Repository files navigation

ModelScope Community Website
中文   ｜   English  

Paper   ｜English Documentation   ｜  中文文档  

• Groups
• Introduction
• News
• Installation
• Quick Start
• Usage
• License
• Citation

You can contact us and communicate with us by adding our group:

Discord Group	WeChat Group

🍲ms-swift is a large model and multimodal large model fine-tuning and deployment framework provided by the ModelScope community. It now supports training (pre-training, fine-tuning, human alignment), inference, evaluation, quantization, and deployment for 600+ text-only large models and 300+ multimodal large models. Large models include: Qwen3, Qwen3-Next, InternLM3, GLM4.5, Mistral, DeepSeek-R1, Llama4, etc. Multimodal large models include: Qwen3-VL, Qwen3-Omni, Llava, InternVL3.5, MiniCPM-V-4, Ovis2.5, GLM4.5-V, DeepSeek-VL2, etc.

🍔 In addition, ms-swift integrates the latest training technologies, including Megatron parallelism techniques such as TP, PP, CP, EP to accelerate training, as well as numerous GRPO algorithm family reinforcement learning algorithms including: GRPO, DAPO, GSPO, SAPO, CISPO, RLOO, Reinforce++, etc. to enhance model intelligence. ms-swift supports a wide range of training tasks, including preference learning algorithms such as DPO, KTO, RM, CPO, SimPO, ORPO, as well as Embedding, Reranker, and sequence classification tasks. ms-swift provides full-pipeline support for large model training, including acceleration for inference, evaluation, and deployment modules using vLLM, SGLang, and LMDeploy, as well as model quantization using GPTQ, AWQ, BNB, and FP8 technologies.

Why Choose ms-swift?

• 🍎Model Types: Supports600+ text-only large models,300+ multimodal large models, and All-to-All full modality models from training to deployment full pipeline, with Day-0 support for popular models.
• Dataset Types: Built-in 150+ datasets for pre-training, fine-tuning, human alignment, multimodal and various other tasks, with support for custom datasets. Users only need to prepare datasets for one-click training.
• Hardware Support: Supports A10/A100/H100, RTX series, T4/V100, CPU, MPS, and domestic hardware Ascend NPU, etc.
• Lightweight Training: Supports lightweight fine-tuning methods such as LoRA, QLoRA, DoRA, LoRA+, LLaMAPro, LongLoRA, LoRA-GA, ReFT, RS-LoRA, Adapter, LISA, etc.
• Quantized Training: Supports training on BNB, AWQ, GPTQ, AQLM, HQQ, EETQ quantized models, requiring only 9GB training resources for 7B models.
• Memory Optimization: GaLore, Q-Galore, UnSloth, Liger-Kernel, Flash-Attention 2/3, andUlysses and Ring-Attention sequence parallelism techniques support, reducing memory consumption for long-text training.
• Distributed Training: Supports distributed data parallelism (DDP), device_map simple model parallelism, DeepSpeed ZeRO2 ZeRO3, FSDP/FSDP2, and Megatron distributed training technologies.
• 🍓Multimodal Training: Supports multimodal packing technology to improve training speed by 100%+, supports mixed modality data training with text, images, video and audio, and supports independent control of vit/aligner/llm.
• Agent Training: Supports Agent templates, allowing one dataset to be used for training different models.
• 🍊Training Tasks: Supports pre-training and instruction fine-tuning, as well as training tasks such as DPO, GKD, KTO, RM, CPO, SimPO, ORPO, and supportsEmbedding/Reranker and sequence classification tasks.
• 🥥Megatron Parallelism: Provides TP/PP/SP/CP/ETP/EP/VPP parallel strategies,MoE model acceleration up to 10x. Supports full-parameter and LoRA training methods for 250+ text-only large models and 100+ multimodal large models. Supports CPT/SFT/GRPO/DPO/KTO/RM training tasks.
• 🍉Reinforcement Learning: Built-inrich GRPO family algorithms, including GRPO, DAPO, GSPO, SAPO, CISPO, CHORD, RLOO, Reinforce++, etc. Supports synchronous and asynchronous vLLM engine inference acceleration, with extensible reward functions, multi-turn inference Schedulers, and environments through plugins.
• Full-Pipeline Capabilities: Covers the entire workflow of training, inference, evaluation, quantization, and deployment.
• UI Training: Provides Web-UI interface for training, inference, evaluation, and quantization, completing the full pipeline for large models.
• Inference Acceleration: Supports PyTorch, vLLM, SGLang, and LmDeploy inference acceleration engines, providing OpenAI interfaces for accelerating inference, deployment, and evaluation modules.
• Model Evaluation: Uses EvalScope as the evaluation backend, supporting 100+ evaluation datasets for evaluating text-only and multimodal models.
• Model Quantization: Supports quantization export for AWQ, GPTQ, FP8, and BNB. Exported models support inference acceleration using vLLM/SGLang/LmDeploy.

• 🎁 2025.11.14: Megatron GRPO is now available! Check out thedocs andexamples.
• 🎁 2025.11.04: Support forMcore-Bridge, making Megatron training as simple and easy to use as transformers.
• 🎁 2025.10.28: Rayhere.
• 🎁 2025.09.07: Added support for CHORD training algorithm. See thedocumentation.
• 🎁 2025.09.06: Ulysses can now be used with ring-attention, allowing sequences to be sharded into any number of chunks (no longer limited by the number of heads). The argument remains--sequence_parallel_size N.
• 🎁 2025.09.02: Megatron-SWIFT now supports multimodal model training. Documentation can be foundhere.
• 🎁 2025.08.12: SupportDynamic Fine-Tuning(DFT) in SFT training, use parameter--enable_dft_loss true. Training scripts can be foundhere.
• 🎁 2025.07.09: Megatron-SWIFT supports LoRA training. Compared to ms-swift, it achieves significant speedup on MoE models. Training scripts can be foundhere.
• 🎁 2025.06.23: Fine-tuning of reranker models is supported. Training scripts can be found here:Reranker.
• 🎁 2025.06.15: Support for GKD training on both pure text large models and multimodal models. Training scripts can be found here:Pure Text,Multimodal.

To install using pip:

pip install ms-swift -U

To install from source:

# pip install git+https://github.com/modelscope/ms-swift.gitgit clone https://github.com/modelscope/ms-swift.gitcd ms-swiftpip install -e.

Running Environment:

For more optional dependencies, you can refer tohere.

10 minutes of self-cognition fine-tuning of Qwen2.5-7B-Instruct on a single 3090 GPU:

# 22GBCUDA_VISIBLE_DEVICES=0 \swift sft \    --model Qwen/Qwen2.5-7B-Instruct \    --train_type lora \    --dataset'AI-ModelScope/alpaca-gpt4-data-zh#500' \'AI-ModelScope/alpaca-gpt4-data-en#500' \'swift/self-cognition#500' \    --torch_dtype bfloat16 \    --num_train_epochs 1 \    --per_device_train_batch_size 1 \    --per_device_eval_batch_size 1 \    --learning_rate 1e-4 \    --lora_rank 8 \    --lora_alpha 32 \    --target_modules all-linear \    --gradient_accumulation_steps 16 \    --eval_steps 50 \    --save_steps 50 \    --save_total_limit 2 \    --logging_steps 5 \    --max_length 2048 \    --output_dir output \    --system'You are a helpful assistant.' \    --warmup_ratio 0.05 \    --dataloader_num_workers 4 \    --model_author swift \    --model_name swift-robot

Tips:

• If you want to train with a custom dataset, you can refer tothis guide to organize your dataset format and specify--dataset <dataset_path>.
• The--model_author and--model_name parameters are only effective when the dataset includesswift/self-cognition.
• To train with a different model, simply modify--model <model_id/model_path>.
• By default, ModelScope is used for downloading models and datasets. If you want to use HuggingFace, simply specify--use_hf true.

After training is complete, use the following command to infer with the trained weights:

• Here,--adapters should be replaced with the last checkpoint folder generated during training. Since the adapters folder contains the training parameter fileargs.json, there is no need to specify--model,--system separately; Swift will automatically read these parameters. To disable this behavior, you can set--load_args false.

# Using an interactive command line for inference.CUDA_VISIBLE_DEVICES=0 \swift infer \    --adapters output/vx-xxx/checkpoint-xxx \    --streamtrue \    --temperature 0 \    --max_new_tokens 2048# merge-lora and use vLLM for inference accelerationCUDA_VISIBLE_DEVICES=0 \swift infer \    --adapters output/vx-xxx/checkpoint-xxx \    --streamtrue \    --merge_loratrue \    --infer_backend vllm \    --vllm_max_model_len 8192 \    --temperature 0 \    --max_new_tokens 2048

Finally, use the following command to push the model to ModelScope:

CUDA_VISIBLE_DEVICES=0 \swiftexport \    --adapters output/vx-xxx/checkpoint-xxx \    --push_to_hubtrue \    --hub_model_id'<your-model-id>' \    --hub_token'<your-sdk-token>' \    --use_hffalse

The Web-UI is azero-threshold training and deployment interface solution based on Gradio interface technology. For more details, you can checkhere.

SWIFT_UI_LANG=en swift web-ui

ms-swift also supports training and inference using Python. Below is pseudocode for training and inference. For more details, you can refer tohere.

Training:

# Retrieve the model and template, and add a trainable LoRA modulemodel,tokenizer=get_model_tokenizer(model_id_or_path, ...)template=get_template(model.model_meta.template,tokenizer, ...)model=Swift.prepare_model(model,lora_config)# Download and load the dataset, and encode the text into tokenstrain_dataset,val_dataset=load_dataset(dataset_id_or_path, ...)train_dataset=EncodePreprocessor(template=template)(train_dataset,num_proc=num_proc)val_dataset=EncodePreprocessor(template=template)(val_dataset,num_proc=num_proc)# Train the modeltrainer=Seq2SeqTrainer(model=model,args=training_args,data_collator=template.data_collator,train_dataset=train_dataset,eval_dataset=val_dataset,template=template,)trainer.train()

Inference:

# Perform inference using the native PyTorch engineengine=PtEngine(model_id_or_path,adapters=[lora_checkpoint])infer_request=InferRequest(messages=[{'role':'user','content':'who are you?'}])request_config=RequestConfig(max_tokens=max_new_tokens,temperature=temperature)resp_list=engine.infer([infer_request],request_config)print(f'response:{resp_list[0].choices[0].message.content}')

Here is a minimal example of training to deployment using ms-swift. For more details, you can check theexamples.

• If you want to use other models or datasets (including multimodal models and datasets), you only need to modify--model to specify the corresponding model's ID or path, and modify--dataset to specify the corresponding dataset's ID or path.
• By default, ModelScope is used for downloading models and datasets. If you want to use HuggingFace, simply specify--use_hf true.

Supported Training Methods:

Pre-training:

# 8*A100NPROC_PER_NODE=8 \CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \swift pt \    --model Qwen/Qwen2.5-7B \    --dataset swift/chinese-c4 \    --streamingtrue \    --train_type full \    --deepspeed zero2 \    --output_dir output \    --max_steps 10000 \    ...

Fine-tuning:

CUDA_VISIBLE_DEVICES=0 swift sft \    --model Qwen/Qwen2.5-7B-Instruct \    --dataset AI-ModelScope/alpaca-gpt4-data-en \    --train_type lora \    --output_dir output \    ...

RLHF:

CUDA_VISIBLE_DEVICES=0 swift rlhf \    --rlhf_type dpo \    --model Qwen/Qwen2.5-7B-Instruct \    --dataset hjh0119/shareAI-Llama3-DPO-zh-en-emoji \    --train_type lora \    --output_dir output \    ...

ms-swift supports using Megatron parallelism techniques to accelerate training, including large-scale cluster training and MoE model training. The following training methods are supported:

NPROC_PER_NODE=2 CUDA_VISIBLE_DEVICES=0,1 megatron sft \    --model Qwen/Qwen2.5-7B-Instruct \    --load_safetensorstrue \    --save_safetensorstrue \    --dataset AI-ModelScope/alpaca-gpt4-data-zh \    --train_type lora \    --save output \    ...

ms-swift supports a rich set of GRPO family algorithms:

CUDA_VISIBLE_DEVICES=0,1,2,3 NPROC_PER_NODE=4 \swift rlhf \    --rlhf_type grpo \    --model Qwen/Qwen2.5-7B-Instruct \    --train_type lora \    --use_vllmtrue \    --vllm_mode colocate \    --dataset AI-MO/NuminaMath-TIR#10000 \    --output_dir output \    ...

CUDA_VISIBLE_DEVICES=0 swift infer \    --model Qwen/Qwen2.5-7B-Instruct \    --streamtrue \    --infer_backend pt \    --max_new_tokens 2048# LoRACUDA_VISIBLE_DEVICES=0 swift infer \    --model Qwen/Qwen2.5-7B-Instruct \    --adapters swift/test_lora \    --streamtrue \    --infer_backend pt \    --temperature 0 \    --max_new_tokens 2048

CUDA_VISIBLE_DEVICES=0 swift app \    --model Qwen/Qwen2.5-7B-Instruct \    --streamtrue \    --infer_backend pt \    --max_new_tokens 2048

CUDA_VISIBLE_DEVICES=0 swift deploy \    --model Qwen/Qwen2.5-7B-Instruct \    --infer_backend vllm

CUDA_VISIBLE_DEVICES=0 swift sample \    --model LLM-Research/Meta-Llama-3.1-8B-Instruct \    --sampler_engine pt \    --num_return_sequences 5 \    --dataset AI-ModelScope/alpaca-gpt4-data-zh#5

CUDA_VISIBLE_DEVICES=0 swifteval \    --model Qwen/Qwen2.5-7B-Instruct \    --infer_backend lmdeploy \    --eval_backend OpenCompass \    --eval_dataset ARC_c

CUDA_VISIBLE_DEVICES=0 swiftexport \    --model Qwen/Qwen2.5-7B-Instruct \    --quant_bits 4 --quant_method awq \    --dataset AI-ModelScope/alpaca-gpt4-data-zh \    --output_dir Qwen2.5-7B-Instruct-AWQ

swiftexport \    --model<model-path> \    --push_to_hubtrue \    --hub_model_id'<model-id>' \    --hub_token'<sdk-token>'

This framework is licensed under theApache License (Version 2.0). For models and datasets, please refer to the original resource page and follow the corresponding License.

@misc{zhao2024swiftascalablelightweightinfrastructure,title={SWIFT:A Scalable lightWeight Infrastructure for Fine-Tuning},author={Yuze Zhao and Jintao Huang and Jinghan Hu and Xingjun Wang and Yunlin Mao and Daoze Zhang and Zeyinzi Jiang and Zhikai Wu and Baole Ai and Ang Wang and Wenmeng Zhou and Yingda Chen},year={2024},eprint={2408.05517},archivePrefix={arXiv},primaryClass={cs.CL},url={https://arxiv.org/abs/2408.05517},}