Skip to content

V0.4.dev #58

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 11 commits into from
Jun 4, 2024
Merged

V0.4.dev #58

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
b7c1f85
readme
chencyudel May 19, 2024
b45edf2
readme
chencyudel May 20, 2024
e8d03f9
readme
chencyudel May 20, 2024
13ad050
v0.4.2 refactor
chencyudel May 20, 2024
c866f64
reformat
chencyudel May 20, 2024
c9fd185
scripts
chencyudel May 20, 2024
44ddee1
add multinode launcher script
chencyudel May 20, 2024
a4ad4f3
readme
chencyudel May 20, 2024
2423b67
readme
chencyudel May 20, 2024
685696f
reformat
chencyudel Jun 4, 2024
a2a0bff
add chat template to tokenizer
chencyudel Jun 4, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 23 additions & 17 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,14 +46,19 @@


## News
🔥🔥🔥 [2024/05/20] We released **MFTCoder v0.4.2**, mainly for MFTCoder-accelerate. It supports **QLoRA + DeepSpeed Zero3** and **QLoRA + FSDP** as options allowing you training very large models. It now supports new models like Qwen2, Qwen2-MoE, Starcoder2, Gemma, etc.

🔥🔥🔥 [2024/01/30] The model [CodeFuse-DeepSeek-33B](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B) fine-tuned with MFTCoder ranks first in HuggingFace [Big Code Models LeaderBoard](https://huggingface.co/spaces/bigcode/bigcode-models-leaderboard)
🔥🔥🔥 [2024/05/20] Our paper [MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning](https://arxiv.org/abs/2311.02303) has been accepted by KDD2024.

🔥🔥🔥 [2024/01/17] We released MFTCoder v0.3.0, mainly for MFTCoder-accelerate. It now supports new models like Mixtral(MoE), DeepSeek-coder, chatglm3. It supports FSDP as an option. It also supports Self-paced Loss as a solution for convergence balance in Multitask Fine-tuning.
🔥🔥🔥 [2024/05/20] [CodeFuse-StarCoder2-15B](https://huggingface.co/codefuse-ai/CodeFuse-StarCoder2-15B) has been released, achieving a pass@1 (greedy decoding) score of 73.2% on HumanEval.

🔥🔥🔥 [2024/01/17] [CodeFuse-DeepSeek-33B](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B) has been released, achieving a pass@1 (greedy decoding) score of 78.7% on HumanEval. It lists as top-1 LLM on Bigcode Leardboard in terms of win-rate, the official result is going to be published later.
🔥🔥 [2024/01/30] The model [CodeFuse-DeepSeek-33B](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B) fine-tuned with MFTCoder ranks first in HuggingFace [Big Code Models LeaderBoard](https://huggingface.co/spaces/bigcode/bigcode-models-leaderboard)

🔥🔥🔥 [2024/01/17] [CodeFuse-Mixtral-8x7B](https://huggingface.co/codefuse-ai/CodeFuse-Mixtral-8X7B) has been released, achieving a pass@1 (greedy decoding) score of 56.1% on HumanEval.
🔥🔥 [2024/01/17] We released MFTCoder v0.3.0, mainly for MFTCoder-accelerate. It now supports new models like Mixtral(MoE), DeepSeek-coder, chatglm3. It supports FSDP as an option. It also supports Self-paced Loss as a solution for convergence balance in Multitask Fine-tuning.

🔥🔥 [2024/01/17] [CodeFuse-DeepSeek-33B](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B) has been released, achieving a pass@1 (greedy decoding) score of 78.7% on HumanEval. It lists as top-1 LLM on Bigcode Leardboard in terms of win-rate, the official result is going to be published later.

🔥🔥 [2024/01/17] [CodeFuse-Mixtral-8x7B](https://huggingface.co/codefuse-ai/CodeFuse-Mixtral-8X7B) has been released, achieving a pass@1 (greedy decoding) score of 56.1% on HumanEval.

🔥🔥 [2023/11/07] [MFTCoder Paper](https://arxiv.org/abs/2311.02303) has been released on Arxiv, which discloses technique details of multi-task-fine-tuning.

Expand All @@ -73,6 +78,7 @@
| **CodeFuse-DeepSeek-33B** | **78.7%** | 2024/01 |
| **CodeFuse-CodeLlama-34B** | **74.4%** | 2023/09 |
| **CodeFuse-CodeLlama-34B-4bits** | **73.8%** | 2023/09 |
| **CodeFuse-StarCoder2-15B** | **73.2%** | 2023/05 |
| WizardCoder-Python-34B-V1.0 | 73.2% | 2023/08 |
| GPT-4(zero-shot) | 67.0% | 2023/03 |
| PanGu-Coder2 15B | 61.6% | 2023/08 |
Expand All @@ -88,7 +94,7 @@


## Articles
[MFT Arxiv paper](https://arxiv.org/abs/2311.02303)
[MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning (KDD2024)](https://arxiv.org/abs/2311.02303)

## Introduction

Expand Down Expand Up @@ -125,13 +131,13 @@ The main components of this project include:


## Requirements
To begin, ensure that you have successfully installed CUDA (version >= 11.4, preferably 11.7) along with the necessary drivers. Additionally, make sure you have installed torch (version 2.0.1).
To begin, ensure that you have successfully installed CUDA (version >= 11.4, preferably 12.1) along with the necessary drivers. Additionally, make sure you have installed torch (version >= 2.1.0).

Next, we have provided an init_env.sh script to simplify the installation of required packages. Execute the following command to run the script:
```bash
sh init_env.sh
```
We highly recommend training with flash attention(version >= 2.1.0, preferably 2.3.6), please refer to the following link for installation instructions: https://github.com/Dao-AILab/flash-attention
We highly recommend training with flash attention(version >= 2.3.0), please refer to the following link for installation instructions: https://github.com/Dao-AILab/flash-attention


## Training
Expand All @@ -152,16 +158,16 @@ If you want to explore some new framework like atorch, you could check:
We are excited to release the following two CodeLLMs trained by MFTCoder, now available on both HuggingFace and ModelScope:


| Model | HuggingFace Links | ModelScope Links | Base Model | Num of examples trained | Batch Size | Seq Length |
|--------------------------------------|---------------------------------------------------------------------------|---------------------------------------------------------------------------------|----------------------|------|------------|------------|
| 🔥 CodeFuse-DeepSeek-33B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-DeepSeek-33B) | DeepSeek-coder-33B | 60万 | 80 | 4096 |
| 🔥 CodeFuse-Mixtral-8x7B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-Mixtral-8x7B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-Mixtral-8x7B) | Mixtral-8x7B | 60万 | 80 | 4096 |
| 🔥 CodeFuse-CodeLlama-34B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeLlama-34B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeLlama-34B) | CodeLlama-34b-Python | 60万 | 80 | 4096 |
| 🔥 CodeFuse-CodeLlama-34B-4bits | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeLlama-34B-4bits) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeLlama-34B-4bits) | CodeLlama-34b-Python | | | 4096 |
| 🔥 CodeFuse-StarCoder-15B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-StarCoder-15B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-StarCoder-15B) | StarCoder-15B | 60万 | 80 | 4096 |
| 🔥 CodeFuse-QWen-14B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-QWen-14B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-QWen-14B) | Qwen-14b | 110万 | 256 | 4096 |
| 🔥 CodeFuse-CodeGeex2-6B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeGeex2-6B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeGeex2-6B) | CodeGeex2-6B | 110万 | 256 | 4096 |

| Model | HuggingFace Links | ModelScope Links | Base Model | Num of examples trained | Batch Size | Seq Length |
|----------------------------------|---------------------------------------------------------------------------|---------------------------------------------------------------------------------|----------------------|-------------------------|------------|------------|
| 🔥 CodeFuse-DeepSeek-33B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-DeepSeek-33B) | DeepSeek-coder-33B | 600K | 80 | 4096 |
| 🔥 CodeFuse-Mixtral-8x7B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-Mixtral-8x7B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-Mixtral-8x7B) | Mixtral-8x7B | 600K | 80 | 4096 |
| 🔥 CodeFuse-CodeLlama-34B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeLlama-34B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeLlama-34B) | CodeLlama-34b-Python | 600K | 80 | 4096 |
| 🔥 CodeFuse-CodeLlama-34B-4bits | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeLlama-34B-4bits) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeLlama-34B-4bits) | CodeLlama-34b-Python | | | 4096 |
| 🔥 CodeFuse-StarCoder-15B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-StarCoder-15B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-StarCoder-15B) | StarCoder-15B | 600K | 80 | 4096 |
| 🔥 CodeFuse-QWen-14B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-QWen-14B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-QWen-14B) | Qwen-14b | 1.1 Million | 256 | 4096 |
| 🔥 CodeFuse-CodeGeex2-6B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeGeex2-6B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeGeex2-6B) | CodeGeex2-6B | 1.1 Million | 256 | 4096 |
| 🔥 CodeFuse-StarCoder2-15B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-StarCoder2-15B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-StarCoder2-15B) | Starcoder2-15B | 700K | 128 | 4096 |

## Datasets
We are also pleased to release two code-related instruction datasets, meticulously selected from a range of datasets to facilitate multitask training. Moving forward, we are committed to releasing additional instruction datasets covering various code-related tasks.
Expand Down
25 changes: 16 additions & 9 deletions README_cn.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,11 +45,17 @@


## 新闻
🔥🔥🔥 [2024/01/17] **MFTCoder-v0.3.0**发布。新增对Mixtral(MoE), DeepSeek等模型的支持；新增支持FSDP(Fully Sharded Data Parallel)；新增Self-paced Loss, 支持多任务收敛均衡。 感兴趣详见微信公众号CodeFuse的文章[MFTCoder 重磅升级v0.3.0发布](https://mp.weixin.qq.com/s/xI3f0iUKq9TIIKZ_kMtcQg)
🔥🔥🔥 [2024/05/20] **MFTCoder-v0.4.2**发布。新增支持**QLoRA+ DeepSpeed Zero3**, **QLoRA + FSDP**训练模式，可以更好的支持微调更大的模型，比如Qwen1.5-70B等。新增对Qwen2, Qwen2-MoE, Starcoder2, Gemma等模型的支持。

🔥🔥🔥 [2024/01/17] 开源了[CodeFuse-DeepSeek-33B](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B)模型，在HumanEval pass@1(greedy decoding)上可以达到78.7%。该模型在Big Code榜单的结果近期发布，请关注公众号获取最新信息。
🔥🔥🔥 [2024/05/20] 我们的论文 [MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning](https://arxiv.org/abs/2311.02303) 已被 KDD 2024 接收.

🔥🔥🔥 [2024/01/17] 开源了[CodeFuse-Mixtral-8x7B](https://huggingface.co/codefuse-ai/CodeFuse-Mixtral-8x7B)模型，在HumanEval pass@1(greedy decoding)上可以达到56.1%。感兴趣详见微信公众号CodeFuse的文章[MFTCoder提升Mixtral-8x7B混合专家模型的代码能力实践](https://mp.weixin.qq.com/s/xI3f0iUKq9TIIKZ_kMtcQg)
🔥🔥🔥 开源了[CodeFuse-StarCoder2-15B](https://huggingface.co/codefuse-ai/CodeFuse-StarCoder2-15B)模型，在HumanEval上可以达到73.2%，多代码语言能力均衡.

🔥🔥 [2024/01/17] **MFTCoder-v0.3.0**发布。新增对Mixtral(MoE), DeepSeek等模型的支持；新增支持FSDP(Fully Sharded Data Parallel)；新增Self-paced Loss, 支持多任务收敛均衡。 感兴趣详见微信公众号CodeFuse的文章[MFTCoder 重磅升级v0.3.0发布](https://mp.weixin.qq.com/s/xI3f0iUKq9TIIKZ_kMtcQg)

🔥🔥 [2024/01/17] 开源了[CodeFuse-DeepSeek-33B](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B)模型，在HumanEval pass@1(greedy decoding)上可以达到78.7%。该模型在Big Code榜单的结果近期发布，请关注公众号获取最新信息。

🔥🔥 [2024/01/17] 开源了[CodeFuse-Mixtral-8x7B](https://huggingface.co/codefuse-ai/CodeFuse-Mixtral-8x7B)模型，在HumanEval pass@1(greedy decoding)上可以达到56.1%。感兴趣详见微信公众号CodeFuse的文章[MFTCoder提升Mixtral-8x7B混合专家模型的代码能力实践](https://mp.weixin.qq.com/s/xI3f0iUKq9TIIKZ_kMtcQg)

🔥🔥 [2023/11/07] [MFTCoder论文](https://arxiv.org/abs/2311.02303)在Arxiv公布，介绍了多任务微调的技术细节。

Expand All @@ -69,6 +75,7 @@
| **CodeFuse-DeepSeek-33B** | **78.7%** | 2024/01 |
| **CodeFuse-CodeLlama-34B** | **74.4%** | 2023/09 |
| **CodeFuse-CodeLlama-34B-4bits** | **73.8%** | 2023/09 |
| **CodeFuse-StarCoder2-15B** | **73.2%** | 2023/05 |
| WizardCoder-Python-34B-V1.0 | 73.2% | 2023/08 |
| GPT-4(zero-shot) | 67.0% | 2023/03 |
| PanGu-Coder2 15B | 61.6% | 2023/08 |
Expand Down Expand Up @@ -118,12 +125,12 @@


## 环境
首先, 你需要将CUDA(>=11.4, 推荐11.7)及其相关驱动安装成功，并确保其工作正常, 并且安装基本的torch（>=2.0.0）
首先, 你需要将CUDA(>=11.4, 推荐12.1)及其相关驱动安装成功，并确保其工作正常, 并且安装基本的torch（>=2.1.0）
在requirements.txt下固定了几个主要的python包的版本，执行如下脚本即可：
```bash
sh init_env.sh
```
我们强烈建议您安装flash attention（>=2.1.0, 推荐2.3.6）, 安装请参考 https://github.com/Dao-AILab/flash-attention
我们强烈建议您安装flash attention（>=2.3.0）, 安装请参考 https://github.com/Dao-AILab/flash-attention

## 训练
如果你熟悉大模型训练的各种主流开源资源，例如 ```transformers```, ```DeepSpeed```, ```FSDP```等， 为了用开源项目快速上手高性能微调，我们建议您尝试：
Expand All @@ -145,11 +152,11 @@ sh init_env.sh
| 🔥🔥🔥 CodeFuse-DeepSeek-33B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-DeepSeek-33B) | DeepSeek-coder-33B | 60万 | 80 | 4096 |
| 🔥🔥🔥 CodeFuse-Mixtral-8x7B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-Mixtral-8x7B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-Mixtral-8x7B) | Mixtral-8x7B | 60万 | 80 | 4096 |
| 🔥🔥🔥 CodeFuse-CodeLlama-34B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeLlama-34B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeLlama-34B) | CodeLlama-34b-Python | 60万 | 80 | 4096 |
| 🔥🔥🔥 CodeFuse-CodeLlama-34B-4bits | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeLlama-34B-4bits) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeLlama-34B-4bits) | CodeLlama-34b-Python | | | 4096 |
| 🔥🔥🔥 CodeFuse-CodeLlama-34B-4bits | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeLlama-34B-4bits) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeLlama-34B-4bits) | CodeLlama-34b-Python | | | 4096 |
| 🔥🔥🔥 CodeFuse-StarCoder-15B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-StarCoder-15B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-StarCoder-15B) | StarCoder-15B | 60万 | 80 | 4096 |
| 🔥🔥🔥 CodeFuse-QWen-14B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-QWen-14B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-QWen-14B) | Qwen-14b | 110万 | 256 | 4096 |
| 🔥🔥🔥 CodeFuse-CodeGeex2-6B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeGeex2-6B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeGeex2-6B) | CodeGeex2-6B | 110万 | 256 | 4096 |

| 🔥🔥🔥 CodeFuse-QWen-14B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-QWen-14B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-QWen-14B) | Qwen-14b | 110万 | 256 | 4096 |
| 🔥🔥🔥 CodeFuse-CodeGeex2-6B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-CodeGeex2-6B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-CodeGeex2-6B) | CodeGeex2-6B | 110万 | 256 | 4096 |
| 🔥🔥🔥 CodeFuse-StarCoder2-15B | [h-link](https://huggingface.co/codefuse-ai/CodeFuse-StarCoder2-15B) | [m-link](https://modelscope.cn/models/codefuse-ai/CodeFuse-StarCoder2-15B) | Starcoder2-15B | 70万 | 128 | 4096 |



Expand Down
Loading