[v4] refactor ms-swift v4 (pipelines/arguments/swiftmixin/callback/tuner_plugin)

README.md

@@ -161,7 +161,7 @@ For more optional dependencies, you can refer to [here](https://github.com/model
 CUDA_VISIBLE_DEVICES=0 \
 swift sft \
     --model Qwen/Qwen2.5-7B-Instruct \
-    --train_type lora \
+    --tuner_type lora \
     --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
               'AI-ModelScope/alpaca-gpt4-data-en#500' \
               'swift/self-cognition#500' \
@@ -248,25 +248,28 @@ ms-swift also supports training and inference using Python. Below is pseudocode
 Training:
 
 ```python
-from swift import get_model_processor, get_template, Swift, load_dataset, EncodePreprocessor, Seq2SeqTrainer
+from peft import LoraConfig, get_peft_model
+from swift import get_model_processor, get_template, load_dataset, EncodePreprocessor
+from swift.trainers import Seq2SeqTrainer, Seq2SeqTrainingArguments
 # Retrieve the model and template, and add a trainable LoRA module
 model, tokenizer = get_model_processor(model_id_or_path, ...)
 template = get_template(tokenizer, ...)
-model = Swift.prepare_model(model, lora_config)
+lora_config = LoraConfig(...)
+model = get_peft_model(model, lora_config)
 
 # Download and load the dataset, and encode the text into tokens
 train_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 model
+training_args = Seq2SeqTrainingArguments(...)
 trainer = Seq2SeqTrainer(
     model=model,
     args=training_args,
-    data_collator=template.data_collator,
+    template=template,
     train_dataset=train_dataset,
     eval_dataset=val_dataset,
-    template=template,
 )
 trainer.train()
 ```
@@ -329,7 +332,7 @@ swift pt \
     --model Qwen/Qwen2.5-7B \
     --dataset swift/chinese-c4 \
     --streaming true \
-    --train_type full \
+    --tuner_type full \
     --deepspeed zero2 \
     --output_dir output \
     --max_steps 10000 \
@@ -341,7 +344,7 @@ Fine-tuning:
 CUDA_VISIBLE_DEVICES=0 swift sft \
     --model Qwen/Qwen2.5-7B-Instruct \
     --dataset AI-ModelScope/alpaca-gpt4-data-en \
-    --train_type lora \
+    --tuner_type lora \
     --output_dir output \
     ...
 ```
@@ -352,7 +355,7 @@ 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 \
+    --tuner_type lora \
     --output_dir output \
     ...
 ```
@@ -379,7 +382,7 @@ NPROC_PER_NODE=2 CUDA_VISIBLE_DEVICES=0,1 megatron sft \
     --load_safetensors true \
     --save_safetensors true \
     --dataset AI-ModelScope/alpaca-gpt4-data-zh \
-    --train_type lora \
+    --tuner_type lora \
     --save output \
     ...
 ```
@@ -404,7 +407,7 @@ 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 \
+    --tuner_type lora \
     --use_vllm true \
     --vllm_mode colocate \
     --dataset AI-MO/NuminaMath-TIR#10000 \

README_CN.md

@@ -155,7 +155,7 @@ pip install -e .
 CUDA_VISIBLE_DEVICES=0 \
 swift sft \
     --model Qwen/Qwen2.5-7B-Instruct \
-    --train_type lora \
+    --tuner_type lora \
     --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
               'AI-ModelScope/alpaca-gpt4-data-en#500' \
               'swift/self-cognition#500' \
@@ -236,25 +236,28 @@ ms-swift也支持使用python的方式进行训练和推理。下面给出训练
 
 训练：
 ```python
-from swift import get_model_processor, get_template, Swift, load_dataset, EncodePreprocessor, Seq2SeqTrainer
+from peft import LoraConfig, get_peft_model
+from swift import get_model_processor, get_template, load_dataset, EncodePreprocessor
+from swift.trainers import Seq2SeqTrainer, Seq2SeqTrainingArguments
 # 获取模型和template，并加入可训练的LoRA模块
 model, tokenizer = get_model_processor(model_id_or_path, ...)
 template = get_template(tokenizer, ...)
-model = Swift.prepare_model(model, lora_config)
+lora_config = LoraConfig(...)
+model = get_peft_model(model, lora_config)
 
 # 下载并载入数据集，并将文本encode成tokens
 train_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)
 
 # 进行训练
+training_args = Seq2SeqTrainingArguments(...)
 trainer = Seq2SeqTrainer(
     model=model,
     args=training_args,
-    data_collator=template.data_collator,
+    template=template,
     train_dataset=train_dataset,
     eval_dataset=val_dataset,
-    template=template,
 )
 trainer.train()
 ```
@@ -317,7 +320,7 @@ swift pt \
     --model Qwen/Qwen2.5-7B \
     --dataset swift/chinese-c4 \
     --streaming true \
-    --train_type full \
+    --tuner_type full \
     --deepspeed zero2 \
     --output_dir output \
     --max_steps 10000 \
@@ -329,7 +332,7 @@ swift pt \
 CUDA_VISIBLE_DEVICES=0 swift sft \
     --model Qwen/Qwen2.5-7B-Instruct \
     --dataset AI-ModelScope/alpaca-gpt4-data-zh \
-    --train_type lora \
+    --tuner_type lora \
     --output_dir output \
     ...
 ```
@@ -340,7 +343,7 @@ 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 \
+    --tuner_type lora \
     --output_dir output \
     ...
 ```
@@ -366,7 +369,7 @@ NPROC_PER_NODE=2 CUDA_VISIBLE_DEVICES=0,1 megatron sft \
     --load_safetensors true \
     --save_safetensors true \
     --dataset AI-ModelScope/alpaca-gpt4-data-zh \
-    --train_type lora \
+    --tuner_type lora \
     --save output \
     ...
 ```
@@ -391,7 +394,7 @@ 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 \
+    --tuner_type lora \
     --use_vllm true \
     --vllm_mode colocate \
     --dataset AI-MO/NuminaMath-TIR#10000 \

docs/source/BestPractices/GRPO-Code-Training.md

@@ -62,7 +62,7 @@ swift rlhf \
     --use_vllm true \
     --vllm_server_host 127.0.0.1 \
     --vllm_server_port 8000 \
-    --train_type lora \
+    --tuner_type lora \
     --lora_rank 16 \
     --lora_alpha 32 \
     --torch_dtype bfloat16 \
@@ -114,7 +114,7 @@ swift rlhf \
     --use_vllm true \
     --vllm_server_host 127.0.0.1 \
     --vllm_server_port 8000 \
-    --train_type lora \
+    --tuner_type lora \
     --torch_dtype bfloat16 \
     --dataset 'open-r1/verifiable-coding-problems-python-10k' \
     --load_from_cache_file true \

docs/source/BestPractices/GRPO-Multi-Modal-Training.md

@@ -126,7 +126,7 @@ swift rlhf \
     --vllm_mode server \
     --vllm_server_host 127.0.0.1 \
     --vllm_server_port 8000 \
-    --train_type full \
+    --tuner_type full \
     --torch_dtype bfloat16 \
     --dataset 'AI-ModelScope/clevr_cogen_a_train' \
     --load_from_cache_file true \
@@ -201,7 +201,7 @@ swift rlhf \
     --vllm_mode server \
     --vllm_server_host 127.0.0.1 \
     --vllm_server_port 8000 \
-    --train_type full \
+    --tuner_type full \
     --torch_dtype bfloat16 \
     --dataset 'AI-ModelScope/GEOQA_R1V_Train_8K' \
     --load_from_cache_file true \
@@ -269,7 +269,7 @@ swift rlhf \
     --vllm_mode server \
     --vllm_server_host 127.0.0.1 \
     --vllm_server_port 8000 \
-    --train_type full \
+    --tuner_type full \
     --torch_dtype bfloat16 \
     --dataset 'lmms-lab/multimodal-open-r1-8k-verified' \
     --load_from_cache_file true \

docs/source/BestPractices/GRPO.md

@@ -131,7 +131,7 @@ swift rlhf \
     --vllm_mode server \
     --vllm_server_host 127.0.0.1 \
     --vllm_server_port 8000 \
-    --train_type full \
+    --tuner_type full \
     --torch_dtype bfloat16 \
     --dataset 'zouxuhong/Countdown-Tasks-3to4#50000' \
     --load_from_cache_file true \

docs/source/BestPractices/MLLM-Registration.md

@@ -543,7 +543,7 @@ if __name__ == '__main__':
         template='my_qwen2_5_omni',
         load_from_cache_file=True,
         split_dataset_ratio=0.01,
-        train_type='lora',
+        tuner_type='lora',
         torch_dtype='bfloat16',
         attn_impl='flash_attn',
         padding_free=True,
@@ -589,7 +589,7 @@ swift sft \
               'swift/VideoChatGPT:all#2000' \
     --load_from_cache_file true \
     --split_dataset_ratio 0.01 \
-    --train_type lora \
+    --tuner_type lora \
     --torch_dtype bfloat16 \
     --attn_impl flash_attn \
     --padding_free true \

docs/source/BestPractices/NPU-support.md

@@ -153,7 +153,7 @@ Legend:
 
 ## 微调
 
-以下介绍LoRA的微调, 全参数微调设置参数`--train_type full`即可. **更多训练脚本**参考[这里](https://github.com/modelscope/ms-swift/tree/main/examples/ascend/train).
+以下介绍LoRA的微调, 全参数微调设置参数`--tuner_type full`即可. **更多训练脚本**参考[这里](https://github.com/modelscope/ms-swift/tree/main/examples/ascend/train).
 
 | 模型大小 | NPU数量 | deepspeed类型 | 最大显存占用量 |
 | -------- | ------- | ------------- | -------------- |
@@ -181,7 +181,7 @@ swift sft \
     --dataset AI-ModelScope/blossom-math-v2 \
     --split_dataset_ratio 0.01 \
     --num_train_epochs 5 \
-    --train_type lora \
+    --tuner_type lora \
     --output_dir output \
     --learning_rate 1e-4 \
     --gradient_accumulation_steps 16 \
@@ -206,7 +206,7 @@ swift sft \
     --dataset AI-ModelScope/blossom-math-v2 \
     --split_dataset_ratio 0.01 \
     --num_train_epochs 5 \
-    --train_type lora \
+    --tuner_type lora \
     --output_dir output \
     ...
 ```
@@ -227,7 +227,7 @@ swift sft \
     --dataset AI-ModelScope/blossom-math-v2 \
     --split_dataset_ratio 0.01 \
     --num_train_epochs 5 \
-    --train_type lora \
+    --tuner_type lora \
     --output_dir output \
     --deepspeed zero2 \
     ...
@@ -246,7 +246,7 @@ swift sft \
     --dataset AI-ModelScope/blossom-math-v2 \
     --split_dataset_ratio 0.01 \
     --num_train_epochs 5 \
-    --train_type lora \
+    --tuner_type lora \
     --output_dir output \
     --deepspeed zero3 \
     ...

docs/source/BestPractices/Qwen3-Best-Practice.md

@@ -142,7 +142,7 @@ swift infer \
 CUDA_VISIBLE_DEVICES=0 \
 swift sft \
     --model Qwen/Qwen3-8B \
-    --train_type lora \
+    --tuner_type lora \
     --dataset 'swift/Qwen3-SFT-Mixin#2000' \
               'swift/self-cognition:qwen3#600' \
     --load_from_cache_file true \
@@ -221,7 +221,7 @@ NPROC_PER_NODE=4 \
 CUDA_VISIBLE_DEVICES=0,1,2,3 \
 swift sft \
     --model Qwen/Qwen3-8B \
-    --train_type full \
+    --tuner_type full \
     --dataset '<your-dataset>' \
     --load_from_cache_file true \
     --split_dataset_ratio 0.01 \
@@ -292,7 +292,7 @@ NPROC_PER_NODE=8 \
 swift rlhf \
     --rlhf_type grpo \
     --model Qwen/Qwen3-8B \
-    --train_type full \
+    --tuner_type full \
     --dataset 'AI-MO/NuminaMath-TIR#5000' \
     --load_from_cache_file true \
     --torch_dtype bfloat16 \

docs/source/BestPractices/Qwen3-VL-Best-Practice.md

@@ -192,7 +192,7 @@ swift sft \
               'swift/VideoChatGPT:Generic#2000' \
     --load_from_cache_file true \
     --split_dataset_ratio 0.01 \
-    --train_type lora \
+    --tuner_type lora \
     --torch_dtype bfloat16 \
     --num_train_epochs 1 \
     --per_device_train_batch_size 1 \

docs/source/BestPractices/Rapidly-Training-VL-model.md

@@ -112,7 +112,7 @@ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
 swift sft \
     --model /path/to/new_vl_model \
     --model_type qwen2_5_vl \
-    --train_type full \
+    --tuner_type full \
     --dataset xxx  \
     --load_from_cache_file true \
     --split_dataset_ratio 0.01 \
@@ -149,7 +149,7 @@ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
 swift sft \
     --model /path/to/stage1_checkpoint \
     --model_type qwen2_5_vl \
-    --train_type full \
+    --tuner_type full \
     --dataset xxx \
     --load_from_cache_file true \
     --split_dataset_ratio 0.01 \

docs/source/Customization/Pluginization.md

@@ -7,7 +7,7 @@
 
 ## callback回调
 
-example在[这里](https://github.com/modelscope/ms-swift/blob/main/swift/plugins/callback.py).
+example在[这里](https://github.com/modelscope/ms-swift/blob/main/swift/callbacks).
 
 `callback`机制是transformers Trainer中的一种训练定制化机制。开发者可以在callback中控制训练流程。通常来说，callback的定制化类似下面的样子：
 ```python
@@ -114,7 +114,7 @@ example在[这里](https://github.com/modelscope/ms-swift/blob/main/swift/agent_
 
 ## 定制化tuner
 
-example在[这里](https://github.com/modelscope/ms-swift/blob/main/swift/plugins/tuner.py).
+example在[这里](https://github.com/modelscope/ms-swift/blob/main/swift/tuner_plugin).
 - 多模态模型对ViT部分使用全参数训练，LLM部分使用LoRA训练，参考[这里](https://github.com/modelscope/ms-swift/tree/main/examples/train/multimodal/lora_llm_full_vit)。
 - Phi4-multimodal，直接对其已有LoRA进行训练而不额外附加LoRA，参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/train/plugins/tuner_phi4_mm.sh)。
 

docs/source/GetStarted/Quick-start.md

@@ -36,7 +36,7 @@ ms-swift的安装请参考[安装文档](./SWIFT-installation.md)。
 CUDA_VISIBLE_DEVICES=0 \
 swift sft \
     --model Qwen/Qwen2.5-7B-Instruct \
-    --train_type lora \
+    --tuner_type lora \
     --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
               'AI-ModelScope/alpaca-gpt4-data-en#500' \
               'swift/self-cognition#500' \