Metadata-Version: 2.1
Name: moe_peft
Version: 2.0.1
Summary: An Efficient LLM Fine-Tuning Factory Optimized for MoE PEFT
Project-URL: Homepage, https://github.com/TUDB-Labs/MoE-PEFT
Project-URL: Bug Tracker, https://github.com/TUDB-Labs/MoE-PEFT/issues
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.8
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: torch<2.6.0,>=2.4.0
Requires-Dist: datasets
Requires-Dist: evaluate
Requires-Dist: accelerate
Requires-Dist: transformers<4.47.0,>=4.44.0
Requires-Dist: sentencepiece
Requires-Dist: huggingface_hub
Requires-Dist: scikit-learn
Requires-Dist: tiktoken

# MoE-PEFT: An Efficient LLM Fine-Tuning Factory for Mixture of Expert (MoE) Parameter-Efficient Fine-Tuning.
[![](https://github.com/TUDB-Labs/MoE-PEFT/actions/workflows/python-test.yml/badge.svg)](https://github.com/TUDB-Labs/MoE-PEFT/actions/workflows/python-test.yml)
[![](https://img.shields.io/github/stars/TUDB-Labs/MoE-PEFT?logo=GitHub&style=flat)](https://github.com/TUDB-Labs/MoE-PEFT/stargazers)
[![](https://img.shields.io/github/v/release/TUDB-Labs/MoE-PEFT?logo=Github)](https://github.com/TUDB-Labs/MoE-PEFT/releases/latest)
[![](https://img.shields.io/pypi/v/moe_peft?logo=pypi)](https://pypi.org/project/moe_peft/)
[![](https://img.shields.io/docker/v/mikecovlee/moe_peft?logo=Docker&label=docker)](https://hub.docker.com/r/mikecovlee/moe_peft/tags)
[![](https://img.shields.io/github/license/TUDB-Labs/MoE-PEFT)](http://www.apache.org/licenses/LICENSE-2.0)

MoE-PEFT is an open-source *LLMOps* framework built on [m-LoRA](https://github.com/TUDB-Labs/mLoRA). It is designed for high-throughput fine-tuning, evaluation, and inference of Large Language Models (LLMs) using techniques such as MoE + Others (like LoRA, DoRA). Key features of MoE-PEFT include:

- Concurrent fine-tuning, evaluation, and inference of multiple adapters with a shared pre-trained model.

- **MoE PEFT** optimization, mainly for [MixLoRA](https://github.com/TUDB-Labs/MixLoRA) and other MoLE implementation.

- Support for multiple PEFT algorithms and various pre-trained models.

- Seamless integration with the [HuggingFace](https://huggingface.co) ecosystem.

You can try MoE-PEFT with [Google Colab](https://colab.research.google.com/github/TUDB-Labs/MoE-PEFT/blob/main/misc/finetune-demo.ipynb) before local installation.

## Supported Platform

| OS      | Executor | Model Precision        | Quantization  | Flash Attention |
|---------|---------|------------------------|---------------|-----------------|
| Linux   | CUDA    | FP32, FP16, TF32, BF16 | 8bit and 4bit | &check;         |
| Windows | CUDA    | FP32, FP16, TF32, BF16 | 8bit and 4bit | -               |
| macOS   | MPS     | FP32, FP16, BF16       | &cross;       | &cross;         |
| All     | CPU     | FP32, FP16, BF16       | &cross;       | &cross;         |

You can use the `MOE_PEFT_EXECUTOR_TYPE` environment variable to force MoE-PEFT to use a specific executor. For example, if you want MoE-PEFT to run only on CPU, you can set `MOE_PEFT_EXECUTOR_TYPE=CPU` before importing `moe_peft`.

## Supported Pre-trained Models

|         | Model                                            | Model Size  |
|---------|--------------------------------------------------|-------------|
| &check; | [LLaMA 1/2](https://huggingface.co/meta-llama)   | 7B/13B/70B  |
| &check; | [LLaMA 3/3.1](https://huggingface.co/meta-llama) | 8B/70B      |
| &check; | [Yi 1/1.5](https://huggingface.co/01-ai)         | 6B/9B/34B   |
| &check; | [TinyLLaMA](https://huggingface.co/TinyLlama)    | 1.1B        |
| &check; | [Qwen 1.5/2](https://huggingface.co/Qwen)        | 0.5B ~ 72B  |
| &check; | [Gemma](https://huggingface.co/google)           | 2B/7B       |
| &check; | [Gemma 2](https://huggingface.co/google)         | 9B/27B      |
| &check; | [Mistral](https://huggingface.co/mistralai)      | 7B          |
| &check; | [Phi 1.5/2](https://huggingface.co/microsoft)    | 2.7B        |
| &check; | [Phi 3/3.5](https://huggingface.co/microsoft)    | 3.8B/7B/14B |
| &check; | [ChatGLM 1/2/3](https://huggingface.co/THUDM)    | 6B          |
| &check; | [GLM 4](https://huggingface.co/THUDM)            | 6B          |


## Supported PEFT Methods

|         | PEFT Methods                                             | Arguments*                                                |
|---------|----------------------------------------------------------|-----------------------------------------------------------|
| &check; | [MoLA](https://arxiv.org/abs/2402.08562)                 | `"routing_strategy": "mola", "num_experts": 8`            |
| &check; | [LoRAMoE](https://arxiv.org/abs/2312.09979)              | `"routing_strategy": "loramoe", "num_experts": 8`         |
| &check; | [MixLoRA](https://arxiv.org/abs/2404.15159)              | `"routing_strategy": "mixlora", "num_experts": 8`         |
| &check; | [LoRA](https://arxiv.org/abs/2106.09685)                 | `"r": 8, "lora_alpha": 16, "lora_dropout": 0.05`          |
| &check; | [QLoRA](https://arxiv.org/abs/2402.12354)                | See *Quantize Methods*                                    |
| &check; | [LoRA+](https://arxiv.org/abs/2402.12354)                | `"loraplus_lr_ratio": 20.0`                               |
| &check; | [DoRA](https://arxiv.org/abs/2402.09353)                 | `"use_dora": true`                                        |
| &check; | [rsLoRA](https://arxiv.org/abs/2312.03732)               | `"use_rslora": true`                                      |

*: Arguments of configuration file

### Notice of PEFT supports
1. MoE-PEFT supports specific optimized operators for these PEFT methods, which can effectively improve the computing performance during training, evaluation and inference. However, these operators may cause a certain degree of accuracy loss (less than 5%). You can disable the optimized operators by defining the `MOE_PEFT_EVALUATE_MODE` environment variable in advance.
2. Auxiliary Loss is not currently supported for MoE PEFT methods other than MixLoRA.
3. You can check detailed arguments of MixLoRA in [TUDB-Labs/MixLoRA](https://github.com/TUDB-Labs/MixLoRA).

## Supported Attention Methods

|         | Attention Methods                                            | Name           | Arguments*               |
|---------|--------------------------------------------------------------|----------------|--------------------------|
| &check; | [Scaled Dot Product](https://arxiv.org/abs/1706.03762)       | `"eager"`      | `--attn_impl eager`      |
| &check; | [Flash Attention 2](https://arxiv.org/abs/2307.08691)        | `"flash_attn"` | `--attn_impl flash_attn` |
| &check; | [Sliding Window Attention](https://arxiv.org/abs/2004.05150) | -              | `--sliding_window`       |

*: Arguments of `moe_peft.py`

MoE-PEFT only supports scaled-dot product attention (eager) by default. Additional requirements are necessary for flash attention.

For flash attention, manual installation of the following dependencies is required:

```bash
pip3 install ninja
pip3 install flash-attn==2.5.8 --no-build-isolation
```

If any attention method is not specified, flash attention is used if available.

## Supported Quantize Methods

|         | Quantize Methods      | Arguments*    |
|---------|-----------------------|---------------|
| &check; | Full Precision (FP32) | by default    |
| &check; | Tensor Float 32       | `--tf32`      |
| &check; | Half Precision (FP16) | `--fp16`      |
| &check; | Brain Float 16        | `--bf16`      |
| &check; | 8bit Quantize         | `--load_8bit` |
| &check; | 4bit Quantize         | `--load_4bit` |

*: Arguments of `moe_peft.py`

MoE-PEFT offers support for various model accuracy and quantization methods. By default, MoE-PEFT utilizes full precision (Float32), but users can opt for half precision (Float16) using `--fp16` or BrainFloat16 using `--bf16`. Enabling half precision reduces the model size by half, and for further reduction, quantization methods can be employed.

Quantization can be activated using `--load_4bit` for 4-bit quantization or `--load_8bit` for 8-bit quantization. However, when only quantization is enabled, MoE-PEFT utilizes Float32 for calculations. To achieve memory savings during training, users can combine quantization and half-precision modes.

To enable quantization support, please manually install `bitsandbytes`:

```bash
pip3 install bitsandbytes==0.43.1
```

It's crucial to note that regardless of the settings, **LoRA weights are always calculated and stored at full precision**. For maintaining calculation accuracy, MoE-PEFT framework mandates the use of full precision for calculations when accuracy is imperative.

For users with NVIDIA Ampere or newer GPU architectures, the `--tf32` option can be utilized to enable full-precision calculation acceleration.

## Offline Configuration

MoE-PEFT relies on **HuggingFace Hub** to download necessary models, datasets, etc. If you cannot access the Internet or need to deploy MoE-PEFT in an offline environment, please refer to the following guide.

1. Use `git-lfs` manually downloads models and datasets from [HuggingFace Hub](https://huggingface.co).
2. Set `--data_path` to the local path to datasets when executing `launch.py gen`.
3. Clone the [evaluate](https://github.com/huggingface/evaluate) code repository locally.
4. Set environment variable `MOE_PEFT_METRIC_PATH` to the local path to `metrics` folder of evaluate code repository.
5. Set `--base_model` to the local path to models when executing `launch.py run`.

Example of (4): `export MOE_PEFT_METRIC_PATH=/path-to-your-git-repo/evaluate/metrics`

## Known issues

 + Quantization with Qwen2 have no effect (same with transformers).
 + Applying quantization with DoRA will result in higher memory and computation cost (same with PEFT).
 + Sliding window attention with generate cache may product abnormal output.

## Installation

Please refer to [MoE-PEFT Install Guide](./Install.md).

## Quickstart

You can conveniently utilize MoE-PEFT via `launch.py`. The following example demonstrates a streamlined approach to training a dummy model with MoE-PEFT.

```bash
# Generating configuration
python launch.py gen --template lora --tasks ./tests/dummy_data.json

# Running the training task
python launch.py run --base_model TinyLlama/TinyLlama_v1.1

# Try with gradio web ui
python inference.py \
  --base_model TinyLlama/TinyLlama_v1.1 \
  --template alpaca \
  --lora_weights ./casual_0
```

For further detailed usage information, please refer to the `help` command:

```bash
python launch.py help
```

## MoE-PEFT

The `moe_peft.py` code is a starting point for finetuning on various datasets.

Basic command for finetuning a baseline model on the [Alpaca Cleaned](https://github.com/gururise/AlpacaDataCleaned) dataset:
```bash
# Generating configuration
python launch.py gen \
  --template lora \
  --tasks yahma/alpaca-cleaned

python moe_peft.py \
  --base_model meta-llama/Llama-2-7b-hf \
  --config moe_peft.json \
  --bf16
```

You can check the template finetune configuration in [templates](./templates/) folder.

For further detailed usage information, please use `--help` option:
```bash
python moe_peft.py --help
```

## Use Docker

Firstly, ensure that you have installed Docker Engine and NVIDIA Container Toolkit correctly.

After that, you can launch the container using the following typical command:

```
docker run --gpus all -it --rm mikecovlee/moe_peft
```

You can check all available tags from: [mikecovlee/moe_peft/tags](https://hub.docker.com/r/mikecovlee/moe_peft/tags)

Please note that this container only provides a proper environment to run MoE-PEFT. The codes of MoE-PEFT are not included.

## Copyright

This project is licensed under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0).
