V2-pretrain).
 
-我们完成了300B token的预训练，总共训练80 K step，Global Batch Size和Llama中一致为4M。
-使用总共7部分数据构成Instruction-tuning数据，模型具有一定的编程能力、数学能力和多轮对话能力，具体数据见Instruction-Tuning部分。
+We have completed 330B token pre-training, training a total of 80 K steps. The Global Batch Size is consistent with Llama at 4M.
+Using a total of 7 parts of data to constitute the Instruction-tuning data, the model has certain programming abilities, mathematical abilities, and multi-turn dialogue abilities. Specific data can be found in the Instruction-Tuning section.
 
-[Demo](http://home.ustc.edu.cn/~sl9292/)
+Below is a display of the model's multi-turn dialogue ability regarding code:
 
-我们参考一些对文心一言的测试也简单测试一下我们的模型，原始报道 [百度“文心一言”测试：国内生成式 AI 什么水平？](https://www.8btc.com/article/6809666)
+![image4](assets/multiturn_chat_en.jpg)
 
-本模型的效果如下图，更多结果还待进一步测试。由于国内网络问题，使用上面的Demo可能出现请求丢失的情况，如长时间无响应可刷新重试
-![image1](assets/image1.png)![image2](assets/image2.png)![image3](assets/image3.png)
+## **Updates**
 
-下面是一个关于代码的多轮对话能力的展示
+**[2023.5.8] Release v2.1**
 
-![image4](assets/multiturn_chat.jpeg)
-<!-- 我们简单预估一下达到上面效果的一个花费，训练40K step使用了1.5亿条预训练数据，大约为110B token，总共训练时间76h，按Google Cloud的A100报价花费大约为19152美元。后续的Instruction-tuning训练了12k Step，使用1.6M条数据，总共训练时间3.4h，大约花费342美元。因此从0开始训练一个这样的模型总花费不到20000美元。 -->
+- This update adds support for larger model training. Using DeepSpeed stage3 + offload + activation checkpoint, you can **train a 65B model with A100-80G**. 
 
-目前模型在数学方面和代码方面表现明显较差，这一方面和训练数据有关，另一方面我认为也是模型大小所造成的，然而这方面的逻辑推理能力是一个可用的模型所必备，因此后续更新会关注提升相关能力。
-## **特性**
+- The peft library is introduced to **support training such as lora**.
 
-### 易用性
+- The following table compares the training speed of Open-Llama and the original Llama, and the performance data of Llama is quoted from the original Llama paper.
 
-我们认为易用性是构建大型语言模型时最重要的特性之一。为了使 Open-LLAMA 更加易于使用，我们特别注重了以下几点：
 
-- **最简实现**：我们采用了最简单的实现方式，降低了入门的门槛，让初学者也能轻松上手。
-- **流程完整**：我们发布了从数据集构建到训练的完整代码，使得构建一个大语言模型的每一步流程都清晰可见。
+|                | DeepSpeed Stage | Offload | Activation Checkpoint | Total Token | GPU hours | Speed token/s/gpu | Batch Size |
+|----------------|-----------------|---------|-----------------------|-------------|-----------|-------------------|------------|
+| Open-Llama 7B  | 1               | False   | False                 | 173.7B      | 13412     | 3620              | 2          |
+| Open-Llama 13B | 3               | False   | True                  | -           | -         | 1856              | 24         |
+| Open-Llama 33B | 3               | False   | True                  | -           | -         | 708               | 12         |
+| Open-Llama 65B | 3               | True    | True                  | -           | -         | 369               | 12         |
+| Llama 7B       | -               | -       | -                     | 1T          | 82432     | 3370              | -          |
+| Llama 13B      | -               | -       | -                     | 1T          | 135168    | 2055              | -          |
+| Llama 33B      | -               | -       | -                     | 1.4T        | 530432    | 733               | -          |
+| Llama 65B      | -               | -       | -                     | 1.4T        | 1022362   | 380               | -          |
 
-### 高性能
+**[2023.4.28] Release v2.0**
 
-由于训练大语言模型的成本高昂，因此在构建大型语言模型时，高性能也是非常重要的。为了实现高性能的训练，我们发布使用了以下技术：
+This update mainly includes the following aspects, increasing the effective training speed by **50%** compared to the v1 version, reducing padding from **30%** to **5%**, and improving training speed from **3200 tokens/s** to **3620 tokens/s**. 0.95 * 3620 / (0.7 * 3200) = 1.521
 
-- **Fused CUDA kernel**：使用[xformers](https://github.com/facebookresearch/xformers)中提供的 fused CUDA kernel 可以将多个操作融合在一起，减少了 GPU 和 CPU 之间的数据传输，从而提高了训练效率。
-- **并行化训练**：我们使用[Accelerate](https://huggingface.co/docs/accelerate/index)库支持在多个 GPU 上进行并行化训练，以加快训练速度。
+1. Use Hugging Face's datasets library for data reading, with the process as follows:
+   1. Use the transform function to unify data formats from different datasets to {'text': 'xxx'}
+   2. Tokenize using Tokenizer
+   3. Sample long sequences; currently, three modes are provided: truncation, sampling (refer to the [Gopher paper](https://arxiv.org/abs/2112.11446)), and splitting
+   4. Optional: concatenate texts from different docs, reducing padding in the data and accelerating training. In the v1 version, padding accounted for **30%**; after concatenation, padding is reduced to **5%**.
+2. Add Trainer, which can be reused for both pre-training and instruction fine-tuning, see solver/trainer.py
+3. Unify the pre-training and instruction fine-tuning training entry to train_lm.py
+4. Provide more convenient configuration, see configs/pretrain_config.yaml
+5. Provide functionality to continue pre-training based on other pre-trained models and supplementing vocabulary
+6. Resuming training from a checkpoint is supported, including loading optimizer parameters/learning rate and skipping duplicate data
 
-对于7B模型，使用Transformers中Pytorch原生版本的Llama模型训练训练速度为1378 token/s/gpu，使用本代码库训练速度达到3290 token/s/gpu，基本达到[Llama原文](https://arxiv.org/pdf/2302.13971.pdf)中的3370 token/s/gpu。
-如果使用500B token进行预训练，需要训练43000 GPU时。按照Google Cloud上A100-80G Spot的价格计算，8卡每小时价格为12.6美元，则总价格为67725美元。
-当使用未加速版本训练时，价格为158744美元。最终降低训练成本9万美元。
-更多测试可见[和其他开源模型性能对比](https://github.com/Bayes-Song/Open-Llama#%E5%92%8C%E5%85%B6%E4%BB%96%E5%BC%80%E6%BA%90%E6%A8%A1%E5%9E%8B%E6%80%A7%E8%83%BD%E5%AF%B9%E6%AF%94)。
-### 通用性
+[2023.4.16] Release v1.0
 
-在训练语言模型时，我们希望能够构建一个通用的模型，可以适用于不同的语言和不同的领域。为了实现这一点，我们采用了以下策略：
+Basic pre-training and instruction fine-tuning codes are provided, with a training speed comparable to that of the original Llama. The pre-trained and fine-tuned models are already open-sourced on Hugging Face.
 
-- **多语言支持**：我们支持多种语言的语料库，包括英语、中文、日语等多种语言，让用户可以根据自己的需求进行选择。
-- **领域通用性**：我们希望模型不仅能在日常问题上能产生帮助，同时希望在专业领域如科学、法律等也能帮助人类。
+v1 version code can be seen at https://github.com/s-JoL/Open-Llama/tree/v1.0
 
-## **要求**
+## **Features**
 
-- Python 3.7 或更高版本
+### Easy to use
+
+We believe that ease of use is one of the most important features when building large language models. To make Open-LLAMA more accessible, we have focused on the following aspects:
+
+- **Minimal implementation**: We have adopted the simplest implementation methods, lowering the entry threshold and allowing beginners to get started with ease.
+- **Complete pipeline**: We have published the complete code from dataset construction to training, making every step in the process of building a large language model clear and visible.
+
+### High performance
+
+Due to the high cost of training large language models, high performance is also crucial when building them. To achieve high-performance training, we have employed the following techniques:
+
+- **Fused CUDA kernel**: Using the fused CUDA kernel provided in [xformers](https://github.com/facebookresearch/xformers) can fuse multiple operations, reducing data transfer between the GPU and CPU, thereby improving training efficiency.
+- **Parallelized training**: We employ the [Accelerate](https://huggingface.co/docs/accelerate/index) library to support parallelized training on multiple GPUs to speed up the training process.
+
+For a 7B model, the training speed with the native PyTorch Llama model in Transformers is **1378 tokens/s/GPU**. Using this codebase, the training speed reaches **3626 tokens/s/GPU**, exceeding **3370 tokens/s/GPU** reported in the [original Llama paper](https://arxiv.org/pdf/2302.13971.pdf).
+
+If pre-training with 500B tokens, 38300 GPU hours are required. According to the hourly price for 8 A100-80G Spot GPUs on Google Cloud, which is 12.6 US dollars, the total cost is 60,300 US dollars.
+When using the unaccelerated version for training, the cost is 158,744 US dollars. The final training cost is reduced by 98,000 US dollars.
+
+For more testing, see [performance comparison with other open-source models](https://github.com/s-JoL/Open-Llama#%E5%92%8C%E5%85%B6%E4%BB%96%E5%BC%80%E6%BA%90%E6%A8%A1%E5%9E%8B%E6%80%A7%E8%83%BD%E5%AF%B9%E6%AF%94).
+
+### Versatility
+
+When training language models, our goal is to build a versatile model that can handle different languages and domains. To achieve this, we have employed the following strategies:
+
+- **Multi-language support**: We support multiple language corpora, including English, Chinese, Japanese, and many other languages, allowing users to choose according to their requirements.
+- **Domain versatility**: We hope that the model can not only help with everyday questions but also assist in professional domains such as science, law, etc.
+- **Interaction with the world**: By incorporating reinforcement learning (RL), we hope to give the model the ability to interact with the world.
+
+## **Requirements**
+
+- Python 3.7 or higher
 - PyTorch 1.13
-- 特殊版本的[Transformers库](https://github.com/Bayes-Song/transformers)
-- [Accelerate库](https://huggingface.co/docs/accelerate/index)
-- CUDA 11.6 或更高版本（用于 GPU 加速，基于CUDA11.7进行测试）
+- [Transformers library](https://github.com/huggingface/transformers)
+- [Accelerate library](https://huggingface.co/docs/accelerate/index)
+- CUDA 11.6 or higher (for GPU acceleration)
+- Hardware configuration: currently using (64 CPU, 1000G Memory, 8xA100-80G) x N. There is a rather curious phenomenon that when more CPUs are used, the system runs slightly slower. I speculate this may have something to do with the multi-processing of dataloader.
 
-## **入门指南**
-### 安装
+## **Getting Started**
+### Installation
 
-使用下面的命令安装相关依赖
+Use the following command to install related dependencies:
 
 ```bash
 pip install -r requirements.txt
 ```
 
-### 数据集准备
+### Dataset Preparation
 
-目前给出了智源开源的悟道数据集和EleutherAI开源的the pile数据集。数据集下载和处理代码在data目录下。
-其中悟道数据集由于需要同意一些协议才能下载因此可能需要修改一下download_wudao中的链接，[悟道](https://data.baai.ac.cn/details/WuDaoCorporaText)。
+Currently provided are the Wudao dataset open-sourced by Zhiyuan and the Pile dataset open-sourced by EleutherAI. Dataset download and processing scripts are located in the data directory.
+Due to the required agreement for downloading the Wudao dataset, you may need to modify the link in download_wudao. [Wudao](https://data.baai.ac.cn/details/WuDaoCorporaText).
 
-运行下面的命令进行数据下载并进行分片
+Thanks to [@skepsun](https://github.com/skepsun)'s suggestion, using scidb to download the wudao dataset does not require login, and the download is more stable. https://github.com/s-JoL/Open-Llama/issues/42. 
+
+**Note that data download may fail. It is recommended to divide the download and processing in the script into two parts for multiple attempts, which will automatically resume downloads from breakpoints.**
+
+Run the following commands to download the data and perform partitioning:
 ```bash
 bash data/download_the_pile.sh
 bash data/download_wudao.sh
 ```
-数据将按照每个文件最大16384行存储为小文件，便于后续使用多进程训练时进行读取。存储格式为jsonl.zst，使用zstd进行压缩，最终数据大小为519.5G，合计16466个文件。
+The data will be stored as small files, with a maximum of 16384 lines per file, for easy reading during multi-process training. The storage format is jsonl.zst, compressed using zstd, with a final data size of 519.5 GB, consisting of 16,466 files in total.
 
-其中the pile数据集包含210607728行json line，悟道数据集包含59132213行json line。
+The Pile dataset contains 210,607,728 JSON lines, while the Wudao dataset contains 59,132,213 JSON lines.
 
-具体数据格式如下
+The specific data format is as follows:
 ```
 WuDao
 {'id': 1, 'dataType': '百科', 'title': 'some title', 'content': 'some content'}
@@ -106,36 +175,68 @@ WuDao
 The Pile
 {'text': 'some text', 'meta': {'pile_set_name': 'Github'}}
 ```
-验证数据完整性可见 [issue](https://github.com/s-JoL/Open-Llama/issues/5)
+Check the data integrity in [issue](https://github.com/s-JoL/Open-Llama/issues/5).
 
-### 数据读取
-数据读取相关代码可见dataset目录，其中包含根据下载的数据集使用SentencePiece训练分词模型，以及根据分词器构建DataLoader。
+### Related Tools
+
+In the utils directory, training tokenizer/supplementing existing tokenizer models and conversion checkpoint code are provided.
+
+Use SentencePiece to train a tokenizer with the following command:
 
-训练分词器使用如下命令
 ```bash
-python3 dataset/train_tokenizer.py
+python3 utils/train_tokenizer.py
 ```
 
-使用如下命令查看DataLoader输出的结果
+In configs, a tokenizer model with a 40k vocabulary, trained only using the Wudao dataset (4w_cn_vocab_wudao15.model), is provided.
+
+To supplement the vocabulary based on an existing tokenizer model, refer to:
+
 ```bash
-python3 dataset/pretrain_dataset.py
+python3 utils/merge_tokenizer.py
 ```
-### 模型结构
-我们基于Transformers库中的[Llama](https://github.com/facebookresearch/llama)参考论文原文中的2.4 Efficient implementation一节进行了修改，
-同时还参考了一些其他论文引入了一些优化。具体来说，我们引入了由META开源的[xformers库](https://github.com/facebookresearch/xformers)中的memory_efficient_attention操作来进行
-Self Attention的计算，这对于性能有明显的提升，提升大约30%。
-具体可以参见[modeling_llama.py](https://github.com/Bayes-Song/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L240)
 
-同时我们还参考了[Bloom](https://huggingface.co/bigscience/bloom)，对于Token Embedding引入了Stable Embedding以更好的稳定训练。
+A bilingual English and Chinese tokenizer model (llama_tokenizer_extended.model) is created by merging the META official tokenizer model with the 40k Chinese tokenizer mentioned above.
 
-最后我们参考[PALM](https://arxiv.org/abs/2204.02311)，使用了Shared Input-Output Embeddings。
+To convert existing Llama model checkpoints, refer to:
 
-### 预训练
-我们基于Accelerate库进行多GPU并行训练，启动命令如下
 ```bash
-accelerate launch --config_file configs/default_config.yaml pretrain_llama.py
+python3 utils/convert_ckpt.py
 ```
-某些情况下可能需要指定下列参数
+
+### Data Loading
+
+Data loading-related code can be found in dataset/dataset.py, which includes pre-training and instruction fine-tuning data processing. To add other datasets, only the transform function needs to be modified.
+
+The data loading process is as follows:
+
+1. Use the transform function to unify data formats from different datasets to {'text': 'xxx'}
+2. Tokenize using Tokenizer
+3. Sample long sequences; currently, three modes are provided: truncation, sampling (refer to the Gopher paper), and splitting
+4. Optional: concatenate texts from different docs, reducing padding in the data and accelerating training. In the v1 version, padding accounted for 30%; after concatenation, padding is reduced to 5%.
+
+Use the following command to view the output of DataLoader and check the correctness of tokenization:
+
+```bash
+python3 dataset/dataset.py
+```
+
+### Model Structure
+
+We modified according to the section 2.4 Efficient implementation of the [Llama](https://github.com/facebookresearch/llama) paper in the Transformers library, and also referenced other papers to introduce some optimizations. Specifically, we used the memory_efficient_attention operation from the [xformers library](https://github.com/facebookresearch/xformers) open-sourced by META for Self Attention computation, which has a significant performance improvement of approximately 30%. Further details can be found in [modeling_llama.py](https://github.com/huggingface/transformers/blob/main/src/transformers/models/open_llama/modeling_open_llama.py#L229).
+
+Additionally, we referred to [Bloom](https://huggingface.co/bigscience/bloom) and introduced Stable Embedding for Token Embedding to better stabilize training.
+
+Finally, we referenced [PALM](https://arxiv.org/abs/2204.02311) and employed Shared Input-Output Embeddings.
+
+### Pre-training
+
+We use multi-GPU parallel training based on the Accelerate library, with the following start command:
+
+```bash
+accelerate launch --config_file configs/accelerate_configs/ds_stage1.yaml train_lm.py --train_config configs/pretrain_config.yaml --model_config configs/model_configs/7B.json 
+```
+In some cases, you may need to specify the following parameters:
+
 ```
 --main_process_ip
 --main_process_port
@@ -143,64 +244,68 @@ accelerate launch --config_file configs/default_config.yaml pretrain_llama.py
 --num_machines
 --machine_rank
 ```
-我们使用[Wandb](https://wandb.ai/)进行训练的可视化，需要自行修改环境变量 WANDB_API_KEY 。
 
-其中我们使用了DeepSpeed stage1以减少显存占用。accelerate相关配置可见configs/default_config.yaml。
+We use [Wandb](https://wandb.ai/) for visualizing training. You need to modify the WANDB_API_KEY environment variable yourself.
 
-训练相关超参数可见configs/train_config.py，目前我们使用10W词表的7B Llama模型进行训练，具体配置如下
+Among them, we use DeepSpeed stage1 to reduce memory usage. For Accelerate-related configurations, see configs/accelerate_configs.
+
+Training related hyperparameters can be found in configs/pretrain_config.yaml.
+
+The default parameters use LlamaTokenizer with a supplemented 40k Chinese vocabulary tokenizer model, and the model size is 7B. The specific configuration is as follows:
 
 | max_length | batch_size | learning_rate | weight_decay | params | dimension | n heads | n layer | vocab_size |
 |------------|------------------|---------------|--------------|--------|-----------|---------|---------|------------|
-| 1024       | 2                | 2e-4          | 1e-1         | 6.88B  | 4096      | 32      | 32      | 100000     |
+| 2048       | 2                | 2e-4          | 1e-1         | 7.03B  | 4096      | 32      | 32      | 68762      |
 
 ```
-=========================================================================================================
-Layer (type:depth-idx)                                  Output Shape              Param #
-=========================================================================================================
-LlamaForCausalLM                                        [1, 64, 32, 128]          --
-├─LlamaModel: 1-1                                       [1, 64, 32, 128]          --
-│    └─Embedding: 2-1                                   [1, 64, 4096]             409,600,000
-│    └─LayerNorm: 2-2                                   [1, 64, 4096]             8,192
-│    └─ModuleList: 2-3                                  --                        --
-│    │    └─LlamaDecoderLayer: x32                      [1, 64, 4096]             202,383,360 x 32
-│    └─LlamaRMSNorm: 2-4                                [1, 64, 4096]             4,096
-=========================================================================================================
-Total params: 6,885,879,808
-Trainable params: 6,885,879,808
+==============================================================================================================
+Layer (type:depth-idx)                                       Output Shape              Param #
+==============================================================================================================
+OpenLlamaForCausalLM                                         [1, 32, 64, 128]          --
+├─OpenLlamaModel: 1-1                                        [1, 32, 64, 128]          --
+│    └─Embedding: 2-1                                        [1, 64, 4096]             281,649,152
+│    └─ModuleList: 2-2                                       --                        --
+│    │    └─OpenLlamaDecoderLayer: 3x32                      [1, 64, 4096]             202,383,360
+│    └─OpenLlamaRMSNorm: 2-3                                 [1, 64, 4096]             4,096
+├─Linear: 1-2                                                [1, 64, 68762]            281,649,152
+==============================================================================================================
+Total params: 7,039,569,920
+Trainable params: 7,039,569,920
 Non-trainable params: 0
-Total mult-adds (G): 6.89
+Total mult-adds (G): 7.04
 ```
 
-目前的进展
-![](assets/pretrain_loss.png)
+Pre-training loss from scratch is shown below:
+
+![loss](assets/pretrain_loss.png)
 
 ### Instruction-Tuning
 
-我们使用目前开源的三个数据集进行Instruction-tuning，后续会加入更多的任务以及自己构建的数据集。
+We use the currently available seven datasets for Instruction-tuning, and more tasks and our own datasets will be added later.
+
 - [yizhongw/self_instruct](https://huggingface.co/datasets/yizhongw/self_instruct)
 - [BelleGroup/train_0.5M_CN](https://huggingface.co/datasets/BelleGroup/train_0.5M_CN)
 - [BelleGroup/train_1M_CN](https://huggingface.co/datasets/BelleGroup/train_1M_CN)
 - [BelleGroup/multiturn_chat_0.8M](https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M)
 - [BelleGroup/school_math_0.25M](https://huggingface.co/datasets/BelleGroup/school_math_0.25M)
-- [RyokoAI/ShareGPT52K](https://huggingface.co/datasets/RyokoAI/ShareGPT52K)
+- [anon8231489123/ShareGPT_Vicuna_unfiltered](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered)
 - [Graverman/Instruct-to-Code](https://huggingface.co/datasets/Graverman/Instruct-to-Code)
 
-其中ShareGPT52K数据在datastes的处理有些问题，我们直接下载原数据重新进行了处理。
-我们对原始数据进行了一些预处理，格式如下
+The ShareGPT_Vicuna_unfiltered dataset has some issues in the datastes processing, so we directly downloaded the original data and reprocessed it.
+We performed some preprocessing on the original data, with the format as follows:
+
 ```
 user: {prompt}\nsystem: {completion}</s>
 ```
 
-具体训练代码和预训练基本一样，代码可见
-```
-instruction_tuning.py
+The startup command is basically the same as pre-training:
+
+```bash
+accelerate launch --config_file configs/accelerate_configs/ds_stage1.yaml train_lm.py --train_config configs/instruct_config.yaml --model_config configs/model_configs/7B.json 
 ```
 
-启动命令也基本一致
-```bash
-accelerate launch --config_file configs/default_config.yaml instruction_tuning.py
-```
-某些情况下可能需要指定下列参数
+In some cases, you may need to specify the following parameters:
+
 ```
 --main_process_ip
 --main_process_port
@@ -209,163 +314,89 @@ accelerate launch --config_file configs/default_config.yaml instruction_tuning.p
 --machine_rank
 ```
 
-过程中Loss如下，基本在波动不怎么下降
+The loss during the process is shown below, with a total of 3 epochs:
+
 ![loss](assets/instruct_loss.png)
+
 ### RLHF
-暂无
+
+Not available yet.
+
 ### Server
 
-单轮对话使用server.py，对于多轮对话使用chat_server.py
+For multi-turn dialogue, use chat_server.py.
 
-基于Gradio开发。 
-## 性能对比
+Developed based on Gradio.
 
-### 训练框架
-在训练框架方面我们测试了HuggingFace开源的Accelerate库和HPC-AI开源的ColossalAI，我们测试在打满显卡时性能差异较小。因此最终选择了实现相对简单的Accelerate库作为训练框架
+## Performance Comparison
 
-测试数据如下，测试过程中使用的模型结构为
+### Training Framework
+
+In terms of training frameworks, we tested Hugging Face's open-source Accelerate library, PyTorch Lightning, and HPC-AI's open-source ColossalAI. We found that their performance differences are relatively small when fully utilizing GPUs. Therefore, we chose the relatively simple-to-implement Accelerate library as the training framework.
+
+The test code can be found in utils/speed_test.py.
+
+The model structure used during the testing process is:
 | Model | n gpu | n layer | n heads | hidden size | vocab size | seq length |
 |-------|-------|---------|---------|-------------|------------|------------|
 | GPT2  | 2     | 6       | heads   | 4096        | 250100     | 1024       |
 
-测试结果如下，可以看到当打满时速度和显存相差不大
-|                 | HuggingFace                       | HuggingFace                        | ColossalAI                                             | ColossalAI                                             | ColossalAI                         |
+The test results are shown below, indicating that when the GPUs are fully utilized, the differences in speed and memory consumption are not significant.
+|                 | Hugging Face                       | Hugging Face                        | ColossalAI                                             | ColossalAI                                             | ColossalAI                         |
 |-----------------|-----------------------------------|------------------------------------|--------------------------------------------------------|--------------------------------------------------------|------------------------------------|
 | config          | without activation ckpt, bs2      | without activation ckpt, max_bs=12 | with activation ckpt, bs2                              | without activation ckpt, bs2                           | without activation ckpt, max_bs=10 |
 | second pre step | 0.336, fw=0.033, bw=0.3, opt=5e-6 | 1.25                               | 0.347                                                  | 0.308, fw=0.067, bw=0.152, opt=0.088                   | 1.055                              |
-| gpu memory      | nvidia-smi 45445                  |                                    | fw+bw+opt=21053.63+22064.12+17987.52, nvidia-smi 40961 | fw+bw+opt=24684.74+21087.13+17987.52, nvidia-smi 46821 | oom after 10 steps, 疑似有内存泄漏 |
+| gpu memory      | nvidia-smi 45445                  |                                    | fw+bw+opt=21053.63+22064.12+17987.52, nvidia-smi 40961 | fw+bw+opt=24684.74+21087.13+17987.52, nvidia-smi 46821 | oom after 10 steps |
 
-### 性能优化
-在最早版本中我们使用DeepSpeed stage2 + Transformers中的原生Llama实现进行训练但是速度和论文中所说的相差较大，因此后续我们进行了一系列的优化，我们将每一步的性能提升列在下面可供参考。
+### Performance Optimization
 
-论文中提到对于6.7B模型使用了1T token进行训练，最终的gpu时为82432，因此可以计算出他的训练速度大致为3370 token/s/gpu。
-当使用下面的优化后速度开源基本和论文中速度一致，使用20x8 A100-80G进行测试。预计加入更多融合算子开源取得更好的性能。
+In the earliest version, we used the native Llama implementation from DeepSpeed stage2 + Transformers for training. However, the speed was significantly different from what was claimed in the paper. Therefore, we carried out a series of optimizations afterwards, and we list each step of the performance improvement below for reference.
 
-|                     | V1           | V2                    |
-|---------------------|--------------|-----------------------|
-| Model               | Transformers | Transformers+xformers |
-| Optimizer           | Pytorch Adam | Fused Adam            |
-| DeepSpeed           | stage2       | stage1                |
-| Grad Accumulation   | 4            | 12                    |
-| Return Padding Mask | yes          | no                    |
-| Speed token/s/gpu   | 1378         | 3290                  |
+The paper mentioned that for the 6.7B model, 1T token was used for training and the final GPU time was 82432, from which the training speed was roughly calculated as 3370 token/s/gpu. After using the following optimizations, the speed is now basically consistent with what was claimed in the paper when tested on 20x8 A100-80G. It is expected that more fusion operators will be added in the future to achieve better performance.
 
-### 和其他开源模型性能对比
-下表是一个对目前开源模型性能的一个总结，使用GPU device均为A100，由于模型大小各不相同结构也有一定差异，难以准确的对比性能，作为一个粗略估计可以认为速度和模型参数量基本呈反比关系，这一点看Llama不同大小的模型可以得到印证。基于这个粗略估计可以看到使用本项目的性能明显由于其他项目。
+|                     | V1           | V2                                 |
+|---------------------|--------------|------------------------------------|
+| Dataset             | self implemented | datasets                       |
+| Model               | Transformers     | Transformers+xformers          |
+| Optimizer           | Pytorch Adam     | Fused Adam                     |
+| DeepSpeed           | stage2           | stage1                         |
+| Grad Accumulation   | 4                | 12                             |
+| Return Padding Mask | yes              | no                             |
+| Speed token/s/gpu   | 1378             | 3637                           |
+
+### Comparison with Other Open-source Models
+
+The following table summarizes the performance of currently available open-source models. In all cases, the GPU device used is A100. Due to differences in the size and structure of the models, it is difficult to make accurate performance comparisons. As a rough estimate, it can be assumed that the speed is generally inversely proportional to the size of the model parameters, which is confirmed by the performance of Llama with models of different sizes. Based on this rough estimate, it can be seen that the performance using our project is significantly better than that of other projects.
 
 | Model               | Open-Llama | LLAMA    | LLAMA   | LLAMA     | OPT     | Bloom              | GLM   | GPT-NEOX | CPM-ANT | CodeGeeX  |
 |---------------------|------------|----------|---------|-----------|---------|--------------------|-------|----------|---------|-----------|
-| Model size          | 6.9B       | 6.7B     | 13B     | 65B       | 175B    | 175B               | 130B  | 20B      | 10B     | 13B       |
+| Model size          | 7.0B       | 6.7B     | 13B     | 65B       | 175B    | 175B               | 130B  | 20B      | 10B     | 13B       |
 | Token               |            | 1T       | 1T      | 1.4T      | 180B    | 366B               | 400B  | 402B     | 200B    | 13.9B     |
 | GPU Hour            |            | 82,432   | 135,168 | 1,022,362 | 809,472 | 1,082,990          | 43776 | 175680   | 47040   | 3072      |
-| speed token/s/gpu   | 3290       | 3370     | 2055    | 380       | 61.8    | 93.9               | 105.7 | 635.6    | 1181    | 1257      |
-| 相关依赖            | xformers   | xformers |         |           | measeq  | Megatron-DeepSpeed |       |          | BMtrain | MindSpore |
-| speed token/s/gpu/B | 22701      | 22579    | 26715   | 24700     | 10815   | 16432              | 13741 | 12712    | 11810   | 16341     |
+| speed token/s/gpu   | 3637       | 3370     | 2055    | 380       | 61.8    | 93.9               | 105.7 | 635.6    | 1181    | 1257      |
+| 相关依赖             | xformers   | xformers |         |           | measeq  | Megatron-DeepSpeed |       |          | BMtrain | MindSpore |
+| speed token*params B/s/gpu | 25728      | 22579    | 26715   | 24700     | 10815   | 16432              | 13741 | 12712    | 11810   | 16341     |
 
-## 后续计划
+## Future Plans
 
-1. 加入更多训练监控，比如训练数据类别的分布等，加入继续训练相关代码
-2. 开源预训练好的多语言Llama 6.9B的checkpoint
-3. 实现Instruction-tuning代码，并开源相关checkpoint
-4. 使用Gradio搭建在线Demo
-5. 使用[Triton](https://github.com/openai/triton)加入更多高性能算子，进一步提升性能
-6. 加入根据Common Crawl构建预训练数据集相关代码，并开源相关数据集
-7. 加入多模态训练代码
+1. Integrate RLHF code.
+2. Use Triton to add more high-performance operators to further improve performance.
+3. Add code for building pre-training datasets based on Common Crawl and open related datasets.
+4. Add code for multimodal training.
 
-## 引用
+## Citation
 
 ```
 @misc{openllama,
   title={Open-Llama},
-  author={Liang Song},
+  author={s-JoL},
   year={2023},
-  howpublished={\url{https://github.com/Bayes-Song/Open-Llama}},
+  howpublished={\url{https://github.com/s-JoL/Open-Llama}},
 }
 ```
 
-<!-- 一些之前没注意到的部分
-
-1. [GPT3](https://arxiv.org/pdf/2005.14165.pdf), Details of Model Training
-
-During training we always train on sequences of the full nctx = 2048 token context window, packing multiple documents into a single sequence when documents are shorter than 2048, in order to increase computational efficiency. Sequences with multiple documents are not masked in any special way but instead documents within a sequence are delimited with a special end of text token, giving the language model the information necessary to infer that context separated by the end of text token is unrelated. This allows for efficient training without need for any special sequence-specific masking.
-
-在[PALM](https://arxiv.org/pdf/2204.02311.pdf)中也有类似的说法
-Sequence length – A sequence length of 2048 was used for all models. Input examples are concatenated together and then split into sequences of exactly 2048 tokens, so that there are no padding tokens, but examples may be split in the middle. Input examples are differentiated from one another with a special [eod] token.
-
-2. GPT3,  Common Crawl Filtering
-
-使用高质量文本作为正例，其他所有样本作为负例。根据判为正例的概率作为筛选np.random.pareto(α) > 1 − document_score。
-思想是尽量使用和高质量样本相似的数据。
-The classifier is trained using logistic regression classifier with features from Spark’s standard tokenizer and HashingTF.
-
-3. GPT3, fuzzy deduplication
-
-使用MinHashLSH进行去重，同时把CC中的WebText部分数据去掉。这些特征和分类器使用的一致。
-we fuzzily deduplicated documents (i.e. removed documents with high overlap with other documents) within each dataset using Spark’s MinHashLSH implementation with 10 hashes
-
-4. GPT3, Test Set Contamination
-
-5. [The pile](https://arxiv.org/pdf/2101.00027.pdf), BPB(bits per UTF-8 encoded byte)/bits per character/perplexity
-
-$
-BPB = = (L_T /L_B)l/ ln(2) \\
-perplexity = P(w1, w2, w3, w4, ...)^{-\frac{1}{N}} \\
-bpc=-\frac{1}{T}\sum_i log_2 P(w_i|w1, w2, ..., w_{i-1}) \\
-2^{bpc}=(\prod_i P(w_i|w1, w2, ..., w_{i-1}))^{-\frac{1}{T}}=perplexity
-$
-bpc是字符粒度，和分词算法相关。而bpb为byte粒度，与分词算法无关。
-
-可以使用bpb的差异衡量不同数据的难度。
-
-6. The pile, diversity of the collected data
-
-数据多样性
-
-We hypothesize that this is due to the perplexity based filtering used in CC-100, where a language model is trained on Wikipedia and all data with a perplexity too high or too low is discarded. This effectively discards any data too similar to or too different from Wikipedia, which severely limits the diversity of the collected data. 
-
-7. The pile, bytes per token
-
-Since the GPT-2 BPE tokenizer is trained on WebText, the mean bytes per token is also a very rough indicator of how syntactically different each Pile component is from WebText.
-
-8. The pile, Deduplication
-
-We used 10 hash functions for each Minhash and an approximate Jaccard similarity of 0.5.
-
-9. GLM, Embedding Layer Gradient Shrink 
-
-和stable embedding类似
-$
-word-embedding = word-embedding*\alpha+word-embedding.detach() ∗ (1−\alpha)
-$
-
-10. PALM, Training Instability
-
-训练中的loss尖峰是由特定的数据和特定的参数共同造成，使用模型回滚+跳过部分数据解决。
-Instead, we found that a simple strategy to effectively mitigate the issue: We re-started training from a checkpoint roughly 100 steps before the spike started, and skipped roughly 200–500 data batches, which cover the batches that were seen before and during the spike. With this mitigation, the loss did not spike again at the same point. We do not believe that the spikes were caused by “bad data” per se, because we ran several ablation experiments where we took the batches of data that were surrounding the spike, and then trained on those same data batches starting from a different, earlier checkpoint. In these cases, we did not see a spike. This implies that spikes only occur due to the combination of specific data batches with a particular model parameter state
-
-
-11. [Chinchilla](https://arxiv.org/pdf/2203.15556.pdf), Optimal model scaling
-
-20 tokens per parameter, for example 10B model should use 200B tokens to pretrain
-
-12. [Gopher](https://arxiv.org/pdf/2112.11446.pdf), Quality Filtering
-
-Quality Filtering (MassiveWeb only) The vast majority of text found on the web is of insufficient
-quality to be useful for language model training. For example, many web pages contain primarily
-automatically generated content, or text that is not intended for human consumption (such as keywords
-for search-engine optimisation). Much of the web also comprises social media content, which can
-variously lack context, coherence, or substance. To remove low-quality data while minimising potential
-for bias, we apply a number of simple, easily understood heuristic filters: we remove any document
-that does not contain between 50 and 100,000 words, or whose mean word length is outside the
-range of 3 to 10 characters; we remove any document with a symbol-to-word ratio greater than 0.1
-for either the hash symbol or the ellipsis; and we remove any document with more than 90% of lines
-starting with a bullet point, or more than 30% ending with an ellipsis. We also require that 80%
-of words in a document contain at least one alphabetic character, and apply a "stop word" filter, to
-remove documents that do not contain at least two of the following English words: the, be, to, of, and,
-that, have, with; this adequately deals with ostensibly English documents that contain no coherent
-English text.
-
-13. Gopher, Constructing Token Sequences
-
-和GPT3中的避免mask的方法类似
- -->
+<p align="center">
+  <a href="https://star-history.com/#s-JoL/Open-Llama&Date">
+    <img src="https://api.star-history.com/svg?repos=s-JoL/Open-Llama&type=Date" alt="Star History Chart">
+  </a>
+</p>

README_en.md

@@ -1,274 +0,0 @@
-<!--
- * @Author: LiangSong(sl12160010@gmail.com)
- * @Date: 2023-03-10 21:18:35
- * @LastEditors: LiangSong(sl12160010@gmail.com)
- * @LastEditTime: 2023-04-16 23:49:28
- * @FilePath: /Open-Llama/README_en.md
- * @Description: 
- * 
- * Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
--->
-# Open-Llama
-
-Translated by ChatGPT.
-
-Open-Llama is an open source project that provides a complete set of training processes for building large-scale language models, from data preparation to tokenization, pre-training, instruction tuning, and reinforcement learning techniques such as RLHF.
-
-## Progress
-
-The checkpoint after Instruct-tuning has been open-sourced on [HuggingFace](https://huggingface.co/s-JoL/Open-Llama-V1).
-To use the checkpoint, you need to first install the latest version of Transformers using the following command.
-``` base
-pip install git+https://github.com/s-JoL/transformers.git@dev
-```
-
-We completed pre-training on 300 billion tokens, with a total of 80,000 steps trained, using a global batch size of 4 million, consistent with Llama. We constructed the instruction-tuning dataset using a total of 7 parts of data, which the model has certain programming ability, mathematical ability, and multi-turn dialogue ability. For specific data, please refer to the instruction-tuning section.
-
-[Demo](http://home.ustc.edu.cn/~sl9292/)
-
-We tested our model by referring to some tests for Wenxin Yiyuan. Original report can be found at Baidu ["Wenxin Yiyan" Test: What is the level of domestic generative AI?](https://www.8btc.com/article/6809666)
-
-The results of our model are shown in the following figure, and more results are yet to be further tested. Due to domestic network problems, the use of the above Demo may result in a request loss situation. If there is no response for a long time, please refresh and try again.
-
-![image1](assets/eng1.png)![image2](assets/eng2.png)![image3](assets/eng3.png)
-
-Here is a demonstration of the model's ability in multi-turn dialogue about code.
-
-![image4](assets/multiturn_chat_en.jpeg)
-
-We roughly estimate the cost to achieve the above results. The 40K-step pre-training used 150 million pre-training data, which is about 110B tokens. The total training time is 76 hours, and the cost is about $19,152 according to Google Cloud's A100 quotation. The Instruction-tuning training was carried out for 12k steps, using 1.6 million data, and the total training time was 3.4 hours, costing about $342. Therefore, the total cost of training such a model from scratch is less than $20,000.
-
-Currently, the model's performance in both mathematical and code-related tasks is noticeably poor. This is partially due to the training data used, but I also believe it is due to the size of the model. However, the ability to perform logical reasoning is essential for any usable model. Therefore, future updates will focus on improving this aspect of the model's capabilities.
-## **Features**
-### Ease of Use
-We believe that ease of use is one of the most important features when building large-scale language models. To make Open-Llama more accessible, we focus on the following:
-
-- **Minimal implementation**: We use the simplest implementation approach to reduce the barrier to entry and make it easy for beginners to get started.
-- **Complete workflow**: We provide complete code from data set construction to training, making each step of building a large language model clear and visible.
-
-### High Performance
-Since training large language models is costly, high performance is also crucial when building large-scale language models. To achieve high-performance training, we employ the following techniques:
-
-- **Fused CUDA kernel**: Using fused CUDA kernels provided by [xformers](https://github.com/facebookresearch/xformers) can fuse multiple operations together, reducing data transfer between GPU and CPU, and improving training efficiency.
-- **Parallel training**: We use the [Accelerate](https://huggingface.co/docs/accelerate/index) library to support parallel training on multiple GPUs, accelerating the training process.
-
-
-For 7B mode, the training speed of the Llama model using the PyTorch native version in the Transformers library is 1378 tokens/s/GPU. With our code, the training speed reaches 3290 tokens/s/GPU, which is close to the reported 3370 tokens/s/GPU in the [Llama paper](https://arxiv.org/pdf/2302.13971.pdf).
-If we pretrain with 500 billion tokens, it will take 43,000 GPU hours. Assuming the price of A100-80G Spot on Google Cloud is $12.6 per hour for 8 GPUs, the total cost will be $67,725.
-Without acceleration, the cost would be $158,744. Our method reduces the training cost by $90,019 in total.
-More comparison can be found in [Comparison of Performance with Other Open-Source Models](https://github.com/Bayes-Song/Open-Llama/blob/main/README_en.md#performance-comparison-with-other-open-source-models).
-### Universality
-When training language models, we aim to build a universal model that can be used for different languages and fields. To achieve this, we adopt the following strategies:
-
-- **Multi-language support**: We support a variety of language corpora, including English, Chinese, Japanese, and other languages, allowing users to choose according to their needs.
-- **Field universality**: We hope that the model can not only help with everyday problems but also assist in professional fields such as science and law.
-## **Requirements**
-- Python 3.7 or higher
-- PyTorch 1.13
-- Customized [Transformers library](https://github.com/Bayes-Song/transformers)
-- [Accelerate library](https://huggingface.co/docs/accelerate/index)
-- CUDA 11.6 or higher version (for GPU acceleration, tested based on CUDA 11.7)
-## **Getting Started**
-### Installation
-Use the following command to install the required dependencies:
-
-```bash
-pip install -r requirements.txt
-```
-
-### Dataset Preparation
-Currently, we provide the Wudao dataset from ZhuiyiAI and The Pile dataset from EleutherAI. The code for downloading and processing the datasets can be found in the data directory. Please note that the Wudao dataset requires agreeing to some agreements before downloading, so you may need to modify the link in download_wudao.sh. [WuDao](https://data.baai.ac.cn/details/WuDaoCorporaText)
-
-Use the following commands to download and shard the data:
-
-```bash
-bash data/download_the_pile.sh
-bash data/download_wudao.sh
-```
-The data will be stored as small files with a maximum of 16,384 lines per file for efficient multi-processing training. The storage format is jsonl.zst compressed with zstd, resulting in a total data size of 519.5 GB and 16,466 files.
-
-The Pile dataset contains 210,607,728 rows of JSON lines, and the Wudao dataset contains 59,132,213 rows of JSON lines.
-
-The specific data format is as follows:
-
-```
-WuDao
-{'id': 1, 'dataType': '百科', 'title': 'some title', 'content': 'some content'}
-
-The Pile
-{'text': 'some text', 'meta': {'pile_set_name': 'Github'}}
-```
-
-Verification of data intergrity can be foud in this [issue]((https://github.com/s-JoL/Open-Llama/issues/5)
-
-### Data Loading
-The code for loading data can be found in the dataset directory, which includes training a tokenizer using SentencePiece and constructing a DataLoader based on the tokenizer.
-
-Train the tokenizer with the following command:
-
-```bash
-python3 dataset/train_tokenizer.py
-```
-Check the DataLoader output with the following command:
-
-```bash
-python3 dataset/pretrain_dataset.py
-```
-### Model Structure
-We modified the [Llama](https://github.com/facebookresearch/llama) model in the Transformers library based on section 2.4 "Efficient Implementation" in the original paper and introduced some optimizations from other papers. Specifically, we introduced the memory_efficient_attention operation from the [xformers library](https://github.com/facebookresearch/xformers) by META for computing self-attention, which significantly improves performance by about 30%. Please refer to modeling_llama.py for details.
-
-We also referred to Bloom for introducing stable embeddings for better training of token embeddings.
-
-Finally, we referred to PALM and used shared input-output embeddings.
-
-### Pretraining
-We use the Accelerate library for multi-GPU parallel training. Launch training with the following command:
-
-```bash
-accelerate launch --config_file configs/default_config.yaml pretrain_llama.py
-```
-In some cases, it may be necessary to specify the following parameters.
-```
---main_process_ip
---main_process_port
---num_processes
---num_machines
---machine_rank
-```
-We use [Wandb](https://wandb.ai/) for training visualization and you need to modify the environment variable WANDB_API_KEY.
-
-We use DeepSpeed stage 1 to reduce GPU memory usage. Accelerate-related configurations can be found in configs/default_config.yaml.
-
-The training-related hyperparameters can be found in configs/train_config.py. We currently train a 7B Llama model with a vocabulary size of 100,000, and the specific configuration is as follows:
-
-| max_length | batch_size | learning_rate | weight_decay | params | dimension | n heads | n layer | vocab_size |
-|------------|------------------|---------------|--------------|--------|-----------|---------|---------|------------|
-| 1024       | 2                | 2e-4          | 1e-1         | 6.88B  | 4096      | 32      | 32      | 100000     |
-
-```
-=========================================================================================================
-Layer (type:depth-idx)                                  Output Shape              Param #
-=========================================================================================================
-LlamaForCausalLM                                        [1, 64, 32, 128]          --
-├─LlamaModel: 1-1                                       [1, 64, 32, 128]          --
-│    └─Embedding: 2-1                                   [1, 64, 4096]             409,600,000
-│    └─LayerNorm: 2-2                                   [1, 64, 4096]             8,192
-│    └─ModuleList: 2-3                                  --                        --
-│    │    └─LlamaDecoderLayer: x32                      [1, 64, 4096]             202,383,360 x 32
-│    └─LlamaRMSNorm: 2-4                                [1, 64, 4096]             4,096
-=========================================================================================================
-Total params: 6,885,879,808
-Trainable params: 6,885,879,808
-Non-trainable params: 0
-Total mult-adds (G): 6.89
-```
-Current Progress
-![](assets/loss.png)
-### Instruction-Tuning
-
-We performed instruction-tuning on three currently available open-source datasets, and we plan to add more tasks and our own constructed datasets in the future.
-- [yizhongw/self_instruct](https://huggingface.co/datasets/yizhongw/self_instruct)
-- [BelleGroup/generated_train_0.5M_CN](https://huggingface.co/datasets/BelleGroup/generated_train_0.5M_CN)
-- [BelleGroup/generated_train_1M_CN](https://huggingface.co/datasets/BelleGroup/generated_train_1M_CN)
-- [BelleGroup/train_0.5M_CN](https://huggingface.co/datasets/BelleGroup/train_0.5M_CN)
-- [BelleGroup/train_1M_CN](https://huggingface.co/datasets/BelleGroup/train_1M_CN)
-- [BelleGroup/multiturn_chat_0.8M](https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M)
-- [BelleGroup/school_math_0.25M](https://huggingface.co/datasets/BelleGroup/school_math_0.25M)
-- [RyokoAI/ShareGPT52K](https://huggingface.co/datasets/RyokoAI/ShareGPT52K)
-- [Graverman/Instruct-to-Code](https://huggingface.co/datasets/Graverman/Instruct-to-Code)
-There were some issues with the handling of ShareGPT52K dataset in the processing of the datasets. We downloaded the original data again and reprocessed it.
-
-We did some preprocessing on the raw data, the format is as follows:
-```
-user: {prompt}\nsystem: {completion}</s>
-```
-The training code is similar to pre-training and can be seen in
-```
-instruction_tuning.py
-```
-
-The launch command is also similar to pre-training:
-```bash
-accelerate launch --config_file configs/default_config.yaml instruction_tuning.py
-```
-In some cases, the following parameters may need to be specified:
-```
---main_process_ip
---main_process_port
---num_processes
---num_machines
---machine_rank
-```
-
-The loss during the process is as follows, basically fluctuating and not decreasing much:
-![loss](assets/instruct_loss.png)
-### RLHF
-N/A
-
-### Server
-
-Use server.py for single-turn conversation, and chat_server.py for multi-turn conversation.
-
-Developed based on Gradio.
-## Performance Comparison
-
-### Training Framework
-In terms of the training framework, we tested the HuggingFace's Accelerate library and HPC-AI's ColossalAI, and found that there was little difference in performance when running on fully utilized GPUs. Therefore, we ultimately chose the relatively simple Accelerate library as our training framework.
-
-The test data is shown below, and the model structure used during testing is:
-
-| Model | n gpu | n layer | n heads | hidden size | vocab size | seq length |
-|-------|-------|---------|---------|-------------|------------|------------|
-| GPT2  | 2     | 6       | heads   | 4096        | 250100     | 1024       |
-
-The test results are shown below, and we can see that there is little difference in speed and memory utilization when running on fully utilized GPUs:
-
-|                 | HuggingFace                       | HuggingFace                        | ColossalAI                                             | ColossalAI                                             | ColossalAI                         |
-|-----------------|-----------------------------------|------------------------------------|--------------------------------------------------------|--------------------------------------------------------|------------------------------------|
-| config          | without activation ckpt, bs2      | without activation ckpt, max_bs=12 | with activation ckpt, bs2                              | without activation ckpt, bs2                           | without activation ckpt, max_bs=10 |
-| second pre step | 0.336, fw=0.033, bw=0.3, opt=5e-6 | 1.25                               | 0.347                                                  | 0.308, fw=0.067, bw=0.152, opt=0.088                   | 1.055                              |
-| gpu memory      | nvidia-smi 45445                  |                                    | fw+bw+opt=21053.63+22064.12+17987.52, nvidia-smi 40961 | fw+bw+opt=24684.74+21087.13+17987.52, nvidia-smi 46821 | oom after 10 steps, suspected memory leak |
-
-### Performance Optimization
-In the earliest version, we used DeepSpeed stage2 and the native Llama implementation in Transformers for training. However, the speed was significantly different from what was reported in the paper. Therefore, we conducted a series of optimizations and list the performance improvements for each step below.
-
-The paper mentions that they trained the 6.7B model with 1T tokens, and the GPU utilization was 82432, so the training speed was approximately 3370 tokens/s/GPU. After implementing the following optimizations, our speed is now comparable to that reported in the paper, using 20x8 A100-80G for testing. We expect to achieve better performance by adding more fusion operators in the future.
-
-|                     | V1           | V2                    |
-|---------------------|--------------|-----------------------|
-| Model               | Transformers | Transformers+xformers |
-| Optimizer           | Pytorch Adam | Fused Adam            |
-| DeepSpeed           | stage2       | stage1                |
-| Grad Accumulation   | 4            | 12                    |
-| Return Padding Mask | yes          | no                    |
-| Speed token/s/gpu   | 1378         | 3290                  |
-
-### Performance Comparison with Other Open-source Models
-The following table summarizes the performance of current open-source models, all tested on A100 GPUs. Due to differences in model sizes and structures, it is difficult to make accurate performance comparisons. As a rough estimate, it can be assumed that speed and model parameter count are inversely proportional, as evidenced by Llama models of different sizes. Based on this rough estimate, it can be seen that the performance using our project is significantly better than other projects.
-
-| Model               | Open-Llama | LLAMA    | LLAMA   | LLAMA     | OPT     | Bloom              | GLM   | GPT-NEOX | CPM-ANT | CodeGeeX  |
-|---------------------|------------|----------|---------|-----------|---------|--------------------|-------|----------|---------|-----------|
-| Model size          | 6.9B       | 6.7B     | 13B     | 65B       | 175B    | 175B               | 130B  | 20B      | 10B     | 13B       |
-| Token               |            | 1T       | 1T      | 1.4T      | 180B    | 366B               | 400B  | 402B     | 200B    | 13.9B     |
-| GPU Hour            |            | 82,432   | 135,168 | 1,022,362 | 809,472 | 1,082,990          | 43776 | 175680   | 47040   | 3072      |
-| speed token/s/gpu   | 3290       | 3370     | 2055    | 380       | 61.8    | 93.9               | 105.7 | 635.6    | 1181    | 1257      |
-| Dependencies        | xformers   | xformers |         |           | measeq  | Megatron-DeepSpeed |       |          | BMtrain | MindSpore |
-| speed token/s/gpu/B | 22701      | 22579    | 26715   | 24700     | 10815   | 16432              | 13741 | 12712    | 11810   | 16341     |
-
-## Future Plans
-1. Add more training monitoring, such as the distribution of training data categories, and add code for continuing training.
-2. Realease the pre-trained checkpoint for the multi-lingual Llama 6.9B model.
-3. Implement instruction-tuning code and open-source related checkpoints.
-Build an online demo using Gradio.
-4. Use [Triton](https://github.com/openai/triton) to add more high-performance operators and further improve performance.
-5. Add code for building pre-training datasets based on Common Crawl and open-source related datasets.
-6. Add code for multi-modal training.
-## Citation
-```
-@misc{openllama,
-  title={Open-Llama},
-  author={Liang Song},
-  year={2023},
-  howpublished={\url{https://github.com/Bayes-Song/Open-Llama}},
-}
-```

README_zh.md

@@ -0,0 +1,464 @@
+<!--
+ * @Author: s-JoL(sl12160010@gmail.com)
+ * @Date: 2023-03-10 21:18:35
+ * @LastEditors: s-JoL(sl12160010@gmail.com)
+ * @LastEditTime: 2023-05-17 22:43:46
+ * @FilePath: /Open-Llama/README_zh.md
+ * @Description: 
+ * 
+ * Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
+-->
+[**中文**](./README_zh.md) | [**English**](./README.md)
+
+![camel](assets/logo.png)
+
+# Open-Llama
+
+<p align="center">
+    <img alt="GitHub" src="https://img.shields.io/github/license/s-JoL/Open-Llama.svg?color=blue&style=flat-square">
+    <img alt="GitHub release (latest by date)" src="https://img.shields.io/github/v/release/s-JoL/Open-Llama">
+    <img alt="GitHub top language" src="https://img.shields.io/github/languages/top/s-JoL/Open-Llama">
+    <img alt="GitHub last commit" src="https://img.shields.io/github/last-commit/s-JoL/Open-Llama">
+</p>
+
+Open-Llama是一个开源项目，提供了一整套用于构建大型语言模型的训练流程，从数据集准备到分词、预训练、指令调优，lora, 以及强化学习技术 RLHF。
+
+**可从[Demo](http://home.ustc.edu.cn/~sl9292/)直接试用本模型。**
+
+加入[discord](https://discord.gg/TrKxrTpnab)一起讨论大语言模型的发展。
+
+## **主要内容**
+
+- **支持Transformers/Hugging Face直接调用。** 经过Instruct-tuning的CheckPoint已开源在[Hugging Face: s-JoL/Open-Llama-V2](https://huggingface.co/s-JoL/Open-Llama-V2)。
+
+- **采用FastChat项目相同方法测评Open-Llama的效果和GPT3.5的效果对比，经过测试在中文问题上可以达到GPT3.5 89%的水平。**
+
+- **训练速度达到3620 token/s，快于Llama原文中的3370 token/s，达到目前sota的水平。**
+
+
+``` python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("s-JoL/Open-Llama-V2", use_fast=False)
+model = AutoModelForCausalLM.from_pretrained("s-JoL/Open-Llama-V2", device_map="auto")
+
+inputs = tokenizer('user:implement quick sort in python\nsystem:', return_tensors='pt', return_attention_mask=False, add_special_tokens=False)
+for k, v in inputs.items():
+   inputs[k] = v.cuda()
+pred = model.generate(**inputs, max_new_tokens=512, do_sample=True)
+print(tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))
+
+```
+只经过预训练的CheckPoint也上传至[s-JoL/Open-Llama-V2-pretrain](https://huggingface.co/s-JoL/Open-Llama-V2-pretrain)。
+
+我们完成了330B token的预训练，总共训练80 K step，Global Batch Size和Llama中一致为4M。
+使用总共7部分数据构成Instruction-tuning数据，模型具有一定的编程能力、数学能力和多轮对话能力，具体数据见Instruction-Tuning部分。
+
+如下是一个关于代码的多轮对话能力的展示
+
+![image4](assets/multiturn_chat.jpeg)
+
+## **更新**
+
+**[2023.5.8] Release v2.1**
+
+- 本次更新加入对更大模型训练的支持，使用DeepSpeed stage3 + offload + activation checkpoint可以在**A100-80G训练65B模型**。
+
+- 引入peft库**支持lora**等训练。
+
+- 下表对比了Open-Llama和Llama原文的训练速度，Llama性能数据引自Llama原文。
+
+
+|                | DeepSpeed Stage | Offload | Activation Checkpoint | Total Token | GPU hours | Speed token/s/gpu | Batch Size |
+|----------------|-----------------|---------|-----------------------|-------------|-----------|-------------------|------------|
+| Open-Llama 7B  | 1               | False   | False                 | 173.7B      | 13412     | 3620              | 2          |
+| Open-Llama 13B | 3               | False   | True                  | -           | -         | 1856              | 24         |
+| Open-Llama 33B | 3               | False   | True                  | -           | -         | 708               | 12         |
+| Open-Llama 65B | 3               | True    | True                  | -           | -         | 369               | 12         |
+| Llama 7B       | -               | -       | -                     | 1T          | 82432     | 3370              | -          |
+| Llama 13B      | -               | -       | -                     | 1T          | 135168    | 2055              | -          |
+| Llama 33B      | -               | -       | -                     | 1.4T        | 530432    | 733               | -          |
+| Llama 65B      | -               | -       | -                     | 1.4T        | 1022362   | 380               | -          |
+
+**[2023.4.28] Release v2.0**
+
+本次更新主要包含以下几个方面，相对于v1版本提升有效训练速度**50%**，其中pad从**30%**减少至**5%**，训练速度从**3200token/s**提升至**3620token/s**。0.95 * 3620/(0.7 * 3200)=1.521
+1. 使用Hugging Face的datasets库进行数据读取，具体流程如下
+   1. 使用transform函数将不同数据集的数据统一格式为{'text': 'xxx'}
+   2. 使用Tokenizer进行分词
+   3. 对长序列进行采样，目前提供三种模式，分别是：截断/采样（参考[Gopher论文](https://arxiv.org/abs/2112.11446)）/切分
+   4. 可选：对来自不同doc的文本进行拼接。减少了数据中的pad，加速训练；在v1版本中pad占比为**30%**，使用拼接后pad占比降低为**5%**。
+2. 加入Trainer，对于预训练和指令微调都可以复用，见solver/trainer.py
+3. 统一预训练和指令微调训练入口为train_lm.py
+4. 提供更方便的配置，可见configs/pretrain_config.yaml
+5. 提供基于其他预训练模型补充词表，继续预训练功能
+6. 支持从中断点继续训练，包括加载优化器参数/学习率和跳过重复数据
+
+[2023.4.16] Release v1.0
+
+提供基础的预训练和指令微调代码，训练速度达到Llama原文速度。预训练和指令微调后的模型已经开源在Hugging Face。
+
+v1版代码可见https://github.com/s-JoL/Open-Llama/tree/v1.0
+
+## **特性**
+
+### 易用性
+
+我们认为易用性是构建大型语言模型时最重要的特性之一。为了使 Open-LLAMA 更加易于使用，我们特别注重了以下几点：
+
+- **最简实现**：我们采用了最简单的实现方式，降低了入门的门槛，让初学者也能轻松上手。
+- **流程完整**：我们发布了从数据集构建到训练的完整代码，使得构建一个大语言模型的每一步流程都清晰可见。
+
+### 高性能
+
+由于训练大语言模型的成本高昂，因此在构建大型语言模型时，高性能也是非常重要的。为了实现高性能的训练，我们发布使用了以下技术：
+
+- **Fused CUDA kernel**：使用[xformers](https://github.com/facebookresearch/xformers)中提供的 fused CUDA kernel 可以将多个操作融合在一起，减少了 GPU 和 CPU 之间的数据传输，从而提高了训练效率。
+- **并行化训练**：我们使用[Accelerate](https://huggingface.co/docs/accelerate/index)库支持在多个 GPU 上进行并行化训练，以加快训练速度。
+
+对于7B模型，使用Transformers中Pytorch原生版本的Llama模型训练训练速度为**1378 token/s/gpu**，使用本代码库训练速度达到**3626 token/s/gpu**，超过[Llama原文](https://arxiv.org/pdf/2302.13971.pdf)中的**3370 token/s/gpu**。
+
+如果使用500B token进行预训练，需要训练38300 GPU时。按照Google Cloud上A100-80G Spot的价格计算，8卡每小时价格为12.6美元，则总价格为60300美元。
+当使用未加速版本训练时，价格为158744美元。最终降低训练成本9.8万美元。
+
+更多测试可见[和其他开源模型性能对比](https://github.com/s-JoL/Open-Llama#%E5%92%8C%E5%85%B6%E4%BB%96%E5%BC%80%E6%BA%90%E6%A8%A1%E5%9E%8B%E6%80%A7%E8%83%BD%E5%AF%B9%E6%AF%94)。
+### 通用性
+
+在训练语言模型时，我们希望能够构建一个通用的模型，可以适用于不同的语言和不同的领域。为了实现这一点，我们采用了以下策略：
+
+- **多语言支持**：我们支持多种语言的语料库，包括英语、中文、日语等多种语言，让用户可以根据自己的需求进行选择。
+- **领域通用性**：我们希望模型不仅能在日常问题上能产生帮助，同时希望在专业领域如科学、法律等也能帮助人类。
+- **和世界交互**：希望通过加入RL使得模型具备和世界交互的能力
+
+## **要求**
+
+- Python 3.7 或更高版本
+- PyTorch 1.13
+- [Transformers库](https://github.com/huggingface/transformers)
+- [Accelerate库](https://huggingface.co/docs/accelerate/index)
+- CUDA 11.6 或更高版本（用于 GPU 加速）
+- 硬件配置：目前使用(64 CPU, 1000G Memory, 8xA100-80G) x N，有个比较神奇的现象当使用更多cpu时反而会慢一点，猜测这和dataloader的多进程有一定关系。
+
+## **入门指南**
+### 安装
+
+使用下面的命令安装相关依赖
+
+```bash
+pip install -r requirements.txt
+```
+
+### 数据集准备
+
+目前给出了智源开源的悟道数据集和EleutherAI开源的the pile数据集。数据集下载和处理代码在data目录下。
+其中悟道数据集由于需要同意一些协议才能下载因此可能需要修改一下download_wudao中的链接，[悟道](https://data.baai.ac.cn/details/WuDaoCorporaText)。
+
+感谢[@skepsun](https://github.com/skepsun)的建议，使用scidb下载wudao数据集不需要登陆，并且下载更稳定一些。https://github.com/s-JoL/Open-Llama/issues/42
+
+**注意数据下载可能出现失败，建议将script中的下载和处理分成两部分来运行，可以将下载多运行机会，会自动断点续传。**
+
+运行下面的命令进行数据下载并进行分片
+```bash
+bash data/download_the_pile.sh
+bash data/download_wudao.sh
+```
+数据将按照每个文件最大16384行存储为小文件，便于后续使用多进程训练时进行读取。存储格式为jsonl.zst，使用zstd进行压缩，最终数据大小为519.5G，合计16466个文件。
+
+其中the pile数据集包含210607728行json line，悟道数据集包含59132213行json line。
+
+具体数据格式如下
+```
+WuDao
+{'id': 1, 'dataType': '百科', 'title': 'some title', 'content': 'some content'}
+
+The Pile
+{'text': 'some text', 'meta': {'pile_set_name': 'Github'}}
+```
+验证数据完整性可见 [issue](https://github.com/s-JoL/Open-Llama/issues/5)
+
+### 相关工具
+在utils目录中提供了训练分词/补充现有分词模型和转换ckpt的代码。
+
+使用SentencePiece训练分词器参考如下命令
+```bash
+python3 utils/train_tokenizer.py
+```
+在configs中提供了只使用wudao数据集训练的4w词表的分词模型 4w_cn_vocab_wudao15.model
+
+根据已有分词模型补充词表参考
+```bash
+python3 utils/merge_tokenizer.py
+```
+根据META官方的分词模型和上面的4w中文合并为中英文双语的分词模型 llama_tokenizer_extended.model
+
+转换现有的Llama模型ckpt参考
+```bash
+python3 utils/convert_ckpt.py
+```
+
+### 数据读取
+数据读取相关代码可见dataset/dataset.py，包含了预训练和指令微调数据的处理，如需加入其他数据集只需要修改其中的transform函数。
+
+数据读取流程如下：
+1. 使用transform函数将不同数据集的数据统一格式为{'text': 'xxx'}
+2. 使用Tokenizer进行分词
+3. 对长序列进行采样，目前提供三种模式，分别是：截断/采样（参考Gopher论文）/切分
+4. 可选：对来自不同doc的文本进行拼接。减少了数据中的pad，加速训练；在v1版本中pad占比为30%，使用拼接后pad占比降低为5%。
+
+使用如下命令查看DataLoader输出的结果，并检查分词正确性
+```bash
+python3 dataset/dataset.py
+```
+
+### 模型结构
+我们基于Transformers库中的[Llama](https://github.com/facebookresearch/llama)参考论文原文中的2.4 Efficient implementation一节进行了修改，
+同时还参考了一些其他论文引入了一些优化。具体来说，我们引入了由META开源的[xformers库](https://github.com/facebookresearch/xformers)中的memory_efficient_attention操作来进行
+Self Attention的计算，这对于性能有明显的提升，提升大约30%。
+具体可以参见[modeling_llama.py](https://github.com/huggingface/transformers/blob/main/src/transformers/models/open_llama/modeling_open_llama.py#L229)
+
+同时我们还参考了[Bloom](https://huggingface.co/bigscience/bloom)，对于Token Embedding引入了Stable Embedding以更好的稳定训练。
+
+最后我们参考[PALM](https://arxiv.org/abs/2204.02311)，使用了Shared Input-Output Embeddings。
+
+### 预训练
+我们基于Accelerate库进行多GPU并行训练，启动命令如下
+```bash
+accelerate launch --config_file configs/accelerate_configs/ds_stage1.yaml train_lm.py --train_config configs/pretrain_config.yaml --model_config configs/model_configs/7B.json 
+```
+某些情况下可能需要指定下列参数
+```
+--main_process_ip
+--main_process_port
+--num_processes
+--num_machines
+--machine_rank
+```
+我们使用[Wandb](https://wandb.ai/)进行训练的可视化，需要自行修改环境变量 WANDB_API_KEY 。
+
+其中我们使用了DeepSpeed stage1以减少显存占用。accelerate相关配置可见configs/accelerate_configs。
+
+训练相关超参数可见configs/pretrain_config.yaml
+
+其中默认参数为使用LlamaTokenizer补充4w中文的词表的分词模型，模型大小为7B，具体配置如下
+
+| max_length | batch_size | learning_rate | weight_decay | params | dimension | n heads | n layer | vocab_size |
+|------------|------------------|---------------|--------------|--------|-----------|---------|---------|------------|
+| 2048       | 2                | 2e-4          | 1e-1         | 7.03B  | 4096      | 32      | 32      | 68762      |
+
+```
+==============================================================================================================
+Layer (type:depth-idx)                                       Output Shape              Param #
+==============================================================================================================
+OpenLlamaForCausalLM                                         [1, 32, 64, 128]          --
+├─OpenLlamaModel: 1-1                                        [1, 32, 64, 128]          --
+│    └─Embedding: 2-1                                        [1, 64, 4096]             281,649,152
+│    └─ModuleList: 2-2                                       --                        --
+│    │    └─OpenLlamaDecoderLayer: 3x32                      [1, 64, 4096]             202,383,360
+│    └─OpenLlamaRMSNorm: 2-3                                 [1, 64, 4096]             4,096
+├─Linear: 1-2                                                [1, 64, 68762]            281,649,152
+==============================================================================================================
+Total params: 7,039,569,920
+Trainable params: 7,039,569,920
+Non-trainable params: 0
+Total mult-adds (G): 7.04
+```
+
+从头预训练Loss如下
+![](assets/pretrain_loss.png)
+
+### Instruction-Tuning
+
+我们使用目前开源的七个数据集进行Instruction-tuning，后续会加入更多的任务以及自己构建的数据集。
+- [yizhongw/self_instruct](https://huggingface.co/datasets/yizhongw/self_instruct)
+- [BelleGroup/train_0.5M_CN](https://huggingface.co/datasets/BelleGroup/train_0.5M_CN)
+- [BelleGroup/train_1M_CN](https://huggingface.co/datasets/BelleGroup/train_1M_CN)
+- [BelleGroup/multiturn_chat_0.8M](https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M)
+- [BelleGroup/school_math_0.25M](https://huggingface.co/datasets/BelleGroup/school_math_0.25M)
+- [anon8231489123/ShareGPT_Vicuna_unfiltered](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered)
+- [Graverman/Instruct-to-Code](https://huggingface.co/datasets/Graverman/Instruct-to-Code)
+
+其中ShareGPT_Vicuna_unfiltered数据在datastes的处理有些问题，我们直接下载原数据重新进行了处理。
+我们对原始数据进行了一些预处理，格式如下
+```
+user: {prompt}\nsystem: {completion}</s>
+```
+
+启动命令和预训练基本一致
+```bash
+accelerate launch --config_file configs/accelerate_configs/ds_stage1.yaml train_lm.py --train_config configs/instruct_config.yaml --model_config configs/model_configs/7B.json 
+```
+某些情况下可能需要指定下列参数
+```
+--main_process_ip
+--main_process_port
+--num_processes
+--num_machines
+--machine_rank
+```
+
+过程中Loss如下，总计使用3个epoch
+![loss](assets/instruct_loss.png)
+### RLHF
+暂无
+### Server
+
+多轮对话使用chat_server.py
+
+基于Gradio开发。 
+## 性能对比
+
+### 训练框架
+在训练框架方面我们测试了Hugging Face开源的Accelerate库pytorch-lightning和HPC-AI开源的ColossalAI，我们测试在打满显卡时性能差异较小。因此最终选择了实现相对简单的Accelerate库作为训练框架
+
+测试代码可见utils/speed_test.py
+
+测试过程中使用的模型结构为
+| Model | n gpu | n layer | n heads | hidden size | vocab size | seq length |
+|-------|-------|---------|---------|-------------|------------|------------|
+| GPT2  | 2     | 6       | heads   | 4096        | 250100     | 1024       |
+
+测试结果如下，可以看到当打满时速度和显存相差不大
+|                 | Hugging Face                       | Hugging Face                        | ColossalAI                                             | ColossalAI                                             | ColossalAI                         |
+|-----------------|-----------------------------------|------------------------------------|--------------------------------------------------------|--------------------------------------------------------|------------------------------------|
+| config          | without activation ckpt, bs2      | without activation ckpt, max_bs=12 | with activation ckpt, bs2                              | without activation ckpt, bs2                           | without activation ckpt, max_bs=10 |
+| second pre step | 0.336, fw=0.033, bw=0.3, opt=5e-6 | 1.25                               | 0.347                                                  | 0.308, fw=0.067, bw=0.152, opt=0.088                   | 1.055                              |
+| gpu memory      | nvidia-smi 45445                  |                                    | fw+bw+opt=21053.63+22064.12+17987.52, nvidia-smi 40961 | fw+bw+opt=24684.74+21087.13+17987.52, nvidia-smi 46821 | oom after 10 steps, 疑似有内存泄漏 |
+
+### 性能优化
+在最早版本中我们使用DeepSpeed stage2 + Transformers中的原生Llama实现进行训练但是速度和论文中所说的相差较大，因此后续我们进行了一系列的优化，我们将每一步的性能提升列在下面可供参考。
+
+论文中提到对于6.7B模型使用了1T token进行训练，最终的gpu时为82432，因此可以计算出他的训练速度大致为3370 token/s/gpu。
+当使用下面的优化后速度开源基本和论文中速度一致，使用20x8 A100-80G进行测试。预计加入更多融合算子开源取得更好的性能。
+
+|                     | V1           | V2                                 |
+|---------------------|--------------|------------------------------------|
+| Dataset             | self implemented | datasets                       |
+| Model               | Transformers     | Transformers+xformers          |
+| Optimizer           | Pytorch Adam     | Fused Adam                     |
+| DeepSpeed           | stage2           | stage1                         |
+| Grad Accumulation   | 4                | 12                             |
+| Return Padding Mask | yes              | no                             |
+| Speed token/s/gpu   | 1378             | 3637                           |
+
+### 和其他开源模型性能对比
+下表是一个对目前开源模型性能的一个总结，使用GPU device均为A100，由于模型大小各不相同结构也有一定差异，难以准确的对比性能，作为一个粗略估计可以认为速度和模型参数量基本呈反比关系，这一点看Llama不同大小的模型可以得到印证。基于这个粗略估计可以看到使用本项目的性能明显由于其他项目。
+
+| Model               | Open-Llama | LLAMA    | LLAMA   | LLAMA     | OPT     | Bloom              | GLM   | GPT-NEOX | CPM-ANT | CodeGeeX  |
+|---------------------|------------|----------|---------|-----------|---------|--------------------|-------|----------|---------|-----------|
+| Model size          | 7.0B       | 6.7B     | 13B     | 65B       | 175B    | 175B               | 130B  | 20B      | 10B     | 13B       |
+| Token               |            | 1T       | 1T      | 1.4T      | 180B    | 366B               | 400B  | 402B     | 200B    | 13.9B     |
+| GPU Hour            |            | 82,432   | 135,168 | 1,022,362 | 809,472 | 1,082,990          | 43776 | 175680   | 47040   | 3072      |
+| speed token/s/gpu   | 3637       | 3370     | 2055    | 380       | 61.8    | 93.9               | 105.7 | 635.6    | 1181    | 1257      |
+| 相关依赖             | xformers   | xformers |         |           | measeq  | Megatron-DeepSpeed |       |          | BMtrain | MindSpore |
+| speed token*params B/s/gpu | 25728      | 22579    | 26715   | 24700     | 10815   | 16432              | 13741 | 12712    | 11810   | 16341     |
+
+## 后续计划
+
+1. 加入RLHF代码
+2. 使用[Triton](https://github.com/openai/triton)加入更多高性能算子，进一步提升性能
+3. 加入根据Common Crawl构建预训练数据集相关代码，并开源相关数据集
+4. 加入多模态训练代码
+
+## 引用
+
+```
+@misc{openllama,
+  title={Open-Llama},
+  author={s-JoL},
+  year={2023},
+  howpublished={\url{https://github.com/s-JoL/Open-Llama}},
+}
+```
+
+<p align="center">
+  <a href="https://star-history.com/#s-JoL/Open-Llama&Date">
+    <img src="https://api.star-history.com/svg?repos=s-JoL/Open-Llama&type=Date" alt="Star History Chart">
+  </a>
+</p>
+
+<!-- 一些之前没注意到的部分
+
+1. [GPT3](https://arxiv.org/pdf/2005.14165.pdf), Details of Model Training
+
+During training we always train on sequences of the full nctx = 2048 token context window, packing multiple documents into a single sequence when documents are shorter than 2048, in order to increase computational efficiency. Sequences with multiple documents are not masked in any special way but instead documents within a sequence are delimited with a special end of text token, giving the language model the information necessary to infer that context separated by the end of text token is unrelated. This allows for efficient training without need for any special sequence-specific masking.
+
+在[PALM](https://arxiv.org/pdf/2204.02311.pdf)中也有类似的说法
+Sequence length – A sequence length of 2048 was used for all models. Input examples are concatenated together and then split into sequences of exactly 2048 tokens, so that there are no padding tokens, but examples may be split in the middle. Input examples are differentiated from one another with a special [eod] token.
+
+2. GPT3,  Common Crawl Filtering
+
+使用高质量文本作为正例，其他所有样本作为负例。根据判为正例的概率作为筛选np.random.pareto(α) > 1 − document_score。
+思想是尽量使用和高质量样本相似的数据。
+The classifier is trained using logistic regression classifier with features from Spark’s standard tokenizer and HashingTF.
+
+3. GPT3, fuzzy deduplication
+
+使用MinHashLSH进行去重，同时把CC中的WebText部分数据去掉。这些特征和分类器使用的一致。
+we fuzzily deduplicated documents (i.e. removed documents with high overlap with other documents) within each dataset using Spark’s MinHashLSH implementation with 10 hashes
+
+4. GPT3, Test Set Contamination
+
+5. [The pile](https://arxiv.org/pdf/2101.00027.pdf), BPB(bits per UTF-8 encoded byte)/bits per character/perplexity
+
+$
+BPB = = (L_T /L_B)l/ ln(2) \\
+perplexity = P(w1, w2, w3, w4, ...)^{-\frac{1}{N}} \\
+bpc=-\frac{1}{T}\sum_i log_2 P(w_i|w1, w2, ..., w_{i-1}) \\
+2^{bpc}=(\prod_i P(w_i|w1, w2, ..., w_{i-1}))^{-\frac{1}{T}}=perplexity
+$
+bpc是字符粒度，和分词算法相关。而bpb为byte粒度，与分词算法无关。
+
+可以使用bpb的差异衡量不同数据的难度。
+
+6. The pile, diversity of the collected data
+
+数据多样性
+
+We hypothesize that this is due to the perplexity based filtering used in CC-100, where a language model is trained on Wikipedia and all data with a perplexity too high or too low is discarded. This effectively discards any data too similar to or too different from Wikipedia, which severely limits the diversity of the collected data. 
+
+7. The pile, bytes per token
+
+Since the GPT-2 BPE tokenizer is trained on WebText, the mean bytes per token is also a very rough indicator of how syntactically different each Pile component is from WebText.
+
+8. The pile, Deduplication
+
+We used 10 hash functions for each Minhash and an approximate Jaccard similarity of 0.5.
+
+9. GLM, Embedding Layer Gradient Shrink 
+
+和stable embedding类似
+$
+word-embedding = word-embedding*\alpha+word-embedding.detach() ∗ (1−\alpha)
+$
+
+10. PALM, Training Instability
+
+训练中的loss尖峰是由特定的数据和特定的参数共同造成，使用模型回滚+跳过部分数据解决。
+Instead, we found that a simple strategy to effectively mitigate the issue: We re-started training from a checkpoint roughly 100 steps before the spike started, and skipped roughly 200–500 data batches, which cover the batches that were seen before and during the spike. With this mitigation, the loss did not spike again at the same point. We do not believe that the spikes were caused by “bad data” per se, because we ran several ablation experiments where we took the batches of data that were surrounding the spike, and then trained on those same data batches starting from a different, earlier checkpoint. In these cases, we did not see a spike. This implies that spikes only occur due to the combination of specific data batches with a particular model parameter state
+
+
+11. [Chinchilla](https://arxiv.org/pdf/2203.15556.pdf), Optimal model scaling
+
+20 tokens per parameter, for example 10B model should use 200B tokens to pretrain
+
+12. [Gopher](https://arxiv.org/pdf/2112.11446.pdf), Quality Filtering
+
+Quality Filtering (MassiveWeb only) The vast majority of text found on the web is of insufficient
+quality to be useful for language model training. For example, many web pages contain primarily
+automatically generated content, or text that is not intended for human consumption (such as keywords
+for search-engine optimisation). Much of the web also comprises social media content, which can
+variously lack context, coherence, or substance. To remove low-quality data while minimising potential
+for bias, we apply a number of simple, easily understood heuristic filters: we remove any document
+that does not contain between 50 and 100,000 words, or whose mean word length is outside the
+range of 3 to 10 characters; we remove any document with a symbol-to-word ratio greater than 0.1
+for either the hash symbol or the ellipsis; and we remove any document with more than 90% of lines
+starting with a bullet point, or more than 30% ending with an ellipsis. We also require that 80%
+of words in a document contain at least one alphabetic character, and apply a "stop word" filter, to
+remove documents that do not contain at least two of the following English words: the, be, to, of, and,
+that, have, with; this adequately deals with ostensibly English documents that contain no coherent
+English text.
+
+13. Gopher, Constructing Token Sequences
+
+和GPT3中的避免mask的方法类似
+ -->

chat_server.py

@@ -1,64 +1,30 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-04-06 22:30:10
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-07 23:03:31
+LastEditors: s-JoL(sl12160010@gmail.com)
+LastEditTime: 2023-05-12 15:07:36
 FilePath: /Open-Llama/chat_server.py
 Description: 
 
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 import torch
+import logging
 import gradio as gr
-import sentencepiece as spm
-from dataset.tokenizer import Tokenizer
-from transformers import LlamaForCausalLM, LlamaConfig
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
 
-sp_model = spm.SentencePieceProcessor(
-    model_file="configs/10w_vocab_wudao5_pile10.model"
+tokenizer = AutoTokenizer.from_pretrained("s-JoL/Open-Llama-V2", use_fast=False)
+model = AutoModelForCausalLM.from_pretrained(
+    "s-JoL/Open-Llama-V2", torch_dtype=torch.bfloat16, device_map="auto"
 )
-tokenizer = Tokenizer(sp_model)
-raw_model = LlamaForCausalLM(
-    LlamaConfig(
-        vocab_size=tokenizer.vocab_size,
-        initializer_range=0.01,
-        pad_token_id=tokenizer.pad_id,
-        rms_norm_eps=1e-5,
-        hidden_dropout_prob=0.1,
-        attention_dropout_prob=0.1,
-        use_stable_embedding=True,
-        shared_input_output_embedding=True,
-    )
-)
-ckpt = torch.load(
-    "data/saved_ckpt/instruction_tuning_math_code_multiturn/36001.pt",
-    map_location="cpu",
-)
-raw_model.load_state_dict(ckpt)
-raw_model.eval()
-model = raw_model.cuda()
-print("ready")
-
-
-def parse_codeblock(text):
-    lines = text.split("\n")
-    for i, line in enumerate(lines):
-        if "```" in line:
-            if line != "```":
-                lines[i] = f'<pre><code class="{lines[i][3:]}">'
-            else:
-                lines[i] = "</code></pre>"
-        else:
-            if i > 0:
-                lines[i] = "<br/>" + line.replace("<", "&lt;").replace(">", "&gt;")
-    return "".join(lines)
+logging.warning("ready")
 
 
 with gr.Blocks() as demo:
     gr.Markdown(
         """
-    # [Open-Llama](https://github.com/Bayes-Song/Open-Llama)
+    # [Open-Llama](https://github.com/s-JoL/Open-Llama)
     完全使用Open-Llama项目从0开始训练的Instruct-GPT模型，当长时间无响应（如20s以上）可刷新重试。
 
     Instruct-GPT model is trained from scratch using the Open-Llama project without relying on any other pre-trained models. If there is no response for a long time (such as more than 20 seconds), please refresh and try again. 
@@ -69,7 +35,7 @@ with gr.Blocks() as demo:
     clear = gr.Button("Clear")
 
     def user(user_message, history):
-        print(user_message)
+        logging.warning(user_message)
         return "", history + [[user_message, None]]
 
     def bot(history):
@@ -80,22 +46,30 @@ with gr.Blocks() as demo:
             if completion is None:
                 inputs = "user:{}\nsystem:".format(prompt)
                 inputs = tokenizer(
-                    inputs, return_tensors=True, add_special_tokens=False
+                    inputs,
+                    return_tensors="pt",
+                    add_special_tokens=False,
+                    return_attention_mask=False,
                 )
                 context.append(inputs["input_ids"])
             else:
                 inputs = "user:{}\nsystem:{}".format(prompt, completion)
-                inputs = tokenizer(inputs, return_tensors=True, add_special_tokens=True)
+                inputs = tokenizer(
+                    inputs,
+                    return_tensors="pt",
+                    add_special_tokens=True,
+                    return_attention_mask=False,
+                )
                 context.append(inputs["input_ids"])
         context = torch.cat(context, dim=-1)
         context = context[:, -1024:]
         inputs_len = context.shape[1]
         context = context.cuda()
-        pred = model.generate(input_ids=context, max_new_tokens=512, do_sample=True)
+        pred = model.generate(input_ids=context, max_new_tokens=1024, do_sample=True)
         pred = pred[:, inputs_len:]
-        pred = tokenizer.decode(pred.cpu())[0]
-        print(pred)
-        bot_message = parse_codeblock(pred)
+        pred = tokenizer.decode(pred.cpu()[0], skip_special_tokens=True)
+        logging.warning(pred)
+        bot_message = pred
         history[-1][1] = bot_message
         return history
 
@@ -108,8 +82,9 @@ with gr.Blocks() as demo:
     当前体验服务生成的所有内容都是由人工智能模型生成，我们对其生成内容的准确性、完整性和功能性不做任何保证，并且其生成的内容不代表我们的态度或观点。
 
     联系方式: sl12160010@gmail.com  对于该项目有任何意见和建议都欢迎联系我.
+    
     Contact information: sl12160010@gmail.com. Any opinions or suggestions regarding the project are welcome to be addressed to me through this email.
     """
     )
 
-demo.launch(share=True)
+demo.launch()

configs/accelerate_configs/ds_stage1.yaml

@@ -1,30 +1,18 @@
 compute_environment: LOCAL_MACHINE
 deepspeed_config:
   deepspeed_multinode_launcher: standard
-  gradient_accumulation_steps: 12
   gradient_clipping: 1.0
   offload_optimizer_device: none
   offload_param_device: none
   zero3_init_flag: false
   zero_stage: 1
 distributed_type: DEEPSPEED
-downcast_bf16: 'no'
-dynamo_backend: 'no'
-# dynamo_config: 
-  # dynamo_backend: INDUCTOR
-  # dynamo_mode: default
-  # dynamo_use_dynamic: true
-  # dynamo_use_fullgraph: false
 fsdp_config: {}
 machine_rank: 0
 main_training_function: main
-megatron_lm_config: {}
 mixed_precision: bf16
 num_machines: 1
 num_processes: 8
 rdzv_backend: static
 same_network: true
-tpu_env: []
-tpu_use_cluster: false
-tpu_use_sudo: false
-use_cpu: false
+use_cpu: false

configs/accelerate_configs/ds_stage2.yaml

@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  deepspeed_multinode_launcher: standard
+  gradient_clipping: 1.0
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+use_cpu: false

configs/accelerate_configs/ds_stage3.yaml

@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  deepspeed_multinode_launcher: standard
+  gradient_clipping: 1.0
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+use_cpu: false

configs/accelerate_configs/ds_stage3_offload.yaml

@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  deepspeed_multinode_launcher: standard
+  gradient_clipping: 1.0
+  offload_optimizer_device: cpu
+  offload_param_device: cpu
+  zero3_init_flag: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+use_cpu: false

configs/instruct_config.yaml

@@ -0,0 +1,32 @@
+data:
+  mode: "instruct"
+  data: 
+    mixed: "data/instruction_data/part-*.jsonl.zst"
+  pad_to_max: False
+  sequence_sample_mode: "none"
+  concat_multiple_sequence: True
+  num_sequences: 50
+  seq_length: 2048
+  tokenizer_model_path: "configs/tokenizer_models/llama_tokenizer_extended.model"
+  split_by_shard: False
+train:
+  train_batch_size: 2
+  # 1B token for 1 epoch, 5epoch
+  num_training_steps: 20000
+  num_warmup_steps: 500
+  initializer_range: 1.0e-2
+  lr: 2.0e-4
+  weight_decay: 1.0e-1
+  ckpt: "data/saved_model/ckpt.pth"
+  train_num_workers: 16
+  gradient_accumulation_steps: 1
+  prefetch_factor: 100
+  train_and_eval: False
+  gradient_checkpointing_enable: False
+  use_lora: False
+# global step
+log_interval: 50
+eval_interval: 500
+save_interval: 1000
+work_dir: "data/saved_ckpt/7B_instruction"
+project_name: "Llama Instruction"

configs/instruction_tuning_config.py

@@ -1,25 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-30 21:38:07
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-06 03:37:23
-FilePath: /Open-Llama/configs/instruction_tuning_config.py
-Description: 
-
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-max_length = 1024
-train_batch_size = 2
-num_training_steps = 40000
-num_warmup_steps = 100
-initializer_range = 1e-2
-lr = 2e-4
-weight_decay = 1e-1
-tokenizer_model_path = "configs/10w_vocab_wudao5_pile10.model"
-patterns = ["data/instruction_data/part-*.jsonl.zst"]
-# global step
-log_interval = 50
-eval_interval = 500
-save_interval = 1000
-work_dir = "data/saved_ckpt/"
-ckpt_path = "data/saved_ckpt/83200.pt"

configs/model_configs/13B.json

@@ -0,0 +1,26 @@
+{
+  "architectures": [
+    "OpenLlamaForCausalLM"
+  ],
+  "attention_dropout_prob": 0.1,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 5120,
+  "initializer_range": 1e-2,
+  "intermediate_size": 13824,
+  "max_position_embeddings": 2048,
+  "model_type": "open-llama",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "pad_token_id": 32000,
+  "rms_norm_eps": 1e-05,
+  "shared_input_output_embedding": false,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "use_cache": true,
+  "use_memorry_efficient_attention": true,
+  "use_stable_embedding": false,
+  "vocab_size": 68762
+}

configs/model_configs/33B.json

@@ -0,0 +1,26 @@
+{
+  "architectures": [
+    "OpenLlamaForCausalLM"
+  ],
+  "attention_dropout_prob": 0.1,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 6656,
+  "initializer_range": 1e-2,
+  "intermediate_size": 17920,
+  "max_position_embeddings": 2048,
+  "model_type": "open-llama",
+  "num_attention_heads": 52,
+  "num_hidden_layers": 60,
+  "pad_token_id": 32000,
+  "rms_norm_eps": 1e-05,
+  "shared_input_output_embedding": false,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "use_cache": true,
+  "use_memorry_efficient_attention": true,
+  "use_stable_embedding": false,
+  "vocab_size": 68762
+}

configs/model_configs/65B.json

@@ -0,0 +1,26 @@
+{
+  "architectures": [
+    "OpenLlamaForCausalLM"
+  ],
+  "attention_dropout_prob": 0.1,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 8192,
+  "initializer_range": 1e-2,
+  "intermediate_size": 22016,
+  "max_position_embeddings": 2048,
+  "model_type": "open-llama",
+  "num_attention_heads": 64,
+  "num_hidden_layers": 80,
+  "pad_token_id": 32000,
+  "rms_norm_eps": 1e-05,
+  "shared_input_output_embedding": false,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "use_cache": true,
+  "use_memorry_efficient_attention": true,
+  "use_stable_embedding": false,
+  "vocab_size": 68762
+}

configs/model_configs/7B.json

@@ -0,0 +1,26 @@
+{
+  "architectures": [
+    "OpenLlamaForCausalLM"
+  ],
+  "attention_dropout_prob": 0.1,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 4096,
+  "initializer_range": 1e-2,
+  "intermediate_size": 11008,
+  "max_position_embeddings": 2048,
+  "model_type": "open-llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "pad_token_id": 32000,
+  "rms_norm_eps": 1e-05,
+  "shared_input_output_embedding": false,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "use_cache": true,
+  "use_memorry_efficient_attention": true,
+  "use_stable_embedding": false,
+  "vocab_size": 68762
+}

configs/pretrain_config.py

@@ -1,14 +0,0 @@
-max_length = 1024
-train_batch_size = 2
-num_training_steps = 1000000
-num_warmup_steps = 2000
-initializer_range = 1e-2
-lr = 2e-4
-weight_decay = 1e-1
-tokenizer_model_path = "configs/10w_vocab_wudao5_pile10.model"
-patterns = ["data/pretrain_data/part-*.jsonl.zst"]
-# global step
-log_interval = 5
-eval_interval = 200
-save_interval = 800
-work_dir = "data/saved_ckpt/"

configs/pretrain_config.yaml

@@ -0,0 +1,32 @@
+data:
+  mode: "pretrain"
+  data: 
+    wudao: "data/pretrain_data/part-wudao*.jsonl.zst"
+    the_pile: "data/pretrain_data/part-pile-1*.jsonl.zst"
+  pad_to_max: False
+  sequence_sample_mode: "none"
+  concat_multiple_sequence: True
+  num_sequences: 10
+  seq_length: 2048
+  tokenizer_model_path: "configs/tokenizer_models/llama_tokenizer_extended.model"
+  split_by_shard: False
+train:
+  train_batch_size: 2
+  num_training_steps: 500000
+  num_warmup_steps: 2000
+  initializer_range: 1.0e-2
+  lr: 2.0e-4
+  weight_decay: 1.0e-1
+  ckpt: null
+  train_num_workers: 16
+  gradient_accumulation_steps: 12
+  prefetch_factor: 100
+  train_and_eval: False
+  gradient_checkpointing_enable: False
+  use_lora: False
+# global step
+log_interval: 5
+eval_interval: 500
+save_interval: 1000
+work_dir: "data/saved_ckpt/7B"
+project_name: "Llama Pretrain"

data/download_instruct.sh

@@ -1,15 +1,15 @@
 #!/bin/bash
 ###
- # @Author: LiangSong(sl12160010@gmail.com)
+ # @Author: s-JoL(sl12160010@gmail.com)
  # @Date: 2023-04-05 23:18:10
- # @LastEditors: LiangSong(sl12160010@gmail.com)
- # @LastEditTime: 2023-04-05 23:34:30
+ # @LastEditors: s-JoL(sl12160010@gmail.com)
+ # @LastEditTime: 2023-05-04 08:24:17
  # @FilePath: /Open-Llama/data/download_instruct.sh
  # @Description: 
  # 
- # Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+ # Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 ### 
 mkdir data/instruction_data
-curl -C - --retry 3 'https://huggingface.co/datasets/RyokoAI/ShareGPT52K/resolve/main/sg_90k_part1.json' -o data/sg_90k_part1.json
-curl -C - --retry 3 'https://huggingface.co/datasets/RyokoAI/ShareGPT52K/resolve/main/sg_90k_part2.json' -o data/sg_90k_part2.json
+wget -c --tries 3 'https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/HTML_cleaned_raw_dataset/sg_90k_part1_html_cleaned.json' -O data/sg_90k_part1_html_cleaned.json
+wget -c --tries 3 'https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/HTML_cleaned_raw_dataset/sg_90k_part2_html_cleaned.json' -O data/sg_90k_part2_html_cleaned.json
 python3 data/preprocess_instruction.py

data/download_the_pile.sh

@@ -1,22 +1,23 @@
 #!/bin/bash
 ###
- # @Author: LiangSong(sl12160010@gmail.com)
+ # @Author: s-JoL(sl12160010@gmail.com)
  # @Date: 2023-03-16 21:21:38
- # @LastEditors: LiangSong(sl12160010@gmail.com)
+ # @LastEditors: s-JoL(sl12160010@gmail.com)
  # @LastEditTime: 2023-03-26 22:58:02
  # @FilePath: /Open-Llama/data/download_the_pile.sh
  # @Description: 
  # download the pile dataset and preprocess
- # Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+ # Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 ### 
 start=0
 end=29
 mkdir data/the_pile
 for (( i=$start; i<=$end; i++ ))
 do
-    url="https://the-eye.eu/public/AI/pile/train/$(printf "%02d" $i).jsonl.zst"
+    url="https://huggingface.co/datasets/monology/pile-uncopyrighted/resolve/main/train/$(printf "%02d" $i).jsonl.zst?download=true"
     echo "Downloading file: $url"
-    curl -C - $url -o data/the_pile/"$(printf "%02d" $i).jsonl.zst"
+    #curl -C - $url -o data/the_pile/"$(printf "%02d" $i).jsonl.zst"
+    curl -L $url -o data/the_pile/"$(printf "%02d" $i).jsonl.zst"
 done
 
 wait

data/download_wudao.sh

@@ -1,19 +1,24 @@
 #!/bin/bash
 ###
- # @Author: LiangSong(sl12160010@gmail.com)
+ # @Author: s-JoL(sl12160010@gmail.com)
  # @Date: 2023-03-16 21:21:56
- # @LastEditors: LiangSong(sl12160010@gmail.com)
+ # @LastEditors: s-JoL(sl12160010@gmail.com)
  # @LastEditTime: 2023-03-26 22:58:11
  # @FilePath: /Open-Llama/data/download_wudao.sh
  # @Description: 
  # download wudao dataset and preprocess
- # Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+ # Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 ### 
+apt-add-repository multiverse
 apt install unrar
-for i in {1..100}
-do
-  curl -C - --retry 100 'https://dorc.baai.ac.cn/resources/data/WuDaoCorpora2.0/WuDaoCorpus2.0_base_200G.rar?AccessKeyId=AKLTNasiLRBBTcOgPqzlkPzu1w&Expires=1679127659&Signature=7jh%2FpnJyC2hAeumm9EjaeE5HN9E%3D' -o data/WuDaoCorpus2.0_base_200G.rar
-done
-unrar x data/WuDaoCorpus2.0_base_200G.rar
+
+wget -v -c 'https://download.scidb.cn/download?fileId=63a30383fed6a8a9e8454302&dataSetType=organization&fileName=WuDaoCorporaText-2.0-open.rar' -O data/WuDaoCorpus2.0_base_200G.rar
+
+# for i in {1..100}
+# do
+#   curl -C - --retry 100 'https://dorc.baai.ac.cn/resources/data/WuDaoCorpora2.0/WuDaoCorpus2.0_base_200G.rar?AccessKeyId=AKLTNasiLRBBTcOgPqzlkPzu1w&Expires=1679127659&Signature=7jh%2FpnJyC2hAeumm9EjaeE5HN9E%3D' -o data/WuDaoCorpus2.0_base_200G.rar
+# done
+
+unrar x data/WuDaoCorpus2.0_base_200G.rar data/
 mkdir data/pretrain_data
 python3 data/preprocess_wudao.py

data/preprocess_instruction.py

@@ -1,167 +1,90 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-03-30 20:52:10
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-05 23:51:16
+LastEditors: s-JoL(sl12160010@gmail.com)
+LastEditTime: 2023-05-04 08:32:04
 FilePath: /Open-Llama/data/preprocess_instruction.py
 Description: 
 
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 import json
+from tqdm import tqdm
 import zstandard as zstd
 from datasets import load_dataset
 
 
 root_dir = "data"
+write_path = "data/instruction_data/part-{}-{}.jsonl.zst"
+dataset_map = {
+    "yizhongw/self_instruct": "self_instruct",
+    "BelleGroup/train_0.5M_CN": "belle_0.5M",
+    "BelleGroup/train_1M_CN": "belle_1M",
+    "BelleGroup/train_2M_CN": "belle_2M",
+    "BelleGroup/school_math_0.25M": "belle_school_math_0.25M",
+    "BelleGroup/multiturn_chat_0.8M": "belle_multiturn_chat_0.8M",
+    "Graverman/Instruct-to-Code": "instruct_to_code",
+    "qwedsacf/grade-school-math-instructions": "grade_school_math",
+    "camel-ai/math": "camel_ai_math",
+    "camel-ai/physics": "camel_ai_physics",
+    "camel-ai/chemistry": "camel_ai_chemistry",
+    "camel-ai/biology": "camel_ai_biology",
+    ("bigscience/xP3mt", "code"): "xP3mt_code",
+    ("bigscience/xP3mt", "zh"): "xP3mt_zh",
+}
 
-dataset = load_dataset("yizhongw/self_instruct")
-write_path = root_dir + "/instruction_data/part-self_instruct-{}.jsonl.zst"
-total_num = 0
-file_num = 0
-wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-for line in dataset["train"]:
-    line = json.dumps(line)
-    if total_num % 1024 == 0 and total_num > 0:
-        file_num += 1
-        wfp.close()
-        wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-    wfp.write(line.encode("utf-8"))
-    wfp.write(b"\n")
-    total_num += 1
-wfp.close()
-print(
-    "yizhongw/self_instruct preprocess done. Total line: {}, Total file: {}".format(
-        total_num, file_num
+
+def process_hf_dataset(name, local_name):
+    if isinstance(name, str):
+        dataset = load_dataset(name)
+    else:
+        dataset = load_dataset(*name)
+    total_num = 0
+    file_num = 1
+    wfp = zstd.open(write_path.format(local_name, file_num), "wb", encoding="utf-8")
+    for line in tqdm(dataset["train"]):
+        line = json.dumps(line)
+        if total_num % 1024 == 0 and total_num > 0:
+            file_num += 1
+            wfp.close()
+            wfp = zstd.open(
+                write_path.format(local_name, file_num), "wb", encoding="utf-8"
+            )
+        wfp.write(line.encode("utf-8"))
+        wfp.write(b"\n")
+        total_num += 1
+    wfp.close()
+    print(
+        "{} preprocess done. Total line: {}, Total file: {}".format(
+            name, total_num, file_num
+        )
     )
-)
 
-dataset = load_dataset("BelleGroup/train_0.5M_CN")
-write_path = root_dir + "/instruction_data/part-belle_0.5M-{}.jsonl.zst"
-total_num = 0
-file_num = 0
-wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-for line in dataset["train"]:
-    line = json.dumps(line)
-    if total_num % 1024 == 0 and total_num > 0:
-        file_num += 1
-        wfp.close()
-        wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-    wfp.write(line.encode("utf-8"))
-    wfp.write(b"\n")
-    total_num += 1
-wfp.close()
-print(
-    "BelleGroup/train_0.5M_CN preprocess done. Total line: {}, Total file: {}".format(
-        total_num, file_num
-    )
-)
 
-dataset = load_dataset("BelleGroup/train_1M_CN")
-write_path = root_dir + "/instruction_data/part-belle_1M-{}.jsonl.zst"
-total_num = 0
-file_num = 0
-wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-for line in dataset["train"]:
-    line = json.dumps(line)
-    if total_num % 1024 == 0 and total_num > 0:
-        file_num += 1
-        wfp.close()
-        wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-    wfp.write(line.encode("utf-8"))
-    wfp.write(b"\n")
-    total_num += 1
-wfp.close()
-print(
-    "BelleGroup/train_1M_CN preprocess done. Total line: {}, Total file: {}".format(
-        total_num, file_num
-    )
-)
+for k, v in dataset_map.items():
+    process_hf_dataset(k, v)
 
-dataset = load_dataset("BelleGroup/school_math_0.25M")
-write_path = root_dir + "/instruction_data/part-belle_school_math_0.25M-{}.jsonl.zst"
+local_name = "sharegpt_90K"
 total_num = 0
-file_num = 0
-wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-for line in dataset["train"]:
-    line = json.dumps(line)
-    if total_num % 1024 == 0 and total_num > 0:
-        file_num += 1
-        wfp.close()
-        wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-    wfp.write(line.encode("utf-8"))
-    wfp.write(b"\n")
-    total_num += 1
-wfp.close()
-print(
-    "BelleGroup/school_math_0.25M preprocess done. Total line: {}, Total file: {}".format(
-        total_num, file_num
-    )
-)
-
-dataset = load_dataset("BelleGroup/multiturn_chat_0.8M")
-write_path = root_dir + "/instruction_data/part-belle_multiturn_chat_0.8M-{}.jsonl.zst"
-total_num = 0
-file_num = 0
-wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-for line in dataset["train"]:
-    line = json.dumps(line)
-    if total_num % 1024 == 0 and total_num > 0:
-        file_num += 1
-        wfp.close()
-        wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-    wfp.write(line.encode("utf-8"))
-    wfp.write(b"\n")
-    total_num += 1
-wfp.close()
-print(
-    "BelleGroup/multiturn_chat_0.8M preprocess done. Total line: {}, Total file: {}".format(
-        total_num, file_num
-    )
-)
-
-dataset = load_dataset("Graverman/Instruct-to-Code")
-write_path = root_dir + "/instruction_data/part-instruct_to_code-{}.jsonl.zst"
-total_num = 0
-file_num = 0
-wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-for line in dataset["train"]:
-    line = json.dumps(line)
-    if total_num % 1024 == 0 and total_num > 0:
-        file_num += 1
-        wfp.close()
-        wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-    wfp.write(line.encode("utf-8"))
-    wfp.write(b"\n")
-    total_num += 1
-wfp.close()
-print(
-    "Graverman/Instruct-to-Code preprocess done. Total line: {}, Total file: {}".format(
-        total_num, file_num
-    )
-)
-
-write_path = root_dir + "/instruction_data/part-sharegpt_90K-{}.jsonl.zst"
-total_num = 0
-file_num = 0
-wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
-with open("data/sg_90k_part1.json", "r") as fp:
+file_num = 1
+wfp = zstd.open(write_path.format(local_name, file_num), "wb", encoding="utf-8")
+with open("{}/sg_90k_part1_html_cleaned.json".format(root_dir), "r") as fp:
     data1 = json.load(fp)
-with open("data/sg_90k_part2.json", "r") as fp:
+with open("{}/sg_90k_part2_html_cleaned.json".format(root_dir), "r") as fp:
     data2 = json.load(fp)
 data = data1 + data2
-for line in data:
+for line in tqdm(data):
     line = json.dumps(line)
     if total_num % 1024 == 0 and total_num > 0:
         file_num += 1
         wfp.close()
-        wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
+        wfp = zstd.open(write_path.format(local_name, file_num), "wb", encoding="utf-8")
     wfp.write(line.encode("utf-8"))
     wfp.write(b"\n")
     total_num += 1
 wfp.close()
 print(
-    "RyokoAI/ShareGPT52K preprocess done. Total line: {}, Total file: {}".format(
+    "anon8231489123/ShareGPT_Vicuna_unfiltered preprocess done. Total line: {}, Total file: {}".format(
         total_num, file_num
     )
 )

data/preprocess_the_pile.py

@@ -1,13 +1,13 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-03-16 22:35:38
-LastEditors: LiangSong(sl12160010@gmail.com)
+LastEditors: s-JoL(sl12160010@gmail.com)
 LastEditTime: 2023-03-26 22:59:38
 FilePath: /Open-Llama/data/preprocess_the_pile.py
 Description: 
 Parse the dataset from the raw files and split them into different jsonl files based on the preset maximum number of lines, 
 making it easy for parallel training to perform streaming reads.
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 import json
 from glob import glob
@@ -17,7 +17,7 @@ import zstandard as zstd
 paths = glob("data/the_pile/*.jsonl.zst")
 write_path = "data/pretrain_data/part-pile-{}.jsonl.zst"
 total_num = 0
-file_num = 0
+file_num = 1
 wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
 for path in tqdm(paths, total=len(paths)):
     with zstd.open(path, "r", encoding="utf-8") as fp:

data/preprocess_wudao.py

@@ -1,13 +1,13 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-03-16 22:10:44
-LastEditors: LiangSong(sl12160010@gmail.com)
+LastEditors: s-JoL(sl12160010@gmail.com)
 LastEditTime: 2023-03-26 22:59:55
 FilePath: /Open-Llama/data/preprocess_wudao.py
 Description: 
 Parse the dataset from the raw files and split them into different jsonl files based on the preset maximum number of lines, 
 making it easy for parallel training to perform streaming reads.
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 import json
 from glob import glob
@@ -17,7 +17,7 @@ import zstandard as zstd
 paths = glob("data/WuDaoCorpus2.0_base_200G/part*")
 write_path = "data/pretrain_data/part-wudao-{}.jsonl.zst"
 total_num = 0
-file_num = 0
+file_num = 1
 wfp = zstd.open(write_path.format(file_num), "wb", encoding="utf-8")
 for path in tqdm(paths, total=len(paths)):
     with open(path, "r") as fp:

dataset/collate_fn.py

@@ -1,69 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-30 20:58:16
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-05 22:11:03
-FilePath: /Open-Llama/dataset/collate_fn.py
-Description: 
-
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-import torch
-
-
-def collate_fn_gen(tokenizer, segment_max_length=1024, padding="longest"):
-    """
-    Organize data into tensors by padding based on the preset maximum length.
-    """
-    pad_id = tokenizer.pad_id
-
-    def collate_fn(batch):
-        if padding == "longest":
-            max_length = max([len(i) for i in batch])
-        elif padding == "max_length":
-            max_length = segment_max_length
-        else:
-            raise Exception("Invalid argumet for padding: {}".format(padding))
-        input_ids = []
-        for i in batch:
-            input_len = len(i)
-            input_ids.append(i + [pad_id] * (max_length - input_len))
-        inputs = {
-            "input_ids": torch.tensor(input_ids, dtype=torch.int64),
-        }
-        return inputs
-
-    return collate_fn
-
-
-if __name__ == "__main__":
-    import sentencepiece as spm
-    from torch.utils.data import DataLoader
-
-    from dataset.pretrain_dataset import preprocess_wudao_gen, preprocess_the_pile_gen
-
-    from dataset.tokenizer import Tokenizer
-    from dataset.data_iter import create_shard_kwargs, DataIter
-
-    sp_model = spm.SentencePieceProcessor(
-        model_file="configs/10w_vocab_wudao5_pile10.model"
-    )
-    tokenizer = Tokenizer(sp_model)
-    patterns = ["data/pretrain_data/part-*.jsonl.zst"]
-    paths = create_shard_kwargs(patterns)
-    transform_dict = {
-        "wudao": preprocess_wudao_gen(tokenizer),
-        "pile": preprocess_the_pile_gen(tokenizer),
-    }
-    data_set = DataIter(paths, transform_dict=transform_dict)
-    train_loader = DataLoader(
-        data_set,
-        batch_size=8,
-        num_workers=4,
-        collate_fn=collate_fn_gen(tokenizer),
-        drop_last=True,
-    )
-    for batch in train_loader:
-        for k, v in batch.items():
-            print(k, v.shape)
-        break

dataset/data_iter.py

@@ -1,116 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-17 19:32:20
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-06 03:37:55
-FilePath: /Open-Llama/dataset/data_iter.py
-Description: 
-
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-import json
-from glob import glob
-import zstandard as zstd
-from torch.utils.data import IterableDataset
-
-
-class DataIter(IterableDataset):
-    """
-    Currently, the allowed storage formats are jsonl.zst.
-    Each line of the data is a dictionary, which can be parsed as JSON for subsequent processing after reading.
-    Currently, only single worker is supported.
-    """
-
-    def __init__(
-        self,
-        paths_with_index,
-        transform_dict=None,
-        max_length=None,
-        concat_docs=False,
-        process_index=0,
-        num_processes=1,
-    ):
-        super().__init__()
-        self.paths_with_index = paths_with_index
-        self.max_length = max_length
-        self.transform_dict = transform_dict
-        self.concat_docs = concat_docs
-        self.process_index = process_index
-        self.num_processes = num_processes
-        if self.concat_docs:
-            self.cache = []
-
-    def __iter__(self):
-        past = None
-        for i, path in self.paths_with_index:
-            # part-dataset_name-01.jsonl.zst
-            dataset_name = path.split("-")[-2]
-            # shard to multiple device
-            if self.num_processes > 1 and i % self.num_processes != self.process_index:
-                continue
-            # Log the file name when encountering a new file.
-            if past != dataset_name:
-                print("Loading data from {}".format(path))
-                past = path
-            # Currently, the allowed storage formats are jsonl.zst.
-            assert path.endswith("jsonl.zst")
-            with zstd.open(path, "r", encoding="utf-8") as fp:
-                for line in fp:
-                    # If the length of the cache is greater than max_length.
-                    if self.concat_docs and len(self.cache) >= self.max_length:
-                        seq = self.cache[: self.max_length]
-                        self.cache = self.cache[self.max_length :]
-                        yield seq
-                    if isinstance(line, bytes):
-                        line = line.decode("utf-8")
-                    line = json.loads(line)
-                    line["dataset"] = dataset_name
-                    # Transformation, including sample, tokenize, etc.
-                    if self.transform_dict:
-                        line = self.transform_dict[dataset_name](line)
-                        # skip bad doc
-                        if line is None:
-                            continue
-                        elif isinstance(line, str):
-                            yield line
-                        # must be list of list
-                        elif isinstance(line, list) and isinstance(line[0], list):
-                            for seq in line:
-                                if self.concat_docs:
-                                    # concat seq from multiple docs
-                                    self.cache += seq
-                                else:
-                                    yield seq
-                        else:
-                            raise Exception(
-                                "Unsupported type in Transformation: {}".format(
-                                    self.transform_dict[dataset_name]
-                                )
-                            )
-                    else:
-                        yield line
-
-
-def create_shard_kwargs(patterns, repeat=1):
-    """
-    Assign numbers to different shards of data to ensure that data is not duplicated
-    when allocated to different nodes during distributed training.
-    """
-    all_path = []
-    for p in patterns:
-        all_path.extend(glob(p))
-    all_path *= repeat
-    return [(i, p) for i, p in enumerate(all_path)]
-
-
-if __name__ == "__main__":
-    patterns = ["data/pretrain_data/part-wudao*.jsonl.zst"]
-    paths = create_shard_kwargs(patterns)
-    transform_dict = {"wudao": lambda x: x["title"], "pile": lambda x: [x["text"]]}
-    data_iter = DataIter(
-        paths, transform_dict=transform_dict, max_length=16, concat_docs=True
-    )
-    for i, data in enumerate(data_iter):
-        print(i, data)
-        if i == 20:
-            break

dataset/dataset.py

@@ -0,0 +1,321 @@
+"""
+Author: s-JoL(sl12160010@gmail.com)
+Date: 2023-04-24 20:05:21
+LastEditors: s-JoL(sl12160010@gmail.com)
+LastEditTime: 2023-05-06 23:30:37
+FilePath: /Open-Llama/dataset/dataset.py
+Description: 
+
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
+"""
+import math
+import torch
+import random
+from glob import glob
+from datasets import load_dataset
+
+
+random.seed(42)
+
+
+def pretrain_transform(batch):
+    # wudao preprocess
+    if "title" in batch and "content" in batch:
+        assert len(batch["title"]) == 1
+        batch["text"] = [batch["title"][0] + "\n" + batch["content"][0]]
+    elif "text" in batch:
+        pass
+    else:
+        raise Exception("Unrecognized pretrain dataset format.")
+    return batch
+
+
+def instruct_transform(batch):
+    # self instruct preprocess
+    if "prompt" in batch and "completion" in batch:
+        prompt = batch["prompt"][0]
+        completion = batch["completion"][0]
+        if prompt.endswith("Output:"):
+            prompt = prompt[:-7]
+        text = "user:{}\nsystem:{}".format(prompt.strip(), completion.strip())
+        texts = [text]
+    # belle preprocess
+    elif "instruction" in batch and "output" in batch:
+        prompt = batch["instruction"][0].replace("\\n", "")
+        prompt = prompt.strip("")
+
+        completion = batch["output"][0].replace("\\n", "")
+        completion = completion.strip("")
+        # multi turn chat
+        if "Human:" in prompt:
+            texts = []
+            chats = prompt + completion
+            chats = chats.split("Human:")
+            for chat in chats:
+                if chat.strip() == "":
+                    continue
+                res = chat.split("Assistant:")
+                if len(res) != 2:
+                    continue
+                prompt, completion = res
+                prompt = prompt.strip()
+                completion = completion.strip()
+                chat = "user:{}\nsystem:{}".format(prompt, completion)
+                texts.append(chat)
+            texts = ["[multiturn_sep]".join(texts)]
+        else:
+            text = "user:{}\nsystem:{}".format(prompt, completion)
+            texts = [text]
+    # instruct code preprocess
+    elif "instruction" in batch and "answer" in batch:
+        prompt = batch["instruction"][0].replace("\\n", "")
+        prompt = prompt.strip("")
+
+        completion = batch["answer"][0].replace("\\n", "")
+        completion = completion.strip("")
+        text = "user:{}\nsystem:{}".format(prompt, completion)
+        texts = [text]
+    # share gpt preprocess
+    elif "conversations" in batch:
+        chats = batch["conversations"][0]
+        if chats[0]["from"] != "human":
+            chats = chats[1:]
+        texts = []
+        for i in range(len(chats) // 2):
+            prompt = chats[2 * i]
+            completion = chats[2 * i + 1]
+            if not (prompt["from"] == "human" and completion["from"] == "gpt"):
+                continue
+            prompt = prompt["value"]
+            prompt = prompt.strip()
+            completion = completion["value"]
+            completion = completion.strip()
+            chat = "user:{}\nsystem:{}".format(prompt, completion)
+            texts.append(chat)
+        texts = ["[multiturn_sep]".join(texts)]
+    # xP3 preprocess
+    elif "inputs" in batch and "targets" in batch:
+        inputs = batch["inputs"][0]
+        targets = batch["targets"][0]
+        text = "user:{}\nsystem:{}".format(inputs.strip(), targets.strip())
+        texts = [text]
+    # camel-ai preprocess
+    elif "message_1" in batch and "message_2" in batch:
+        inputs = batch["message_1"][0]
+        targets = batch["message_2"][0]
+        text = "user:{}\nsystem:{}".format(inputs.strip(), targets.strip())
+        texts = [text]
+    # grade-school-math-instructions preprocess
+    elif "INSTRUCTION" in batch and "RESPONSE" in batch:
+        inputs = batch["INSTRUCTION"][0]
+        targets = batch["RESPONSE"][0]
+        text = "user:{}\nsystem:{}".format(inputs.strip(), targets.strip())
+        texts = [text]
+    else:
+        raise Exception("Unrecognized instruct dataset format.")
+    return {"text": texts}
+
+
+def split_multiturn(batch):
+    return {"text": batch["text"][0].split("[multiturn_sep]")}
+
+
+def sample_sequence_gen(seq_length, eos_token_id):
+    def sample_sequence(line):
+        doc_length = line["input_ids"].shape[0]
+        if doc_length <= seq_length:
+            start = 0
+        else:
+            if random.random() < 1 / 4:
+                start = 0
+            else:
+                start = random.randint(0, doc_length - seq_length)
+        input_ids = line["input_ids"][start : start + seq_length]
+        if input_ids[-1] != eos_token_id:
+            input_ids[-1] = eos_token_id
+        return {"input_ids": input_ids}
+
+    return sample_sequence
+
+
+def split_sequence_gen(seq_length):
+    def split_sequence(batch):
+        input_ids = batch["input_ids"][0]
+        out = []
+        while len(input_ids) >= (1 + len(out)) * seq_length:
+            out.append(input_ids[len(out) * seq_length : (1 + len(out)) * seq_length])
+        return {"input_ids": out}
+
+    return split_sequence
+
+
+def concat_multiple_sequence_gen(seq_length, pad_token_id):
+    def concat_multiple_sequence(batch):
+        concat_input_ids = torch.cat(batch["input_ids"], dim=0)
+        length = concat_input_ids.shape[0]
+        chunks = math.ceil(length / seq_length)
+        pad_length = chunks * seq_length - length
+        pad = torch.ones(pad_length, dtype=concat_input_ids.dtype) * pad_token_id
+        concat_input_ids = torch.cat([concat_input_ids, pad], dim=0)
+        input_ids = torch.chunk(concat_input_ids, chunks)
+        return {"input_ids": input_ids}
+
+    return concat_multiple_sequence
+
+
+def get_labels_gen(pad_token_id):
+    def get_labels(line):
+        input_ids = line["input_ids"]
+        labels = input_ids.clone()
+        labels[labels == pad_token_id] = -100
+        return {"labels": labels}
+
+    return get_labels
+
+
+def construct_dataset(
+    dataset_config, tokenizer, return_raw_text=False, world_size=None
+):
+    all_data_files = []
+    for name, pattern in dataset_config["data"].items():
+        data_files = glob(pattern)
+        assert len(data_files) > 0
+        all_data_files.extend(data_files)
+    random.shuffle(all_data_files)
+    # 当shard可以被world_size整除时 split_dataset_by_node 会直接按shard进行划分，否则会读所有数据然后跳过一部分，可能会慢一点
+    # https://huggingface.co/docs/datasets/package_reference/main_classes#datasets.distributed.split_dataset_by_node
+    if world_size is not None:
+        num_shards = len(all_data_files)
+        all_data_files = all_data_files[: num_shards // world_size * world_size]
+    dataset = load_dataset(
+        "json", data_files=all_data_files, split="train", streaming=True
+    )
+    # shuffle
+    dataset = dataset.shuffle(seed=42)
+    # 文本预处理转换为统一格式
+    if dataset_config["mode"] == "pretrain":
+        dataset = dataset.map(pretrain_transform, batched=True, batch_size=1)
+    elif dataset_config["mode"] == "instruct":
+        dataset = dataset.map(instruct_transform, batched=True, batch_size=1)
+        dataset = dataset.select_columns("text")
+        dataset = dataset.map(split_multiturn, batched=True, batch_size=1)
+    else:
+        raise Exception("Dataset mode: {} not found.".format(dataset_config["mode"]))
+
+    full_dataset = dataset
+
+    # to visualize
+    if return_raw_text:
+        return full_dataset
+
+    seq_length = dataset_config["seq_length"]
+    pad_to_max = dataset_config.get("pad_to_max", True)
+    sequence_sample_mode = dataset_config.get("sequence_sample_mode", "truncation")
+    truncation = sequence_sample_mode == "truncation"
+    concat_multiple_sequence = dataset_config.get("concat_multiple_sequence", False)
+    # tokenize
+    if pad_to_max:
+        full_dataset = full_dataset.map(
+            lambda x: tokenizer(
+                x["text"],
+                return_tensors="pt",
+                return_attention_mask=False,
+                padding="max_length",
+                max_length=seq_length,
+                truncation=truncation,
+            )
+        )
+    else:
+        full_dataset = full_dataset.map(
+            lambda x: tokenizer(
+                x["text"],
+                return_tensors="pt",
+                return_attention_mask=False,
+                truncation=truncation,
+            )
+        )
+
+    # format
+    full_dataset = full_dataset.map(lambda x: {"input_ids": x["input_ids"][0]})
+    full_dataset = full_dataset.select_columns("input_ids")
+
+    # sequence_sample
+    if sequence_sample_mode == "truncation":
+        pass
+    elif sequence_sample_mode == "none":
+        pass
+    elif sequence_sample_mode == "sample":
+        assert pad_to_max or concat_multiple_sequence
+        full_dataset = full_dataset.map(
+            sample_sequence_gen(seq_length, tokenizer.eos_token_id)
+        )
+    elif sequence_sample_mode == "split":
+        assert not concat_multiple_sequence
+        full_dataset = full_dataset.map(
+            split_sequence_gen(seq_length), batched=True, batch_size=1
+        )
+    else:
+        raise Exception(
+            "Unknown sequence_sample mode: {}.".format(sequence_sample_mode)
+        )
+
+    # concat multiple sequence
+    if concat_multiple_sequence:
+        num_sequences = dataset_config["num_sequences"]
+        full_dataset = full_dataset.map(
+            concat_multiple_sequence_gen(seq_length, tokenizer.pad_token_id),
+            batched=True,
+            batch_size=num_sequences,
+            drop_last_batch=True,
+        )
+
+    # add label
+    full_dataset = full_dataset.map(get_labels_gen(tokenizer.pad_token_id))
+
+    # shuffle
+    full_dataset = full_dataset.shuffle(seed=42)
+    return full_dataset
+
+
+if __name__ == "__main__":
+    import time
+    from unicodedata import normalize
+    from torch.utils.data import DataLoader
+    from transformers import LlamaTokenizer
+
+    data_config = {
+        "mode": "pretrain",
+        "data": {"mixed": "data/pretrain_data/part-*.jsonl.zst"},
+        "pad_to_max": False,
+        "sequence_sample_mode": "sample",
+        "concat_multiple_sequence": True,
+        "num_sequences": 10,
+        "seq_length": 2048,
+    }
+    tokenizer = LlamaTokenizer(
+        "configs/tokenizer_models/llama_tokenizer_extended.model",
+        pad_token="<pad>",
+        add_bos_token=False,
+        add_eos_token=True,
+    )
+    pretrain_dataset = construct_dataset(data_config, tokenizer, True)
+    start = time.time()
+    for i, line in enumerate(pretrain_dataset):
+        raw_text = line["text"]
+        # raw_text = normalize("NFKC", raw_text)
+        input_ids = tokenizer(
+            line["text"], return_tensors="pt", return_attention_mask=False
+        )["input_ids"][0]
+        decode_text = tokenizer.decode(input_ids, skip_special_tokens=True)
+        if raw_text != decode_text and "▁" not in raw_text:
+            print(raw_text, "\n", decode_text)
+        if i == 3000:
+            break
+    print("all checked in {} seconds.".format(time.time() - start))
+    pretrain_dataset = construct_dataset(data_config, tokenizer)
+    print(pretrain_dataset.n_shards)
+    pretrain_loader = DataLoader(pretrain_dataset, batch_size=2, num_workers=16)
+    for batch in pretrain_loader:
+        for k, v in batch.items():
+            print(k, v.shape, "\n", v)
+        break

dataset/instruction_dataset.py

@@ -1,178 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-30 21:02:00
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-06 03:33:27
-FilePath: /Open-Llama/dataset/instruction_dataset.py
-Description: 
-
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-import math
-
-
-def preprocess_self_instruction_gen(tokenizer, segment_max_length=1024):
-    def preprocess_self_instruction(line):
-        """
-        The format of the data is roughly as follows.
-        {'prompt': 'Explain the origin of life on earth. Output:', 'completion': 'Life on Earth is believed to have'}
-        Split the data based on the tokenized length according to the maximum length.
-        """
-        prompt = line["prompt"]
-        if prompt.endswith("Output:"):
-            prompt = prompt[:-7]
-        total = "user:{}\nsystem:{}".format(prompt.strip(), line["completion"].strip())
-        out = tokenizer(total)
-        input_ids = out["input_ids"]
-        return [
-            input_ids[i * segment_max_length : (i + 1) * segment_max_length]
-            for i in range(math.ceil(len(input_ids) / segment_max_length))
-        ]
-
-    return preprocess_self_instruction
-
-
-def preprocess_belle_gen(tokenizer, segment_max_length=1024):
-    def preprocess_belle(line):
-        """
-        The format of the data is roughly as follows.
-        {'text': 'some text', 'meta': {'pile_set_name': 'Github'}}
-        Split the data based on the tokenized length according to the maximum length.
-        """
-        prompt = line["instruction"].replace("\\n", "")
-        prompt = prompt.strip("")
-
-        completion = line["output"].replace("\\n", "")
-        completion = completion.strip("")
-        total = "user:{}\nsystem:{}".format(prompt, completion)
-        out = tokenizer(total)
-        input_ids = out["input_ids"]
-        return [
-            input_ids[i * segment_max_length : (i + 1) * segment_max_length]
-            for i in range(math.ceil(len(input_ids) / segment_max_length))
-        ]
-
-    return preprocess_belle
-
-
-def preprocess_belle_multiturn_chat_gen(tokenizer, segment_max_length=1024):
-    def preprocess_belle_multiturn_chat(line):
-        """
-        The format of the data is roughly as follows.
-        {'text': 'some text', 'meta': {'pile_set_name': 'Github'}}
-        Split the data based on the tokenized length according to the maximum length.
-        """
-        prompt = line["instruction"].replace("\\n", "")
-        prompt = prompt.strip("")
-
-        completion = line["output"].replace("\\n", "")
-        completion = completion.strip("")
-        chats = prompt + completion
-        chats = chats.split("Human:")
-        input_ids = []
-        for chat in chats:
-            if chat.strip() == "":
-                continue
-            res = chat.split("Assistant:")
-            if len(res) != 2:
-                continue
-            prompt, completion = res
-            prompt = prompt.strip()
-            completion = completion.strip()
-            chat = "user:{}\nsystem:{}".format(prompt, completion)
-            out = tokenizer(chat)
-            input_ids.extend(out["input_ids"])
-        if len(input_ids) == 0:
-            return None
-        return [
-            input_ids[i * segment_max_length : (i + 1) * segment_max_length]
-            for i in range(math.ceil(len(input_ids) / segment_max_length))
-        ]
-
-    return preprocess_belle_multiturn_chat
-
-
-def preprocess_sharegpt_gen(tokenizer, segment_max_length=1024):
-    def preprocess_sharegpt(line):
-        """
-        The format of the data is roughly as follows.
-        {'text': 'some text', 'meta': {'pile_set_name': 'Github'}}
-        Split the data based on the tokenized length according to the maximum length.
-        """
-        chats = line["conversations"]
-        if chats[0]["from"] != "human":
-            chats = chats[1:]
-        input_ids = []
-        for i in range(len(chats) // 2):
-            prompt = chats[2 * i]
-            completion = chats[2 * i + 1]
-            if not (prompt["from"] == "human" and completion["from"] == "gpt"):
-                continue
-            prompt = prompt["value"]
-            prompt = prompt.strip()
-            completion = completion["value"]
-            completion = completion.strip()
-            chat = "user:{}\nsystem:{}".format(prompt, completion)
-            out = tokenizer(chat)
-            input_ids.extend(out["input_ids"])
-        if input_ids == []:
-            return None
-        return [
-            input_ids[i * segment_max_length : (i + 1) * segment_max_length]
-            for i in range(math.ceil(len(input_ids) / segment_max_length))
-        ]
-
-    return preprocess_sharegpt
-
-
-def preprocess_instruct_code_gen(tokenizer, segment_max_length=1024):
-    def preprocess_instruct_code(line):
-        """
-        The format of the data is roughly as follows.
-        {'text': 'some text', 'meta': {'pile_set_name': 'Github'}}
-        Split the data based on the tokenized length according to the maximum length.
-        """
-        prompt = line["instruction"].replace("\\n", "")
-        prompt = prompt.strip("")
-
-        completion = line["answer"].replace("\\n", "")
-        completion = completion.strip("")
-        total = "user:{}\nsystem:{}".format(prompt, completion)
-        out = tokenizer(total)
-        input_ids = out["input_ids"]
-        return [
-            input_ids[i * segment_max_length : (i + 1) * segment_max_length]
-            for i in range(math.ceil(len(input_ids) / segment_max_length))
-        ]
-
-    return preprocess_instruct_code
-
-
-if __name__ == "__main__":
-    import sentencepiece as spm
-
-    from dataset.tokenizer import Tokenizer
-    from dataset.data_iter import create_shard_kwargs, DataIter
-
-    sp_model = spm.SentencePieceProcessor(
-        model_file="configs/10w_vocab_wudao5_pile10.model"
-    )
-    tokenizer = Tokenizer(sp_model)
-    patterns = ["data/instruction_data/part-belle_multiturn_chat_0.8M-*.jsonl.zst"]
-    paths = create_shard_kwargs(patterns)
-    transform_dict = {
-        "self_instruct": preprocess_self_instruction_gen(tokenizer),
-        "belle_1M": preprocess_belle_gen(tokenizer),
-        "belle_0.5M": preprocess_belle_gen(tokenizer),
-        "belle_school_math_0.25M": preprocess_belle_gen(tokenizer),
-        "belle_multiturn_chat_0.8M": preprocess_belle_multiturn_chat_gen(tokenizer),
-        "instruct_to_code": preprocess_instruct_code_gen(tokenizer),
-        "sharegpt_90K": preprocess_sharegpt_gen(tokenizer),
-    }
-    data_set = DataIter(
-        paths, transform_dict=transform_dict, concat_docs=True, max_length=1024
-    )
-    for i, sample in enumerate(data_set):
-        print(sp_model.decode(sample))
-        if i == 1:
-            break

dataset/pretrain_dataset.py

@@ -1,71 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-17 20:41:25
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-05 22:32:39
-FilePath: /Open-Llama/dataset/pretrain_dataset.py
-Description: 
-
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-import math
-
-
-def preprocess_wudao_gen(tokenizer, segment_max_length=1024):
-    def preprocess_wudao(line):
-        """
-        The format of the data is roughly as follows.
-        {'id': 1, 'dataType': '百科', 'title': 'some title', 'content': 'some content'}
-        Split the data based on the tokenized length according to the maximum length.
-        """
-        total = line["title"] + "\n" + line["content"]
-        out = tokenizer(total)
-        input_ids = out["input_ids"]
-        return [
-            input_ids[i * segment_max_length : (i + 1) * segment_max_length]
-            for i in range(math.ceil(len(input_ids) / segment_max_length))
-        ]
-
-    return preprocess_wudao
-
-
-def preprocess_the_pile_gen(tokenizer, segment_max_length=1024):
-    def preprocess_the_pile(line):
-        """
-        The format of the data is roughly as follows.
-        {'text': 'some text', 'meta': {'pile_set_name': 'Github'}}
-        Split the data based on the tokenized length according to the maximum length.
-        """
-        total = line["text"]
-        out = tokenizer(total)
-        input_ids = out["input_ids"]
-        return [
-            input_ids[i * segment_max_length : (i + 1) * segment_max_length]
-            for i in range(math.ceil(len(input_ids) / segment_max_length))
-        ]
-
-    return preprocess_the_pile
-
-
-if __name__ == "__main__":
-    import sentencepiece as spm
-
-    from dataset.tokenizer import Tokenizer
-    from dataset.data_iter import create_shard_kwargs, DataIter
-
-    sp_model = spm.SentencePieceProcessor(
-        model_file="configs/10w_vocab_wudao5_pile10.model"
-    )
-    tokenizer = Tokenizer(sp_model)
-    patterns = ["data/pretrain_data/part-*.jsonl.zst"]
-    paths = create_shard_kwargs(patterns)
-    transform_dict = {
-        "wudao": preprocess_wudao_gen(tokenizer),
-        "pile": preprocess_the_pile_gen(tokenizer),
-    }
-    data_set = DataIter(
-        paths, transform_dict=transform_dict, concat_docs=True, max_length=1024
-    )
-    for sample in data_set:
-        print(sample)
-        break

dataset/tokenizer.py

@@ -1,218 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-20 21:39:47
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-06 23:01:50
-FilePath: /Open-Llama/dataset/tokenizer.py
-Description: 
-
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-import torch
-
-
-class Tokenizer:
-    def __init__(self, sp_model):
-        self.sp_model = sp_model
-        self.bos_id = self.sp_model.bos_id()
-        self.eos_id = self.sp_model.eos_id()
-        self.pad_id = self.sp_model.pad_id()
-        self.vocab_size = self.sp_model.vocab_size()
-
-    def __call__(
-        self,
-        inputs,
-        padding=None,
-        max_length=256,
-        return_tensors=False,
-        truncation=False,
-        add_special_tokens=True,
-        return_mask=False,
-    ):
-        if isinstance(inputs, str):
-            return self.encode(
-                inputs,
-                padding=padding,
-                max_length=max_length,
-                return_tensors=return_tensors,
-                truncation=truncation,
-                add_special_tokens=add_special_tokens,
-                return_mask=return_mask,
-            )
-        else:
-            return self.encode_batch(
-                inputs,
-                padding=padding,
-                max_length=max_length,
-                return_tensors=return_tensors,
-                truncation=truncation,
-                add_special_tokens=add_special_tokens,
-                return_mask=return_mask,
-            )
-
-    def encode(
-        self,
-        inputs,
-        padding=None,
-        max_length=8192,
-        return_tensors=False,
-        truncation=False,
-        add_special_tokens=True,
-        return_mask=False,
-    ):
-        assert isinstance(inputs, str)
-        input_ids = self.sp_model.Encode(inputs)
-        if return_mask:
-            attention_mask = [1] * len(input_ids)
-        if truncation:
-            # https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L780
-            # 参考Transformer中的实现 默认最后一位一定是pad或者eos
-            input_ids = input_ids[: max_length - 1]
-            if return_mask:
-                attention_mask = attention_mask[: max_length - 1]
-        if add_special_tokens:
-            input_ids = input_ids + [self.eos_id]
-            if return_mask:
-                attention_mask = attention_mask + [0]
-        if padding == "max_length":
-            input_ids = input_ids + [self.pad_id] * (max_length - len(input_ids))
-            if return_mask:
-                attention_mask = attention_mask + [0] * (
-                    max_length - len(attention_mask)
-                )
-        if return_tensors:
-            input_ids = torch.tensor([input_ids])
-            out = {
-                "input_ids": input_ids,
-            }
-            if return_mask:
-                attention_mask = torch.tensor([attention_mask])
-                out["attention_mask"] = attention_mask
-        else:
-            out = {
-                "input_ids": input_ids,
-            }
-            if return_mask:
-                out["attention_mask"] = attention_mask
-        return out
-
-    def encode_batch(
-        self,
-        inputs,
-        padding=None,
-        max_length=8192,
-        return_tensors=False,
-        truncation=False,
-        add_special_tokens=True,
-        return_mask=False,
-    ):
-        input_ids = self.sp_model.Encode(inputs)
-        if return_mask:
-            attention_mask = [[1] * len(i) for i in input_ids]
-        if truncation:
-            input_ids = [i[: max_length - 1] for i in input_ids]
-            if return_mask:
-                attention_mask = [i[: max_length - 1] for i in attention_mask]
-        if add_special_tokens:
-            input_ids = [i + [self.eos_id] for i in input_ids]
-            if return_mask:
-                attention_mask = [i + [0] for i in attention_mask]
-        if padding == "max_length":
-            input_ids_pad = []
-            if return_mask:
-                attention_mask_pad = []
-            for idx, i in enumerate(input_ids):
-                input_ids_pad.append(i + [self.pad_id] * (max_length - len(i)))
-                if return_mask:
-                    j = attention_mask[idx]
-                    attention_mask_pad.append(j + [0] * (max_length - len(j)))
-            input_ids = input_ids_pad
-            if return_mask:
-                attention_mask = attention_mask_pad
-        if return_tensors:
-            input_ids = torch.tensor(input_ids)
-            out = {
-                "input_ids": input_ids,
-            }
-            if return_mask:
-                attention_mask = torch.tensor(attention_mask)
-                out["attention_mask"] = attention_mask
-        else:
-            out = {
-                "input_ids": input_ids,
-            }
-            if return_mask:
-                out["attention_mask"] = attention_mask
-        return out
-
-    def decode(self, inputs, max_rounds=None):
-        inputs = inputs.tolist()
-        out = []
-        for i, ids in enumerate(inputs):
-            count = 0
-            flag = False
-            for j, token in enumerate(ids):
-                if token == self.eos_id:
-                    if max_rounds is None:
-                        flag = True
-                        break
-                    elif isinstance(max_rounds, int):
-                        if count < max_rounds:
-                            count += 1
-                        else:
-                            flag = True
-                            break
-                    elif isinstance(max_rounds, list):
-                        if count < max_rounds[i]:
-                            count += 1
-                        else:
-                            flag = True
-                            break
-            if flag:
-                ids = ids[:j]
-            else:
-                ids = ids
-            out.append(ids)
-        out = self.sp_model.Decode(out)
-        return out
-
-
-if __name__ == "__main__":
-    import sentencepiece as spm
-    from unicodedata import normalize
-
-    # Using sentencepiece may not be able to process some reserved keywords like '▁'.
-    sp_model = spm.SentencePieceProcessor(
-        model_file="configs/10w_vocab_wudao5_pile10.model"
-    )
-    tokenizer = Tokenizer(sp_model)
-    tmp = [
-        "hello world",
-        "这是开源项目的V1版本，this is the first version of a open-source project!",
-        "# this is a python script\nfor i in range(10):\n   print(i)\n   for j in range(10):\n       print(j)",
-    ]
-    print(tmp)
-    out = tokenizer(
-        tmp, padding="max_length", return_tensors=True, max_length=64, truncation=True
-    )
-    for k, v in out.items():
-        print(k, v.shape)
-    print(out["input_ids"])
-    out = tokenizer.decode(out["input_ids"])
-    print(out)
-    for i, j in zip(tmp, out):
-        assert normalize("NFKC", i) == j
-
-    from dataset.data_iter import create_shard_kwargs, DataIter
-
-    patterns = ["data/pretrain_data/part-wudao*.jsonl.zst"]
-    paths = create_shard_kwargs(patterns)
-    data_iter = DataIter(paths)
-    for i, data in enumerate(data_iter):
-        assert (
-            normalize("NFKC", data["content"])
-            == sp_model.Decode(sp_model.Encode(data["content"]))
-            or "▁" in data["content"]
-        )
-        if i == 1000:
-            break

dataset/validation.py

@@ -1,12 +1,12 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-03-18 00:06:41
-LastEditors: LiangSong(sl12160010@gmail.com)
+LastEditors: s-JoL(sl12160010@gmail.com)
 LastEditTime: 2023-03-27 01:09:20
 FilePath: /Open-Llama/dataset/validation.py
 Description: 
 
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 val_set = [
     "白日依山尽，",

docker-compose.yml

@@ -0,0 +1,27 @@
+services:
+  app:
+    build: .
+    image: open-llama-image
+    container_name: open-llama-container
+    volumes:
+      - .:/app
+    #runtime: nvidia
+    deploy:
+      resources:
+        limits:
+          memory: 20G
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+          memory: 20G
+    memswap_limit: 60G
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
+      - CUDA_DEVICE_ORDER=PCI_BUS_ID
+      - CUDA_VISIBLE_DEVICES=0
+      - CUDA_LAUNCH_BLOCKING=1
+      - TORCH_USE_CUDA_DSA=1
+      - PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:256

inctruction_tuning.py

@@ -1,184 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-30 21:35:01
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-15 19:34:59
-FilePath: /Open-Llama/inctruction_tuning.py
-Description: 
-
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-import os
-import time
-import wandb
-import torch
-import random
-import sentencepiece as spm
-from torchinfo import summary
-from accelerate import Accelerator
-from torch.utils.data import DataLoader
-from deepspeed.ops.adam import FusedAdam
-from transformers import LlamaForCausalLM, LlamaConfig, get_cosine_schedule_with_warmup
-
-from dataset.validation import val_set
-from dataset.tokenizer import Tokenizer
-from dataset.data_iter import create_shard_kwargs, DataIter
-from dataset.collate_fn import collate_fn_gen
-from dataset.instruction_dataset import (
-    preprocess_belle_gen,
-    preprocess_self_instruction_gen,
-    preprocess_belle_multiturn_chat_gen,
-    preprocess_instruct_code_gen,
-    preprocess_sharegpt_gen,
-)
-from configs.instruction_tuning_config import *
-
-accelerator = Accelerator()
-
-if accelerator.is_main_process:
-    wandb.init(project="LLAMA Instruction")
-
-log_interval *= accelerator.gradient_accumulation_steps
-eval_interval *= accelerator.gradient_accumulation_steps
-save_interval *= accelerator.gradient_accumulation_steps
-
-sp_model = spm.SentencePieceProcessor(model_file=tokenizer_model_path)
-tokenizer = Tokenizer(sp_model)
-
-paths = create_shard_kwargs(patterns, repeat=3)
-random.shuffle(paths)
-transform_dict = {
-    "self_instruct": preprocess_self_instruction_gen(tokenizer),
-    "belle_1M": preprocess_belle_gen(tokenizer),
-    "belle_0.5M": preprocess_belle_gen(tokenizer),
-    "belle_school_math_0.25M": preprocess_belle_gen(tokenizer),
-    "belle_multiturn_chat_0.8M": preprocess_belle_multiturn_chat_gen(tokenizer),
-    "instruct_to_code": preprocess_instruct_code_gen(tokenizer),
-    "sharegpt_90K": preprocess_sharegpt_gen(tokenizer),
-}
-data_set = DataIter(
-    paths,
-    transform_dict=transform_dict,
-    concat_docs=False,
-    process_index=accelerator.process_index,
-    num_processes=accelerator.num_processes,
-)
-train_loader = DataLoader(
-    data_set,
-    batch_size=train_batch_size,
-    # If num_workers is greater than 1, duplicate data may occur.
-    num_workers=0,
-    collate_fn=collate_fn_gen(tokenizer, max_length),
-    drop_last=True,
-)
-# smaller initializer_range make training more stable
-# add stabel embedding to token embedding
-raw_model = LlamaForCausalLM(
-    LlamaConfig(
-        vocab_size=tokenizer.vocab_size,
-        initializer_range=initializer_range,
-        pad_token_id=tokenizer.pad_id,
-        rms_norm_eps=1e-5,
-        hidden_dropout_prob=0.1,
-        attention_dropout_prob=0.1,
-        use_stable_embedding=True,
-        shared_input_output_embedding=True,
-    )
-)
-ckpt = torch.load(ckpt_path, map_location="cpu")
-raw_model.load_state_dict(ckpt)
-raw_model.eval()
-with torch.no_grad():
-    summary(raw_model.cuda(), input_data=torch.ones(1, 64, dtype=torch.int64).cuda())
-no_decay = ["bias", "LayerNorm.weight", "layernorm.weight"]
-optimizer_grouped_parameters = [
-    {
-        "params": [
-            p
-            for n, p in raw_model.named_parameters()
-            if not any(nd in n for nd in no_decay)
-        ],
-        "weight_decay": weight_decay,
-    },
-    {
-        "params": [
-            p
-            for n, p in raw_model.named_parameters()
-            if any(nd in n for nd in no_decay)
-        ],
-        "weight_decay": 0.0,
-    },
-]
-optim = FusedAdam(optimizer_grouped_parameters, lr=lr, betas=(0.9, 0.95))
-optim.zero_grad()
-factor = accelerator.num_processes / accelerator.gradient_accumulation_steps
-scheduler = get_cosine_schedule_with_warmup(
-    optim,
-    num_warmup_steps=num_warmup_steps * factor,
-    num_training_steps=num_training_steps * factor,
-)
-
-_, model, optim, scheduler = accelerator.prepare(
-    train_loader, raw_model, optim, scheduler
-)
-print("start training...")
-train_loader_iter = iter(train_loader)
-global_step = 0
-start_time = time.time()
-for data_step in range(num_training_steps):
-    model.train()
-    with accelerator.accumulate(model):
-        batch = next(train_loader_iter)
-        for k, v in batch.items():
-            batch[k] = v.to(accelerator.device, non_blocking=True)
-        out = model(**batch, labels=batch["input_ids"])
-        total_loss = out.loss
-        losses = {"total_loss": total_loss}
-        accelerator.backward(total_loss)
-        optim.step()
-        scheduler.step()
-        optim.zero_grad()
-        if accelerator.sync_gradients:
-            global_step += 1
-    if data_step % log_interval == 0 and data_step > 0 and accelerator.is_main_process:
-        cost_time = time.time() - start_time
-        start_time = time.time()
-        tokens = train_batch_size * log_interval * max_length
-        wandb.log({"Training/Token per second per gpu": tokens / cost_time})
-        for k, v in losses.items():
-            wandb.log({"Losses/{}".format(k): v})
-        current_lr = optim.param_groups[0]["lr"]
-        wandb.log({"Training/LR": current_lr})
-        if optim.scaler is not None:
-            wandb.log({"Training/Loss Scale": optim.scaler.get_scale()})
-        wandb.log({"Training/Data Step": data_step})
-        wandb.log({"Training/Global Step": global_step})
-        accelerator.print(
-            "Global Step: {}, Data Step: {}, Loss: {}, Token per second per gpu: {}".format(
-                global_step, data_step, losses["total_loss"], tokens / cost_time
-            )
-        )
-    if data_step % eval_interval == 0 and accelerator.is_main_process:
-        text_table = wandb.Table(columns=["question", "pred"])
-        model.eval()
-        with torch.no_grad():
-            for data in val_set:
-                raw_inputs = data
-                inputs_len = len(raw_inputs)
-                inputs = tokenizer(
-                    raw_inputs, return_tensors=True, add_special_tokens=False
-                )
-                for k, v in inputs.items():
-                    inputs[k] = v.to(accelerator.device)
-                pred = model.generate(
-                    **inputs, max_new_tokens=256, do_sample=True, repetition_penalty=2.0
-                )
-                pred = tokenizer.decode(pred.cpu())[0]
-                pred = pred[inputs_len:]
-                text_table.add_data(raw_inputs, pred)
-        wandb.log({"Predictions on {}".format(global_step): text_table})
-    if data_step % save_interval == 0 and data_step > 0 and accelerator.is_main_process:
-        if not os.path.isdir(work_dir):
-            os.mkdir(work_dir)
-        torch.save(raw_model.state_dict(), "{}/{}.pt".format(work_dir, global_step))
-wandb.finish()

models/llama.py

@@ -1,12 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-17 13:21:33
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-03-26 23:13:57
-FilePath: /Open-Llama/models/llama.py
-Description: 
-Building the Llama model proposed by Meta. https://arxiv.org/pdf/2302.13971.pdf
-Performance and effectiveness optimization based on the implementation in the Transformer library.
-https://github.com/Bayes-Song/transformers
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""

pretrain_llama.py

@@ -1,174 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-17 14:27:28
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-15 19:35:06
-FilePath: /Open-Llama/pretrain_llama.py
-Description: 
-pretrain GPT
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-import os
-import time
-import wandb
-import torch
-import random
-import sentencepiece as spm
-from torchinfo import summary
-from accelerate import Accelerator
-from torch.utils.data import DataLoader
-from deepspeed.ops.adam import FusedAdam
-from transformers import LlamaForCausalLM, LlamaConfig, get_cosine_schedule_with_warmup
-
-from dataset.validation import val_set
-from dataset.tokenizer import Tokenizer
-from dataset.data_iter import create_shard_kwargs, DataIter
-from dataset.collate_fn import collate_fn_gen
-from dataset.pretrain_dataset import (
-    preprocess_the_pile_gen,
-    preprocess_wudao_gen,
-)
-from configs.pretrain_config import *
-
-accelerator = Accelerator()
-
-if accelerator.is_main_process:
-    wandb.init(project="LLAMA Pretrain")
-
-log_interval *= accelerator.gradient_accumulation_steps
-eval_interval *= accelerator.gradient_accumulation_steps
-save_interval *= accelerator.gradient_accumulation_steps
-
-sp_model = spm.SentencePieceProcessor(model_file=tokenizer_model_path)
-tokenizer = Tokenizer(sp_model)
-
-paths = create_shard_kwargs(patterns)
-random.shuffle(paths)
-transform_dict = {
-    "wudao": preprocess_wudao_gen(tokenizer, max_length),
-    "pile": preprocess_the_pile_gen(tokenizer, max_length),
-}
-data_set = DataIter(
-    paths,
-    transform_dict=transform_dict,
-    concat_docs=False,
-    process_index=accelerator.process_index,
-    num_processes=accelerator.num_processes,
-)
-train_loader = DataLoader(
-    data_set,
-    batch_size=train_batch_size,
-    # If num_workers is greater than 1, duplicate data may occur.
-    num_workers=0,
-    collate_fn=collate_fn_gen(tokenizer, max_length),
-    drop_last=True,
-)
-# smaller initializer_range make training more stable
-# add stabel embedding to token embedding
-raw_model = LlamaForCausalLM(
-    LlamaConfig(
-        vocab_size=tokenizer.vocab_size,
-        initializer_range=initializer_range,
-        pad_token_id=tokenizer.pad_id,
-        rms_norm_eps=1e-5,
-        hidden_dropout_prob=0.1,
-        attention_dropout_prob=0.1,
-        use_stable_embedding=True,
-        shared_input_output_embedding=True,
-    )
-)
-raw_model.eval()
-with torch.no_grad():
-    summary(raw_model.cuda(), input_data=torch.ones(1, 64, dtype=torch.int64).cuda())
-no_decay = ["bias", "LayerNorm.weight", "layernorm.weight"]
-optimizer_grouped_parameters = [
-    {
-        "params": [
-            p
-            for n, p in raw_model.named_parameters()
-            if not any(nd in n for nd in no_decay)
-        ],
-        "weight_decay": weight_decay,
-    },
-    {
-        "params": [
-            p
-            for n, p in raw_model.named_parameters()
-            if any(nd in n for nd in no_decay)
-        ],
-        "weight_decay": 0.0,
-    },
-]
-optim = FusedAdam(optimizer_grouped_parameters, lr=lr, betas=(0.9, 0.95))
-optim.zero_grad()
-factor = accelerator.num_processes / accelerator.gradient_accumulation_steps
-scheduler = get_cosine_schedule_with_warmup(
-    optim,
-    num_warmup_steps=num_warmup_steps * factor,
-    num_training_steps=num_training_steps * factor,
-)
-
-_, model, optim, scheduler = accelerator.prepare(
-    train_loader, raw_model, optim, scheduler
-)
-print("start training...")
-train_loader_iter = iter(train_loader)
-global_step = 0
-start_time = time.time()
-for data_step in range(num_training_steps):
-    model.train()
-    with accelerator.accumulate(model):
-        batch = next(train_loader_iter)
-        for k, v in batch.items():
-            batch[k] = v.to(accelerator.device, non_blocking=True)
-        out = model(**batch, labels=batch["input_ids"])
-        total_loss = out.loss
-        losses = {"total_loss": total_loss}
-        accelerator.backward(total_loss)
-        optim.step()
-        scheduler.step()
-        optim.zero_grad()
-        if accelerator.sync_gradients:
-            global_step += 1
-    if data_step % log_interval == 0 and data_step > 0 and accelerator.is_main_process:
-        cost_time = time.time() - start_time
-        start_time = time.time()
-        tokens = train_batch_size * log_interval * max_length
-        wandb.log({"Training/Token per second per gpu": tokens / cost_time})
-        for k, v in losses.items():
-            wandb.log({"Losses/{}".format(k): v})
-        current_lr = optim.param_groups[0]["lr"]
-        wandb.log({"Training/LR": current_lr})
-        if optim.scaler is not None:
-            wandb.log({"Training/Loss Scale": optim.scaler.get_scale()})
-        wandb.log({"Training/Data Step": data_step})
-        wandb.log({"Training/Global Step": global_step})
-        accelerator.print(
-            "Global Step: {}, Data Step: {}, Loss: {}, Token per second per gpu: {}".format(
-                global_step, data_step, losses["total_loss"], tokens / cost_time
-            )
-        )
-    if data_step % eval_interval == 0 and accelerator.is_main_process:
-        text_table = wandb.Table(columns=["question", "pred"])
-        model.eval()
-        with torch.no_grad():
-            for data in val_set:
-                raw_inputs = data
-                inputs_len = len(raw_inputs)
-                inputs = tokenizer(
-                    raw_inputs, return_tensors=True, add_special_tokens=False
-                )
-                for k, v in inputs.items():
-                    inputs[k] = v.to(accelerator.device)
-                pred = model.generate(
-                    **inputs, max_new_tokens=256, do_sample=True, repetition_penalty=2.0
-                )
-                pred = tokenizer.decode(pred.cpu())[0]
-                pred = pred[inputs_len:]
-                text_table.add_data(raw_inputs, pred)
-        wandb.log({"Predictions on {}".format(global_step): text_table})
-    if data_step % save_interval == 0 and data_step > 0 and accelerator.is_main_process:
-        if not os.path.isdir(work_dir):
-            os.mkdir(work_dir)
-        torch.save(raw_model.state_dict(), "{}/{}.pt".format(work_dir, global_step))
-wandb.finish()

requirements.txt

@@ -3,7 +3,6 @@ torchvision
 torchaudio
 zstandard
 accelerate
-datasets
 wandb
 deepspeed
 absl-py
@@ -15,6 +14,8 @@ seaborn
 sentencepiece
 triton
 functorch==1.13.1
-xformers
+xformers==0.0.16
 gradio
-git+https://github.com/Bayes-Song/transformers.git
+peft
+transformers
+fsspec==2023.9.2

server.py

@@ -1,75 +0,0 @@
-"""
-Author: LiangSong(sl12160010@gmail.com)
-Date: 2023-03-31 13:26:15
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-06 03:45:44
-FilePath: /Open-Llama/server.py
-Description: 
-
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
-"""
-import torch
-import gradio as gr
-import sentencepiece as spm
-from dataset.tokenizer import Tokenizer
-from transformers import LlamaForCausalLM, LlamaConfig
-
-
-sp_model = spm.SentencePieceProcessor(
-    model_file="configs/10w_vocab_wudao5_pile10.model"
-)
-tokenizer = Tokenizer(sp_model)
-
-raw_model = LlamaForCausalLM(
-    LlamaConfig(
-        vocab_size=tokenizer.vocab_size,
-        initializer_range=0.01,
-        pad_token_id=tokenizer.pad_id,
-        rms_norm_eps=1e-5,
-        hidden_dropout_prob=0.1,
-        attention_dropout_prob=0.1,
-        use_stable_embedding=True,
-        shared_input_output_embedding=True,
-    )
-)
-ckpt = torch.load(
-    "data/saved_ckpt/instruction_tuning_3_epochs/37001.pt", map_location="cpu"
-)
-raw_model.load_state_dict(ckpt)
-raw_model.eval()
-model = raw_model.cuda()
-print("ready")
-
-
-def question_answer(prompt):
-    print(prompt)
-    raw_inputs = "user:{}\nsystem:".format(prompt)
-    inputs_len = len(raw_inputs)
-    inputs = tokenizer(raw_inputs, return_tensors=True, add_special_tokens=False)
-    for k, v in inputs.items():
-        inputs[k] = v.cuda()
-    pred = model.generate(**inputs, max_new_tokens=512, do_sample=True)
-    pred = tokenizer.decode(pred.cpu())[0]
-    pred = pred[inputs_len:]
-    print(pred)
-    return pred
-
-
-demo = gr.Interface(
-    fn=question_answer,
-    inputs="text",
-    outputs="text",
-    examples=[
-        "帮我写一封邮件，内容是咨询教授本学期量子力学课程的时间表？并且希望教授推荐一些相关书籍",
-        "情人节送女朋友什么礼物，预算500",
-        "我今天肚子有点不舒服，晚饭有什么建议么",
-        "可以总结一下小说三体的核心内容么？",
-        "Can you explain to me what quantum mechanics is and how it relates to quantum computing?",
-        "请帮我写一个AI驱动的幼儿教育APP的商业计划书",
-        "用python实现一个快速排序",
-    ],
-    title="Open-Llama",
-    description="不基于其他预训练模型，完全使用[Open-Llama](https://github.com/Bayes-Song/Open-Llama)项目从0开始训练的Instruct-GPT模型，总训练成本不超过2w美元。由于请求需要经Gradio进行转发，可能出现请求丢失的现象，当长时间无响应（如20s以上）可刷新重试。当前体验服务生成的所有内容都是由人工智能模型生成，我们对其生成内容的准确性、完整性和功能性不做任何保证，并且其生成的内容不代表我们的态度或观点。",
-    article="联系方式: sl12160010@gmail.com  对于该项目有任何意见和建议都欢迎联系我",
-).queue(concurrency_count=1)
-demo.launch(share=True)

solver/trainer.py

@@ -0,0 +1,230 @@
+"""
+Author: s-JoL(sl12160010@gmail.com)
+Date: 2023-04-24 20:05:21
+LastEditors: s-JoL(sl12160010@gmail.com)
+LastEditTime: 2023-05-08 22:51:42
+FilePath: /Open-Llama/solver/trainer.py
+Description: 
+
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
+"""
+import time
+import wandb
+import torch
+import logging
+from torchinfo import summary
+from deepspeed.ops.adam import FusedAdam
+from transformers import get_cosine_schedule_with_warmup
+
+from dataset.validation import val_set
+
+
+class Trainer:
+    def __init__(self, config, raw_model, train_loader, tokenizer, accelerator):
+        self.config = config
+        self.raw_model = raw_model
+        self.train_loader = train_loader
+        self.tokenizer = tokenizer
+        self.accelerator = accelerator
+        self.train_and_eval = config["train"].get("train_and_eval", False)
+        self.gradient_accumulation_steps = config["train"].get(
+            "gradient_accumulation_steps", 1
+        )
+        self.lr_scheduler_factor = (
+            accelerator.num_processes / accelerator.gradient_accumulation_steps
+        )
+        self.log_interval = (
+            self.config["log_interval"] * accelerator.gradient_accumulation_steps
+        )
+        self.eval_interval = (
+            self.config["eval_interval"] * accelerator.gradient_accumulation_steps
+        )
+        self.save_interval = (
+            self.config["save_interval"] * accelerator.gradient_accumulation_steps
+        )
+        self.work_dir = self.config["work_dir"]
+        # self.get_model_info()
+        if accelerator.is_main_process:
+            wandb.init(project=self.config["project_name"])
+
+    def get_model_info(self):
+        with torch.no_grad():
+            summary(
+                self.raw_model.cuda(),
+                input_data=torch.ones(1, 64, dtype=torch.int64).cuda(),
+            )
+
+    def get_optimizer(self):
+        no_decay = ["bias", "LayerNorm.weight", "layernorm.weight"]
+        if self.config["train"].get("use_lora", False):
+            optimizer_grouped_parameters = self.raw_model.parameters()
+        else:
+            optimizer_grouped_parameters = [
+                {
+                    "params": [
+                        p
+                        for n, p in self.raw_model.named_parameters()
+                        if not any(nd in n for nd in no_decay)
+                    ],
+                    "weight_decay": self.config["train"]["weight_decay"],
+                },
+                {
+                    "params": [
+                        p
+                        for n, p in self.raw_model.named_parameters()
+                        if any(nd in n for nd in no_decay)
+                    ],
+                    "weight_decay": 0.0,
+                },
+            ]
+        self.optim = FusedAdam(
+            optimizer_grouped_parameters,
+            lr=self.config["train"]["lr"],
+            betas=(0.9, 0.95),
+        )
+
+    def get_lr_scheduler(self):
+        self.scheduler = get_cosine_schedule_with_warmup(
+            self.optim,
+            num_warmup_steps=self.config["train"]["num_warmup_steps"]
+            * self.lr_scheduler_factor,
+            num_training_steps=self.config["train"]["num_training_steps"]
+            * self.lr_scheduler_factor,
+        )
+
+    def prepare(self):
+        (
+            _,
+            self.model,
+            self.optim,
+            self.scheduler,
+        ) = self.accelerator.prepare(
+            self.train_loader, self.raw_model, self.optim, self.scheduler
+        )
+        self.optim.zero_grad()
+        self.global_step = 0
+        try:
+            self.accelerator.load_state(self.work_dir)
+            self.global_step = self.scheduler.scheduler._step_count - 1
+            self.global_step = self.global_step // self.accelerator.num_processes
+            logging.warning("Restored ckpt from {}".format(self.work_dir))
+        except:
+            logging.warning("No ckpt found in {}".format(self.work_dir))
+        if self.global_step > 0:
+            skip_steps = self.global_step * self.gradient_accumulation_steps
+            logging.warning("Skiped {} steps.".format(skip_steps))
+            self.train_loader_skiped = self.accelerator.skip_first_batches(
+                self.train_loader, num_batches=skip_steps
+            )
+        else:
+            self.train_loader_skiped = self.train_loader
+        self.accelerator.wait_for_everyone()
+
+    def train_step(self, batch):
+        out = self.model(**batch)
+        total_loss = out.loss
+        losses = {"total_loss": total_loss}
+        self.accelerator.backward(total_loss)
+        self.optim.step()
+        self.scheduler.step()
+        self.optim.zero_grad()
+        return losses
+
+    def train(self):
+        self.get_optimizer()
+        self.get_lr_scheduler()
+        self.prepare()
+        self.start_time = time.time()
+        self.epoch = 0
+        self.data_step = 0
+        while True:
+            if self.data_step >= self.config["train"]["num_training_steps"]:
+                break
+            if self.epoch == 0:
+                train_loader = self.train_loader_skiped
+            else:
+                train_loader = self.train_loader
+            for batch in train_loader:
+                # end training
+                if self.data_step >= self.config["train"]["num_training_steps"]:
+                    break
+                # data to device
+                for k, v in batch.items():
+                    batch[k] = v.to(self.accelerator.device, non_blocking=True)
+                self.model.train()
+                # train step
+                with self.accelerator.accumulate(self.model):
+                    losses = self.train_step(batch)
+                    if self.accelerator.sync_gradients:
+                        self.global_step += 1
+                # log
+                if (
+                    self.data_step % self.log_interval == 0
+                    and self.data_step > 0
+                    and self.accelerator.is_main_process
+                ):
+                    self.log(losses)
+                # eval/vis model output
+                if (
+                    self.data_step % self.eval_interval == 0
+                    and self.accelerator.is_main_process
+                    and self.train_and_eval
+                ):
+                    self.eval()
+                # save state
+                if self.data_step % self.save_interval == 0 and self.data_step > 0:
+                    self.accelerator.save_state(self.work_dir)
+                self.data_step += 1
+            self.epoch += 1
+        wandb.finish()
+
+    def log(self, losses):
+        cost_time = time.time() - self.start_time
+        self.start_time = time.time()
+        tokens = (
+            self.config["train"]["train_batch_size"]
+            * self.log_interval
+            * self.config["data"]["seq_length"]
+        )
+        wandb.log({"Training/Token per second per gpu": tokens / cost_time})
+        for k, v in losses.items():
+            wandb.log({"Losses/{}".format(k): v})
+        current_lr = self.optim.param_groups[0]["lr"]
+        wandb.log({"Training/LR": current_lr})
+        if self.optim.scaler is not None:
+            wandb.log({"Training/Loss Scale": self.optim.scaler.get_scale()})
+        wandb.log({"Training/Data Step": self.data_step})
+        wandb.log({"Training/Global Step": self.global_step})
+        wandb.log({"Training/Epoch": self.epoch})
+        self.accelerator.print(
+            "Epoch: {}, Global Step: {}, Data Step: {}, Loss: {}, Token per second per gpu: {}".format(
+                self.epoch,
+                self.global_step,
+                self.data_step,
+                losses["total_loss"],
+                tokens / cost_time,
+            )
+        )
+
+    def eval(self):
+        text_table = wandb.Table(columns=["question", "pred"])
+        self.model.eval()
+        with torch.no_grad():
+            for data in val_set:
+                raw_inputs = data
+                inputs = self.tokenizer(
+                    raw_inputs,
+                    return_tensors="pt",
+                    add_special_tokens=False,
+                    return_attention_mask=False,
+                )
+                input_length = inputs["input_ids"].shape[1]
+                for k, v in inputs.items():
+                    inputs[k] = v.to(self.accelerator.device)
+                pred = self.model.generate(
+                    **inputs, max_new_tokens=256, do_sample=True, repetition_penalty=2.0
+                )
+                pred = pred[0, input_length:]
+                pred = self.tokenizer.decode(pred.cpu(), skip_special_tokens=True)
+                text_table.add_data(raw_inputs, pred)
+        wandb.log({"Predictions on {}".format(self.global_step): text_table})

train_lm.py

@@ -0,0 +1,118 @@
+"""
+Author: s-JoL(sl12160010@gmail.com)
+Date: 2023-04-12 19:12:42
+LastEditors: s-JoL(sl12160010@gmail.com)
+LastEditTime: 2023-05-17 22:20:32
+FilePath: /Open-Llama/train_lm.py
+Description: 
+
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
+"""
+import yaml
+import math
+import logging
+from absl import app
+from absl import flags
+from accelerate import Accelerator
+from torch.utils.data import DataLoader
+from peft import LoraConfig, TaskType, get_peft_model
+from datasets.distributed import split_dataset_by_node
+from transformers import AutoConfig, AutoModelForCausalLM, LlamaTokenizer
+
+from dataset.dataset import construct_dataset
+from solver.trainer import Trainer
+
+FLAGS = flags.FLAGS
+flags.DEFINE_string("train_config", None, "Training config path")
+flags.DEFINE_string(
+    "model_config", "configs/model_configs/7B.json", "Model config path"
+)
+
+
+def main(argv):
+    with open(FLAGS.train_config, "r", encoding="utf-8") as fp:
+        config = yaml.load(fp, Loader=yaml.FullLoader)
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=config["train"].get(
+            "gradient_accumulation_steps", 1
+        )
+    )
+    tokenizer = LlamaTokenizer(
+        config["data"]["tokenizer_model_path"],
+        pad_token="<pad>",
+        add_bos_token=False,
+        add_eos_token=True,
+    )
+    data_config = config["data"]
+    if data_config.get("split_by_shard", False):
+        train_dataset = construct_dataset(
+            data_config, tokenizer, world_size=accelerator.num_processes
+        )
+    else:
+        train_dataset = construct_dataset(data_config, tokenizer)
+    train_dataset = split_dataset_by_node(
+        train_dataset,
+        rank=accelerator.process_index,
+        world_size=accelerator.num_processes,
+    )
+    train_loader = DataLoader(
+        train_dataset,
+        batch_size=config["train"]["train_batch_size"],
+        num_workers=config["train"]["train_num_workers"],
+        prefetch_factor=config["train"].get("prefetch_factor", 2),
+        pin_memory=True,
+    )
+    # smaller initializer_range make training more stable
+    # add stabel embedding to token embedding
+    model_config = AutoConfig.from_pretrained(FLAGS.model_config)
+    # Make the vocab size divisible by 16
+    # https://huggingface.co/docs/transformers/main_classes/deepspeed#how-to-choose-which-zero-stage-and-offloads-to-use-for-best-performance
+    # https://developer.nvidia.com/blog/optimizing-gpu-performance-tensor-cores/
+    # vocab_size = math.ceil(tokenizer.vocab_size / 16) * 16
+    # logging.warning(
+    #     "Round vocab_size from {} to {}.".format(tokenizer.vocab_size, vocab_size)
+    # )
+    vocab_size = tokenizer.vocab_size
+    model_config.vocab_size = vocab_size
+    model_config.pad_token_id = tokenizer.pad_token_id
+    # 使用AutoModel可以在Deepspeed.zero.Init()下正确的生效，而直接使用如OpenLlamaModel不能正确生效，导致浪费大量内存空间
+    # https://github.com/huggingface/accelerate/pull/932
+    if config["train"]["ckpt"] is not None:
+        raw_model = AutoModelForCausalLM.from_pretrained(
+            config["train"]["ckpt"], config=model_config
+        )
+        logging.warning("Loaded ckpt from: {}".format(config["train"]["ckpt"]))
+    else:
+        raw_model = AutoModelForCausalLM.from_config(model_config)
+    # lora
+    if config["train"].get("use_lora", False):
+        # gradient ckpt bug, https://github.com/huggingface/transformers/issues/23170
+        if hasattr(raw_model, "enable_input_require_grads"):
+            raw_model.enable_input_require_grads()
+        else:
+
+            def make_inputs_require_grad(module, input, output):
+                output.requires_grad_(True)
+
+            raw_model.get_input_embeddings().register_forward_hook(
+                make_inputs_require_grad
+            )
+        peft_config = LoraConfig(
+            task_type=TaskType.CAUSAL_LM,
+            target_modules=["q_proj", "v_proj"],
+            inference_mode=False,
+            r=1,
+            lora_alpha=32,
+            lora_dropout=0.1,
+        )
+        raw_model = get_peft_model(raw_model, peft_config)
+        raw_model.print_trainable_parameters()
+    if config["train"].get("gradient_checkpointing_enable", False):
+        raw_model.gradient_checkpointing_enable()
+    trainer = Trainer(config, raw_model, train_loader, tokenizer, accelerator)
+    trainer.train()
+
+
+if __name__ == "__main__":
+    app.run(main)

utils/convert_ckpt.py

@@ -0,0 +1,76 @@
+"""
+Author: s-JoL(sl12160010@gmail.com)
+Date: 2023-04-28 19:55:13
+LastEditors: s-JoL(sl12160010@gmail.com)
+LastEditTime: 2023-05-06 23:30:29
+FilePath: /Open-Llama/utils/convert_ckpt.py
+Description: 
+
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
+"""
+import torch
+import sentencepiece as spm
+
+
+sp_model = spm.SentencePieceProcessor(
+    model_file="configs/tokenizer_models/llama_tokenizer_extended.model"
+)
+merged_vocab_size = sp_model.vocab_size()
+ckpt = torch.load("data/llama_raw_ckpt/7B/consolidated.00.pth")
+
+raw_vocab_size, hidden_size = ckpt["tok_embeddings.weight"].shape
+extended_tok_embeddings = torch.randn(merged_vocab_size - raw_vocab_size, hidden_size)
+extended_tok_embeddings = extended_tok_embeddings * 0.001
+ckpt["tok_embeddings.weight"] = torch.cat(
+    [ckpt["tok_embeddings.weight"], extended_tok_embeddings], dim=0
+)
+
+extended_out_embeddings = torch.randn(merged_vocab_size - raw_vocab_size, hidden_size)
+extended_out_embeddings = extended_out_embeddings * 0.001
+ckpt["output.weight"] = torch.cat(
+    [ckpt["output.weight"], extended_out_embeddings], dim=0
+)
+
+rename_map = {
+    "tok_embeddings.weight": "model.embed_tokens.weight",
+    "norm.weight": "model.norm.weight",
+    "output.weight": "lm_head.weight",
+}
+
+for f, t in rename_map.items():
+    v = ckpt.pop(f)
+    ckpt[t] = v
+
+from_names = [
+    "layers.{}.attention.wq.weight",
+    "layers.{}.attention.wk.weight",
+    "layers.{}.attention.wv.weight",
+    "layers.{}.attention.wo.weight",
+    "layers.{}.feed_forward.w1.weight",
+    "layers.{}.feed_forward.w2.weight",
+    "layers.{}.feed_forward.w3.weight",
+    "layers.{}.attention_norm.weight",
+    "layers.{}.ffn_norm.weight",
+    "layers.{}.attention.inner_attention.rope.freqs",
+]
+
+to_names = [
+    "model.layers.{}.self_attn.q_proj.weight",
+    "model.layers.{}.self_attn.k_proj.weight",
+    "model.layers.{}.self_attn.v_proj.weight",
+    "model.layers.{}.self_attn.o_proj.weight",
+    "model.layers.{}.mlp.gate_proj.weight",
+    "model.layers.{}.mlp.down_proj.weight",
+    "model.layers.{}.mlp.up_proj.weight",
+    "model.layers.{}.input_layernorm.weight",
+    "model.layers.{}.post_attention_layernorm.weight",
+    "model.layers.{}.self_attn.rotary_emb.inv_freq",
+]
+
+for layer in range(32):
+    for f, t in zip(from_names, to_names):
+        f = f.format(layer)
+        t = t.format(layer)
+        v = ckpt.pop(f)
+        ckpt[t] = v
+torch.save(ckpt, "data/llama_raw_ckpt/7B/extended.pth")

utils/merge_tokenizer.py

@@ -0,0 +1,27 @@
+from tqdm import tqdm
+import sentencepiece as spm
+from sentencepiece import sentencepiece_model_pb2 as model
+
+raw_model = model.ModelProto()
+raw_model.ParseFromString(
+    open("configs/tokenizer_models/llama_tokenizer.model", "rb").read()
+)
+
+exist_pieces = set([p.piece for p in raw_model.pieces])
+cn_model = model.ModelProto()
+cn_model.ParseFromString(
+    open("configs/tokenizer_models/4w_cn_vocab_wudao15.model", "rb").read()
+)
+
+for p in tqdm(cn_model.pieces, total=len(cn_model.pieces)):
+    if p.piece not in exist_pieces:
+        raw_model.pieces.append(p)
+
+with open("configs/tokenizer_models/llama_tokenizer_extended.model", "wb") as f:
+    f.write(raw_model.SerializeToString())
+
+sp_model = spm.SentencePieceProcessor(
+    model_file="configs/tokenizer_models/llama_tokenizer_extended.model"
+)
+
+print("merged vocab size: {}".format(sp_model.vocab_size()))

utils/speed_test/accelerate/run.py

@@ -1,12 +1,12 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-04-08 22:44:44
-LastEditors: LiangSong(sl12160010@gmail.com)
+LastEditors: s-JoL(sl12160010@gmail.com)
 LastEditTime: 2023-04-08 23:15:57
 FilePath: /Open-Llama/speed_test/accelerate/run.py
 Description: 
 
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 import time
 import torch
@@ -20,13 +20,15 @@ vocab_size = 32000
 total_step = 2
 use_activation_ckpt = True
 
+
 class FakeSet(torch.utils.data.Dataset):
     def __getitem__(self, idx):
-        return torch.randint(0, vocab_size, (seq_length, ))
-    
+        return torch.randint(0, vocab_size, (seq_length,))
+
     def __len__(self):
         return 1000000000
 
+
 accelerator = Accelerator()
 raw_model = LlamaForCausalLM(
     LlamaConfig(
@@ -39,15 +41,18 @@ optimizer = FusedAdam(raw_model.parameters(), lr=1e-5)
 
 train_loader = torch.utils.data.DataLoader(FakeSet(), batch_size=batch_size)
 if accelerator.distributed_type == DistributedType.FSDP:
-    accelerator.print('FSDP')
+    accelerator.print("FSDP")
     model = accelerator.prepare(raw_model)
     optimizer, train_loader = accelerator.prepare(optimizer, train_loader)
 else:
-    model, optimizer, train_loader = accelerator.prepare(raw_model, optimizer, train_loader)
+    model, optimizer, train_loader = accelerator.prepare(
+        raw_model, optimizer, train_loader
+    )
+
 
 def train(model, optimizer, train_loader):
     start_time = time.time()
-    for i,  batch in enumerate(train_loader):
+    for i, batch in enumerate(train_loader):
         if i == total_step:
             break
         optimizer.zero_grad()
@@ -58,4 +63,5 @@ def train(model, optimizer, train_loader):
     end_time = time.time()
     return end_time - start_time
 
-accelerator.print('total time: {}'.format(train(model, optimizer, train_loader)))
+
+accelerator.print("total time: {}".format(train(model, optimizer, train_loader)))

utils/speed_test/accelerate/run.sh

@@ -1,12 +1,12 @@
 ###
- # @Author: LiangSong(sl12160010@gmail.com)
+ # @Author: s-JoL(sl12160010@gmail.com)
  # @Date: 2023-04-08 22:44:27
- # @LastEditors: LiangSong(sl12160010@gmail.com)
+ # @LastEditors: s-JoL(sl12160010@gmail.com)
  # @LastEditTime: 2023-04-11 21:58:43
  # @FilePath: /Open-Llama/speed_test/accelerate/run.sh
  # @Description: 
  # 
- # Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+ # Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 ### 
 total_gpu=8
 accelerate launch --config_file deepspeed_stage2.yaml --main_process_ip 127.0.0.1 --main_process_port 23335 --num_processes $total_gpu run.py

utils/speed_test/colossal-ai/run.py

@@ -1,12 +1,12 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-04-11 20:07:35
-LastEditors: LiangSong(sl12160010@gmail.com)
+LastEditors: s-JoL(sl12160010@gmail.com)
 LastEditTime: 2023-04-11 21:56:23
 FilePath: /Open-Llama/speed_test/colossal-ai/run.py
 Description: 
 
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 import os
 from functools import partial
@@ -23,7 +23,14 @@ from torch.nn.parallel import DistributedDataParallel as DDP
 import colossalai
 from colossalai.logging import disable_existing_loggers, get_dist_logger
 from colossalai.nn.optimizer import HybridAdam
-from colossalai.tensor import ColoParameter, ComputePattern, ComputeSpec, ProcessGroup, ReplicaSpec, ShardSpec
+from colossalai.tensor import (
+    ColoParameter,
+    ComputePattern,
+    ComputeSpec,
+    ProcessGroup,
+    ReplicaSpec,
+    ShardSpec,
+)
 from colossalai.utils import get_current_device
 from colossalai.zero import ColoInitContext, zero_model_wrapper, zero_optim_wrapper
 
@@ -35,7 +42,7 @@ def parse_args():
     parser.add_argument(
         "--distplan",
         type=str,
-        default='CAI_Gemini',
+        default="CAI_Gemini",
         help="The distributed plan [colossalai, zero1, zero2, torch_ddp, torch_zero].",
     )
     parser.add_argument(
@@ -47,14 +54,13 @@ def parse_args():
     parser.add_argument(
         "--placement",
         type=str,
-        default='cpu',
+        default="cpu",
         help="Placement Policy for Gemini. Valid when using colossalai as dist plan.",
     )
     parser.add_argument(
         "--shardinit",
-        action='store_true',
-        help=
-        "Shard the tensors when init the model to shrink peak memory size on the assigned device. Valid when using colossalai as dist plan.",
+        action="store_true",
+        help="Shard the tensors when init the model to shrink peak memory size on the assigned device. Valid when using colossalai as dist plan.",
     )
     parser.add_argument(
         "--batch_size",
@@ -105,7 +111,6 @@ def split_param_col_tp1d(param: ColoParameter, pg: ProcessGroup):
 
 
 class GPTLMLoss(nn.Module):
-
     def __init__(self):
         super().__init__()
         self.loss_fn = nn.CrossEntropyLoss()
@@ -114,7 +119,9 @@ class GPTLMLoss(nn.Module):
         shift_logits = logits[..., :-1, :].contiguous()
         shift_labels = labels[..., 1:].contiguous()
         # Flatten the tokens
-        return self.loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+        return self.loss_fn(
+            shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+        )
 
 
 def get_cpu_mem():
@@ -125,8 +132,8 @@ def get_gpu_mem():
     return torch.cuda.memory_allocated() / 1024**2
 
 
-def get_mem_info(prefix=''):
-    return f'{prefix}GPU memory usage: {get_gpu_mem():.2f} MB, CPU memory usage: {get_cpu_mem():.2f} MB'
+def get_mem_info(prefix=""):
+    return f"{prefix}GPU memory usage: {get_gpu_mem():.2f} MB, CPU memory usage: {get_cpu_mem():.2f} MB"
 
 
 def get_model_size(model: nn.Module):
@@ -142,11 +149,11 @@ def model_size_formatter(numel: int) -> str:
     MB_SIZE = 10**6
     KB_SIZE = 10**3
     if numel >= GB_SIZE:
-        return f'{numel / GB_SIZE:.1f}B'
+        return f"{numel / GB_SIZE:.1f}B"
     elif numel >= MB_SIZE:
-        return f'{numel / MB_SIZE:.1f}M'
+        return f"{numel / MB_SIZE:.1f}M"
     elif numel >= KB_SIZE:
-        return f'{numel / KB_SIZE:.1f}K'
+        return f"{numel / KB_SIZE:.1f}K"
     else:
         return str(numel)
 
@@ -154,7 +161,7 @@ def model_size_formatter(numel: int) -> str:
 def set_cpu_maximum_parallelism():
     conf_str = torch.__config__.parallel_info()
     inter_str = conf_str.split("hardware_concurrency() : ")[1]
-    max_concurrency = inter_str.split('\n')[0]
+    max_concurrency = inter_str.split("\n")[0]
     os.environ["OMP_NUM_THREADS"] = max_concurrency
     print(f"environmental variable OMP_NUM_THREADS is set to {max_concurrency}.")
 
@@ -170,7 +177,7 @@ def tensor_parallelize(model: torch.nn.Module, pg: ProcessGroup):
     for mn, module in model.named_modules():
         for pn, param in module.named_parameters(recurse=False):
             # NOTE() a param maybe shared by two modules
-            if hasattr(param, 'visited'):
+            if hasattr(param, "visited"):
                 continue
 
             # if shard init, then convert param to replica and use the dp-only ProcessGroup
@@ -179,22 +186,22 @@ def tensor_parallelize(model: torch.nn.Module, pg: ProcessGroup):
             param.set_process_group(pg)
 
             # shard it w.r.t tp pattern
-            if 'mlp.c_fc' in mn:
-                if 'weight' in pn or 'bias' in pn:
-                    split_param_col_tp1d(param, pg)    # colmn slice
+            if "mlp.c_fc" in mn:
+                if "weight" in pn or "bias" in pn:
+                    split_param_col_tp1d(param, pg)  # colmn slice
                     # keep the shape of the output from c_fc
                     param.compute_spec.set_output_replicate(False)
                 else:
                     param.set_dist_spec(ReplicaSpec())
-            elif 'mlp.c_proj' in mn:
-                if 'weight' in pn:
-                    split_param_row_tp1d(param, pg)    # row slice
+            elif "mlp.c_proj" in mn:
+                if "weight" in pn:
+                    split_param_row_tp1d(param, pg)  # row slice
                 else:
                     param.set_dist_spec(ReplicaSpec())
-            elif 'wte' in mn or 'wpe' in mn:
-                split_param_col_tp1d(param, pg)    # colmn slice
-            elif 'c_attn' in mn or 'c_proj' in mn:
-                split_param_col_tp1d(param, pg)    # colmn slice
+            elif "wte" in mn or "wpe" in mn:
+                split_param_col_tp1d(param, pg)  # colmn slice
+            elif "c_attn" in mn or "c_proj" in mn:
+                split_param_col_tp1d(param, pg)  # colmn slice
             else:
                 param.set_dist_spec(ReplicaSpec())
             param.visited = True
@@ -209,7 +216,13 @@ def main():
     args = parse_args()
 
     # if args.distplan not in ["colossalai", "torch_ddp", "torch_zero", "zero1", "zero2"]:
-    if args.distplan not in ["CAI_ZeRO1", "CAI_ZeRO2", "CAI_Gemini", "Pytorch_DDP", "Pytorch_ZeRO"]:
+    if args.distplan not in [
+        "CAI_ZeRO1",
+        "CAI_ZeRO2",
+        "CAI_Gemini",
+        "Pytorch_DDP",
+        "Pytorch_ZeRO",
+    ]:
         raise TypeError(f"{args.distplan} is error")
 
     # batch size per DP degree
@@ -221,14 +234,18 @@ def main():
 
     WARMUP_STEPS = 1
     assert WARMUP_STEPS < NUM_STEPS, "warmup steps should smaller than the total steps"
-    assert (NUM_STEPS - WARMUP_STEPS) % 2 == 1, "the number of valid steps should be odd to take the median"
-    PROF_FLAG = False    # The flag of profiling, False by default
+    assert (
+        NUM_STEPS - WARMUP_STEPS
+    ) % 2 == 1, "the number of valid steps should be odd to take the median"
+    PROF_FLAG = False  # The flag of profiling, False by default
 
     disable_existing_loggers()
     colossalai.launch_from_torch(config={})
 
     logger = get_dist_logger()
-    logger.info(f"{args.model_type}, {args.distplan}, batch size {BATCH_SIZE}", ranks=[0])
+    logger.info(
+        f"{args.model_type}, {args.distplan}, batch size {BATCH_SIZE}", ranks=[0]
+    )
 
     # build criterion
     criterion = GPTLMLoss()
@@ -244,10 +261,12 @@ def main():
             raise RuntimeError("You can only use shardinit with CAI_Gemini")
 
         # build GPT model
-        with ColoInitContext(device=get_current_device(),
-                             dtype=torch.half,
-                             default_dist_spec=default_dist_spec,
-                             default_pg=shard_pg):
+        with ColoInitContext(
+            device=get_current_device(),
+            dtype=torch.half,
+            default_dist_spec=default_dist_spec,
+            default_pg=shard_pg,
+        ):
             model = model_builder(VOCAB_SIZE, checkpoint=True)
 
         tp_pg = ProcessGroup(tp_degree=args.tp_degree)
@@ -259,15 +278,21 @@ def main():
         # asign running configurations
         gemini_config = None
         if args.distplan.startswith("CAI_ZeRO"):
-            optim_config = dict(reduce_bucket_size=12 * 1024 * 1024, overlap_communication=True, verbose=True)
+            optim_config = dict(
+                reduce_bucket_size=12 * 1024 * 1024,
+                overlap_communication=True,
+                verbose=True,
+            )
         elif args.distplan == "CAI_Gemini":
-            gemini_config = dict(strict_ddp_mode=args.tp_degree == 1,
-                                 device=get_current_device(),
-                                 placement_policy=args.placement,
-                                 pin_memory=True,
-                                 hidden_dim=model.model.config.hidden_size,
-                                 search_range_mb=128)
-            optim_config = dict(gpu_margin_mem_ratio=0.)
+            gemini_config = dict(
+                strict_ddp_mode=args.tp_degree == 1,
+                device=get_current_device(),
+                placement_policy=args.placement,
+                pin_memory=True,
+                hidden_dim=model.model.config.hidden_size,
+                search_range_mb=128,
+            )
+            optim_config = dict(gpu_margin_mem_ratio=0.0)
         else:
             raise RuntimeError
 
@@ -287,7 +312,7 @@ def main():
         model = zero_model_wrapper(model, zero_stage, gemini_config)
         optimizer = zero_optim_wrapper(model, optimizer, optim_config=optim_config)
 
-        logger.info(get_mem_info(prefix='After init optim, '), ranks=[0])
+        logger.info(get_mem_info(prefix="After init optim, "), ranks=[0])
     elif args.distplan.startswith("Pytorch"):
         assert args.tp_degree == 1, "The degree of TP should be 1 for DDP examples."
         model = model_builder(VOCAB_SIZE, checkpoint=True).cuda()
@@ -296,14 +321,17 @@ def main():
             optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
         elif args.distplan.endswith("ZeRO"):
             from torch.distributed.optim import ZeroRedundancyOptimizer
-            optimizer = ZeroRedundancyOptimizer(model.parameters(), optimizer_class=torch.optim.Adam, lr=1e-3)
+
+            optimizer = ZeroRedundancyOptimizer(
+                model.parameters(), optimizer_class=torch.optim.Adam, lr=1e-3
+            )
     else:
         raise RuntimeError
 
     # model is shared after TP
     numel = get_model_size(model)
     logger.info(f"the size of testing model size is {model_size_formatter(numel)}.")
-    logger.info(get_mem_info(prefix='After init model, '), ranks=[0])
+    logger.info(get_mem_info(prefix="After init model, "), ranks=[0])
 
     # Tflops_per_GPU = global_batch * global_numel * seq_len * 8 / #gpu
     # = (batch_per_DP_group * dp_degree) * (numel * tp_degree) * seq_len * 8 / (tp_degree * dp_degree)
@@ -325,7 +353,7 @@ def main():
         torch.cuda.synchronize()
         fwd_end = time()
         fwd_time = fwd_end - start
-        logger.info(get_mem_info(prefix=f'[{n + 1}/{NUM_STEPS}] Forward '), ranks=[0])
+        logger.info(get_mem_info(prefix=f"[{n + 1}/{NUM_STEPS}] Forward "), ranks=[0])
 
         if args.distplan.startswith("CAI"):
             optimizer.backward(loss)
@@ -337,13 +365,15 @@ def main():
         torch.cuda.synchronize()
         bwd_end = time()
         bwd_time = bwd_end - fwd_end
-        logger.info(get_mem_info(prefix=f'[{n + 1}/{NUM_STEPS}] Backward '), ranks=[0])
+        logger.info(get_mem_info(prefix=f"[{n + 1}/{NUM_STEPS}] Backward "), ranks=[0])
 
         optimizer.step()
         torch.cuda.synchronize()
         optim_time = time() - bwd_end
         step_time = time() - start
-        logger.info(get_mem_info(prefix=f'[{n + 1}/{NUM_STEPS}] Optimizer step '), ranks=[0])
+        logger.info(
+            get_mem_info(prefix=f"[{n + 1}/{NUM_STEPS}] Optimizer step "), ranks=[0]
+        )
 
         step_tflops = get_tflops_func(step_time)
         logger.info(
@@ -353,10 +383,12 @@ def main():
         if n >= WARMUP_STEPS:
             tflops_list.append(step_tflops)
 
-    demo_profiler = get_profile_context(PROF_FLAG,
-                                        WARMUP_STEPS,
-                                        NUM_STEPS - WARMUP_STEPS,
-                                        save_dir=f"profile/{get_time_stamp()}-demo")
+    demo_profiler = get_profile_context(
+        PROF_FLAG,
+        WARMUP_STEPS,
+        NUM_STEPS - WARMUP_STEPS,
+        save_dir=f"profile/{get_time_stamp()}-demo",
+    )
 
     with demo_profiler as prof:
         start_time = time()
@@ -364,7 +396,7 @@ def main():
             train_step()
             prof.step()
         end_time = time()
-        print('total time: {}'.format(end_time - start_time))
+        print("total time: {}".format(end_time - start_time))
 
     tflops_list.sort()
     median_index = ((NUM_STEPS - WARMUP_STEPS) >> 1) + WARMUP_STEPS
@@ -372,5 +404,5 @@ def main():
     torch.cuda.synchronize()
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

utils/speed_test/colossal-ai/utils.py

@@ -2,11 +2,15 @@ import time
 from contextlib import nullcontext
 
 import torch
-from torch.profiler import ProfilerActivity, profile, schedule, tensorboard_trace_handler
+from torch.profiler import (
+    ProfilerActivity,
+    profile,
+    schedule,
+    tensorboard_trace_handler,
+)
 
 
 class DummyProfiler:
-
     def __init__(self):
         self.step_number = 0
 
@@ -16,7 +20,9 @@ class DummyProfiler:
 
 # Randomly Generated Data
 def get_data(batch_size, seq_len, vocab_size):
-    input_ids = torch.randint(0, vocab_size, (batch_size, seq_len), device=torch.cuda.current_device())
+    input_ids = torch.randint(
+        0, vocab_size, (batch_size, seq_len), device=torch.cuda.current_device()
+    )
     attention_mask = torch.ones_like(input_ids)
     return input_ids, attention_mask
 
@@ -27,15 +33,17 @@ def get_tflops(model_numel, batch_size, seq_len, step_time):
 
 def get_profile_context(enable_flag, warmup_steps, active_steps, save_dir):
     if enable_flag:
-        return profile(activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
-                       schedule=schedule(wait=0, warmup=warmup_steps, active=active_steps),
-                       on_trace_ready=tensorboard_trace_handler(save_dir),
-                       record_shapes=True,
-                       profile_memory=True)
+        return profile(
+            activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
+            schedule=schedule(wait=0, warmup=warmup_steps, active=active_steps),
+            on_trace_ready=tensorboard_trace_handler(save_dir),
+            record_shapes=True,
+            profile_memory=True,
+        )
     else:
         return nullcontext(DummyProfiler())
 
 
 def get_time_stamp():
     cur_time = time.strftime("%d-%H:%M", time.localtime())
-    return cur_time
+    return cur_time

utils/speed_test/lightning/run.py

@@ -1,12 +1,12 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-04-11 20:07:35
-LastEditors: LiangSong(sl12160010@gmail.com)
+LastEditors: s-JoL(sl12160010@gmail.com)
 LastEditTime: 2023-04-11 21:56:07
 FilePath: /Open-Llama/speed_test/lightning/run.py
 Description: 
 
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 import time
 import torch
@@ -22,13 +22,15 @@ vocab_size = 32000
 total_step = 100
 use_activation_ckpt = False
 
+
 class FakeSet(torch.utils.data.Dataset):
     def __getitem__(self, idx):
-        return torch.randint(0, vocab_size, (seq_length, ))
-    
+        return torch.randint(0, vocab_size, (seq_length,))
+
     def __len__(self):
         return 1000000000
 
+
 class SpeedTest(pl.LightningModule):
     def __init__(self):
         super().__init__()
@@ -45,7 +47,7 @@ class SpeedTest(pl.LightningModule):
         out = self.model(batch, labels=batch)
         loss = out.loss
         if self.start_time is None:
-            print('start')
+            print("start")
             self.start_time = time.time()
         return loss
 
@@ -53,23 +55,26 @@ class SpeedTest(pl.LightningModule):
         optimizer = FusedAdam(self.trainer.model.parameters(), lr=1e-5)
         return optimizer
 
+
 model = SpeedTest()
 train_loader = torch.utils.data.DataLoader(FakeSet(), batch_size=batch_size)
 
-strategy=DeepSpeedStrategy(
+strategy = DeepSpeedStrategy(
     stage=2,
     offload_optimizer=False,
     offload_parameters=False,
-    process_group_backend="nccl"
+    process_group_backend="nccl",
 )
 trainer = pl.Trainer(
-    limit_train_batches=total_step, 
+    limit_train_batches=total_step,
     max_epochs=1,
     devices=8,
     accelerator="gpu",
     strategy=strategy,
-    precision=16, 
-    enable_checkpointing=False)
+    precision=16,
+    enable_checkpointing=False,
+)
+
 
 def train(model, train_loader):
     start_time = time.time()
@@ -77,4 +82,5 @@ def train(model, train_loader):
     end_time = time.time()
     return end_time - model.start_time
 
-print('total time: {}'.format(train(model, train_loader)))
+
+print("total time: {}".format(train(model, train_loader)))

utils/train_tokenizer.py

@@ -1,33 +1,43 @@
 """
-Author: LiangSong(sl12160010@gmail.com)
+Author: s-JoL(sl12160010@gmail.com)
 Date: 2023-03-24 20:49:03
-LastEditors: LiangSong(sl12160010@gmail.com)
-LastEditTime: 2023-04-05 22:40:29
-FilePath: /Open-Llama/dataset/train_tokenizer.py
+LastEditors: s-JoL(sl12160010@gmail.com)
+LastEditTime: 2023-05-06 23:34:14
+FilePath: /Open-Llama/utils/train_tokenizer.py
 Description: 
 
-Copyright (c) 2023 by LiangSong(sl12160010@gmail.com), All Rights Reserved. 
+Copyright (c) 2023 by s-JoL(sl12160010@gmail.com), All Rights Reserved. 
 """
 import random
-from dataset.data_iter import DataIter, create_shard_kwargs
+from glob import glob
+from datasets import load_dataset
 
-wudao_patterns = [
-    "data/pretrain_data/part-wudao-*.jsonl.zst",
-]
-wudao_paths = create_shard_kwargs(wudao_patterns)
+
+random.seed(42)
+
+wudao_pattern = "data/pretrain_data/part-wudao-*.jsonl.zst"
+wudao_paths = glob(wudao_pattern)
 random.shuffle(wudao_paths)
 
-pile_patterns = [
-    "data/pretrain_data/part-pile-*.jsonl.zst",
-]
-pile_paths = create_shard_kwargs(pile_patterns)
+pile_pattern = "data/pretrain_data/part-pile-*.jsonl.zst"
+pile_paths = glob(pile_pattern)
 random.shuffle(pile_paths)
+
 paths = wudao_paths[:5] + pile_paths[:10]
-transform_dict = {
-    "wudao": lambda line: line["title"] + "\n" + line["content"],
-    "pile": lambda line: line["text"],
-}
-data_iter = iter(DataIter(paths, transform_dict))
+
+dataset = load_dataset("json", data_files=paths, split="train", streaming=True)
+dataset = dataset.shuffle(seed=42)
+
+
+def transform(dataset):
+    for line in dataset:
+        if "title" in line and "content" in line:
+            yield line["title"] + "\n" + line["content"]
+        else:
+            yield line["text"]
+
+
+data_iter = transform(dataset)
 
 import io
 import sentencepiece as spm
@@ -52,11 +62,12 @@ spm.SentencePieceTrainer.train(
     # reserve whitespace and \n and \t etc. for code generation
     allow_whitespace_only_pieces=True,
     remove_extra_whitespaces=False,
+    # Llama use identity instead of nfkc
     normalization_rule_name="nfkc",
 )
 
 # Serialize the model as file.
-with open("configs/10w_vocab_wudao5_pile10.model", "wb") as f:
+with open("configs/tokenizer_models/10w_vocab_wudao5_pile10.model", "wb") as f:
     f.write(model.getvalue())
 
 # Directly load the model from serialized model.