/swift

SWIFT (Scalable lightWeight Infrastructure for Fine-Tuning) is an extensible framwork designed to facilitate lightweight model fine-tuning.

Primary LanguagePythonApache License 2.0Apache-2.0

SWIFT(Scalable lightWeight Infrastructure for Fine-Tuning)



ModelScope Hub
δΈ­ζ–‡  ｜  English

πŸ“– Introduction

SWIFT (Scalable lightWeight Infrastructure for Fine-Tuning) is an extensible framwork designed to faciliate lightweight model fine-tuning and inference. It integrates implementations for various efficient fine-tuning methods, by embracing approaches that is parameter-efficient, memory-efficient, and time-efficient. SWIFT integrates seamlessly into ModelScope ecosystem and offers the capabilities to finetune various models, with a primary emphasis on LLMs and vision models. Additionally, SWIFT is fully compatible with PEFT, enabling users to leverage the familiar Peft interface to finetune ModelScope models.

Currently supported approches (and counting):

  1. LoRA: LORA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS
  2. QA-LoRA:Quantization-Aware Low-Rank Adaptation of Large Language Models.
  3. LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models
  4. Adapter: Parameter-Efficient Transfer Learning for NLP
  5. Prompt Tuning: Visual Prompt Tuning
  6. Side: Side-Tuning: A Baseline for Network Adaptation via Additive Side Networks
  7. ResTuning-Bypass
  8. ROME: Rank-One Editing of Encoder-Decoder Models
  9. All tuners offered on PEFT

Key features:

  1. By integrating the ModelScope library, models can be readily obatined via a model-id.
  2. Tuners provided by SWIFT can be combined together to allow exploration of multiple tuners on a model for best result.
  3. Support calling activate_adapter or deactivate_adapter or set_active_adapters to activate/deactivate tuners. User can inference with one model and multiple tuners in different threads independently.

Users can check the documentation of Swift to get detail tutorials.

πŸŽ‰ News

  • πŸ”₯ 2023.10.30: Support QA-LoRA and LongLoRA to decrease memory usage in training.
  • πŸ”₯ 2023.10.30: Support ROME(Rank One Model Editing) to add/modify knowledges, training is not needed!
  • πŸ”₯ 2023.10.27: Support for chatglm3 series models: chatglm3-6b-base, chatglm3-6b, chatglm3-6b-32k. The corresponding shell script can be found in scripts/chatglm3_6b_32k.
  • πŸ”₯ 2023.10.17: Supported int4, int8 models: qwen-7b-chat-int4, qwen-14b-chat-int4, qwen-vl-chat-int4, baichuan2-7b-chat-int4, baichuan2-13b-chat-int4, qwen-7b-chat-int8, qwen-14b-chat-int8. The corresponding shell script can be found at scripts/qwen_7b_chat_int4, scripts/qwen_14b_chat_int4, scripts/qwen_vl_chat_int4, scripts/qwen_7b_chat_int8, scripts/qwen_14b_chat_int8.
  • 2023.10.15: Supported ziya2-13b model series: ziya2-13b, ziya2-13b-chat. The corresponding shell script can be found at scripts/ziya2_13b_chat.
  • 2023.10.12: Supported mistral-7b model series: openbuddy-mistral-7b-chat, mistral-7b, mistral-7b-chat. The corresponding shell script can be found at scripts/openbuddy_mistral_7b_chat, scripts/mistral_7b_chat.
  • πŸ”₯ 2023.10.7: Supported DeepSpeed ZeRO-2, enabling LoRA (not just QLoRA) to run DDP on 2*A10. The corresponding shell script can be found at scripts/qwen_7b_chat/lora_ddp_ds/sft.sh.
  • πŸ”₯ 2023.9.25: Supported qwen-14b model series: qwen-14b, qwen-14b-chat. The corresponding shell script can be found at scripts/qwen_14b, scripts/qwen_14b_chat.
  • 2023.9.12: Supported training with MP+DDP to accelerate full-parameter fine-tuning speed. The corresponding shell script can be found at scripts/qwen_7b_chat/full_mp_ddp/sft.sh.

✨ LLM SFT Example

Press this link to view the detail documentation of these examples.

Basic Usage

Quickly fine-tune, infer with LLM, and build a Web-UI.

Run using Python

git clone https://github.com/modelscope/swift.git
cd swift
pip install .[llm]
# Experimental environment: A10, 3090, A100, ...
# 16GB GPU memory
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'

import torch

from swift.llm import (
    DatasetName, InferArguments, ModelType, SftArguments
)
from swift.llm.run import infer_main, sft_main, web_ui_main

model_type = ModelType.qwen_7b_chat_int4
sft_args = SftArguments(
    model_type=model_type,
    eval_steps=50,
    train_dataset_sample=2000,
    dataset=[DatasetName.leetcode_python_en],
    output_dir='output',
    gradient_checkpointing=True)
best_ckpt_dir = sft_main(sft_args)
print(f'best_ckpt_dir: {best_ckpt_dir}')
torch.cuda.empty_cache()
infer_args = InferArguments(
    ckpt_dir=best_ckpt_dir,
    load_args_from_ckpt_dir=True,
    stream=True,
    show_dataset_sample=5)
infer_main(infer_args)
torch.cuda.empty_cache()
web_ui_main(infer_args)

Run using Swift CLI

SFT:

# Experimental environment: A10, 3090, A100, ...
# 10GB GPU memory
CUDA_VISIBLE_DEVICES=0 swift sft --model_id_or_path qwen/Qwen-7B-Chat-Int4 --dataset blossom-math-zh

# Using DDP
# Experimental environment: 2 * 3090
# 2 * 10GB GPU memory
CUDA_VISIBLE_DEVICES=0,1 \
NPROC_PER_NODE=2 \
swift sft \
    --model_id_or_path qwen/Qwen-7B-Chat-Int4 \
    --dataset blossom-math-zh \

# Using custom dataset
CUDA_VISIBLE_DEVICES=0 swift sft --model_id_or_path qwen/Qwen-7B-Chat-Int4 --custom_train_dataset_path chatml.jsonl

Inference:

CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'

Web-UI:

CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'

Features

πŸ› οΈ Installation

SWIFT is running in Python environment. Please make sure your python version is higher than 3.8.

  • Install SWIFT by the pip command:
pip install ms-swift -U
  • Install SWIFT by source code(for running sft/infer examples), please run:
git clone https://github.com/modelscope/swift.git
cd swift
pip install -e .

SWIFT requires torch>=1.13.

  • Use SWIFT in our docker image:
docker pull registry.cn-hangzhou.aliyuncs.com/modelscope-repo/modelscope:ubuntu20.04-cuda11.8.0-py38-torch2.0.1-tf2.13.0-1.9.1

πŸš€ Getting Started

SWIFT supports multiple tuners, as well as tuners provided by PEFT. To use these tuners, simply call:

from swift import Swift, LoRAConfig
config = LoRAConfig(...)
model = Swift.prepare_model(model, config, extra_state_keys=['...'])

The code snippet above initialized the tuner randomly. The input model is an instance of torch.nn.Module, the config is a subclass instance of SwiftConfig or PeftConfig. extra_state_keys is the extra module weights(like the linear head) to be trained and stored in the output dir.

You may combine multiple tuners by:

from swift import Swift, LoRAConfig, PromptConfig
model = Swift.prepare_model(model, {'lora': LoRAConfig(...), 'prompt': PromptConfig(...)})

Call save_pretrained and push_to_hub after finetuning:

from swift import push_to_hub
model.save_pretrained('some-output-folder')
push_to_hub('my-group/some-repo-id-modelscope', 'some-output-folder', token='some-ms-token')

Assume my-group/some-repo-id-modelscope is the model-id in the hub, and some-ms-token is the token for uploading.

Using the model-id to do later inference:

from swift import Swift
model = Swift.from_pretrained(model, 'my-group/some-repo-id-modelscope')

Here shows a runnable example:

import os
import tempfile

# Please install modelscope by `pip install modelscope`
from modelscope import Model

from swift import LoRAConfig, SwiftModel, Swift, push_to_hub

tmp_dir = tempfile.TemporaryDirectory().name
if not os.path.exists(tmp_dir):
    os.makedirs(tmp_dir)


model = Model.from_pretrained('modelscope/Llama-2-7b-ms', device_map='auto')
lora_config = LoRAConfig(target_modules=['q_proj', 'k_proj', 'v_proj'])
model: SwiftModel = Swift.prepare_model(model, lora_config)
# Do some finetuning here
model.save_pretrained(tmp_dir)

push_to_hub('my-group/swift_llama2', output_dir=tmp_dir)
model = Model.from_pretrained('modelscope/Llama-2-7b-ms', device_map='auto')
model = SwiftModel.from_pretrained(model, 'my-group/swift_llama2', device_map='auto')

This is a example that uses transformers for model creation uses SWIFT for efficient tuning.

from swift import Swift, LoRAConfig, AdapterConfig, PromptConfig
from transformers import AutoModelForImageClassification

# init vit model
model = AutoModelForImageClassification.from_pretrained("google/vit-base-patch16-224")

# init lora tuner config
lora_config = LoRAConfig(
    r=10,  # the rank of the LoRA module
    target_modules=['query', 'key', 'value'],  # the modules to be replaced with the end of the module name
    merge_weights=False  # whether to merge weights
)

# init adapter tuner config
adapter_config = AdapterConfig(
    dim=768,  # the dimension of the hidden states
    hidden_pos=0,  # the position of the hidden state to passed into the adapter
    target_modules=r'.*attention.output.dense$',  # the modules to be replaced with regular expression
    adapter_length=10  # the length of the adapter length
)

# init prompt tuner config
prompt_config = PromptConfig(
    dim=768,  # the dimension of the hidden states
    target_modules=r'.*layer\.\d+$',  # the modules to be replaced with regular expression
    embedding_pos=0,    # the position of the embedding tensor
    prompt_length=10,   # the length of the prompt tokens
    attach_front=False  # Whether prompt is attached in front of the embedding
)

# create model with swift. In practice, you can use any of these tuners or a combination of them.
model = Swift.prepare_model(model, {"lora_tuner": lora_config, "adapter_tuner": adapter_config, "prompt_tuner": prompt_config})

# get the trainable parameters of model
model.get_trainable_parameters()
# 'trainable params: 838,776 || all params: 87,406,432 || trainable%: 0.9596273189597764'

You can use the features offered by Peft in SWIFT:

from swift import LoraConfig, Swift
from peft import TaskType
lora_config = LoraConfig(target_modules=['query', 'key', 'value'], task_type=TaskType.CAUSAL_LM)
model_wrapped = Swift.prepare_model(model, lora_config)

# or call from_pretrained to load weights in the modelhub
model_wrapped = Swift.from_pretrained(model, 'some-id-in-the-modelscope-modelhub')

The saving strategy between Swift tuners and Peft tuners are slightly different. You can name a tuner by:

model = Swift.prepare_model(model, {'default': LoRAConfig(...)})
model.save_pretrained('./output')

In the output dir, you will have a dir structure like this:

output
    |-- default
        |-- adapter_config.json
        |-- adapter_model.bin
    |-- adapter_config.json
    |-- adapter_model.bin

The config/weights stored in the output dir is the config of extra_state_keys and the weights of it. This is different from PEFT, which stores the weights and config of the default tuner.

πŸ” Learn More

License

This project is licensed under the Apache License (Version 2.0).