Initial commit

finetune_demo/configs/sft.yaml

+data_config:
+  train_file: train.jsonl
+  val_file: dev.jsonl
+  test_file: dev.jsonl
+  num_proc: 1
+max_input_length: 256
+max_output_length: 512
+training_args:
+  # see `transformers.Seq2SeqTrainingArguments`
+  output_dir: ./output
+  max_steps: 3000
+  # needed to be fit for the dataset
+  learning_rate: 5e-5
+  # settings for data loading
+  per_device_train_batch_size: 1
+  dataloader_num_workers: 16
+  remove_unused_columns: false
+  # settings for saving checkpoints
+  save_strategy: steps
+  save_steps: 500
+  # settings for logging
+  log_level: info
+  logging_strategy: steps
+  logging_steps: 10
+  # settings for evaluation
+  per_device_eval_batch_size: 16
+  evaluation_strategy: steps
+  eval_steps: 500
+  # settings for optimizer
+  # adam_epsilon: 1e-6
+  # uncomment the following line to detect nan or inf values
+  # debug: underflow_overflow
+  predict_with_generate: true
+  generation_config:
+    max_new_tokens: 512
+  # set your absolute deepspeed path here
+  deepspeed: configs/ds_zero_3.json

finetune_demo/finetune.py

+# -*- coding: utf-8 -*-
+import json
+import os
+import jieba
+import dataclasses as dc
+import functools
+from collections.abc import Callable, Mapping, Sequence
+from pathlib import Path
+from typing import Annotated, Any, Optional, Union
+import numpy as np
+import ruamel.yaml as yaml
+import torch
+import typer
+from datasets import Dataset, NamedSplit, Split
+from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
+from peft import PeftConfig, get_peft_config, get_peft_model
+from rouge_chinese import Rouge
+from torch import nn
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    EvalPrediction,
+    GenerationConfig,
+    PreTrainedTokenizer,
+    Seq2SeqTrainingArguments,
+)
+from transformers import DataCollatorForSeq2Seq as _DataCollatorForSeq2Seq
+from transformers import Seq2SeqTrainer as _Seq2SeqTrainer
+
+app = typer.Typer(pretty_exceptions_show_locals=False)
+
+
+class DataCollatorForSeq2Seq(_DataCollatorForSeq2Seq):
+    def __call__(self, features, return_tensors=None):
+        output_ids = ([feature['output_ids'] for feature in features] if 'output_ids' in features[0].keys() else None)
+        if output_ids is not None:
+            max_output_length = max(len(out) for out in output_ids)
+            if self.pad_to_multiple_of is not None:
+                max_output_length = (
+                        (
+                                max_output_length + self.pad_to_multiple_of - 1) //
+                        self.pad_to_multiple_of * self.pad_to_multiple_of
+                )
+            for feature in features:
+                remainder = [self.tokenizer.pad_token_id] * (
+                        max_output_length - len(feature['output_ids'])
+                )
+                if isinstance(feature['output_ids'], list):
+                    feature['output_ids'] = feature['output_ids'] + remainder
+                else:
+                    feature['output_ids'] = np.concatenate(
+                        [feature['output_ids'], remainder]
+                    ).astype(np.int64)
+        return super().__call__(features, return_tensors)
+
+
+class Seq2SeqTrainer(_Seq2SeqTrainer):
+    def prediction_step(
+            self,
+            model: nn.Module,
+            inputs: dict[str, Any],
+            prediction_loss_only: bool,
+            ignore_keys=None,
+            **gen_kwargs,
+    ) -> tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]:
+        if self.args.predict_with_generate:
+            output_ids = inputs.pop('output_ids')
+        input_ids = inputs['input_ids']
+        loss, generated_tokens, labels = super().prediction_step(
+            model, inputs, prediction_loss_only, ignore_keys, **gen_kwargs
+        )
+        generated_tokens = generated_tokens[:, input_ids.size()[1]:]
+        labels = output_ids
+        return loss, generated_tokens, labels
+
+
+@dc.dataclass
+class DataConfig(object):
+    train_file: Optional[str] = None
+    val_file: Optional[str] = None
+    test_file: Optional[str] = None
+    num_proc: Optional[int] = None
+
+    @property
+    def data_format(self) -> str:
+        return Path(self.train_file).suffix
+
+    @property
+    def data_files(self) -> dict[NamedSplit, str]:
+        return {
+            split: data_file
+            for split, data_file in zip(
+                [Split.TRAIN, Split.VALIDATION, Split.TEST],
+                [self.train_file, self.val_file, self.test_file],
+            )
+            if data_file is not None
+        }
+
+
+@dc.dataclass
+class FinetuningConfig(object):
+    data_config: DataConfig
+
+    max_input_length: int
+    max_output_length: int
+
+    training_args: Seq2SeqTrainingArguments = dc.field(
+        default_factory=lambda: Seq2SeqTrainingArguments(output_dir='./output')
+    )
+    peft_config: Optional[PeftConfig] = None
+
+    def __post_init__(self):
+        if not self.training_args.do_eval or self.data_config.val_file is None:
+            self.training_args.do_eval = False
+            self.training_args.evaluation_strategy = 'no'
+            self.data_config.val_file = None
+        else:
+            self.training_args.per_device_eval_batch_size = (
+                    self.training_args.per_device_eval_batch_size
+                    or self.training_args.per_device_train_batch_size
+            )
+
+    @classmethod
+    def from_dict(cls, **kwargs) -> 'FinetuningConfig':
+        training_args = kwargs.get('training_args', None)
+        if training_args is not None and not isinstance(
+                training_args, Seq2SeqTrainingArguments
+        ):
+            gen_config = training_args.get('generation_config')
+            # TODO: a bit hacky
+            if not isinstance(gen_config, GenerationConfig):
+                training_args['generation_config'] = GenerationConfig(
+                    **gen_config
+                )
+            kwargs['training_args'] = Seq2SeqTrainingArguments(**training_args)
+
+        data_config = kwargs.get('data_config')
+        if not isinstance(data_config, DataConfig):
+            kwargs['data_config'] = DataConfig(**data_config)
+
+        peft_config = kwargs.get('peft_config', None)
+        if peft_config is not None and not isinstance(peft_config, PeftConfig):
+            kwargs['peft_config'] = get_peft_config(config_dict=peft_config)
+        return cls(**kwargs)
+
+    @classmethod
+    def from_file(cls, path: Union[str, Path]) -> 'FinetuningConfig':
+        path = Path(path)
+        parser = yaml.YAML(typ='safe', pure=True)
+        parser.indent(mapping=2, offset=2, sequence=4)
+        parser.default_flow_style = False
+        kwargs = parser.load(path)
+        return cls.from_dict(**kwargs)
+
+
+from datasets import load_dataset, DatasetDict, NamedSplit
+from typing import Optional
+
+
+def _load_datasets(
+        data_dir: str,
+        data_format: str,
+        data_files: dict[NamedSplit, str],
+        num_proc: Optional[int],
+) -> DatasetDict:
+    if data_format == '.jsonl':
+        dataset_dct = load_dataset(
+            data_dir,
+            data_files=data_files,
+            split=None,
+            num_proc=num_proc,
+        )
+    else:
+        raise NotImplementedError(f"Cannot load dataset in the '{data_format}' format.")
+    return dataset_dct
+
+
+class DataManager(object):
+    def __init__(self, data_dir: str, data_config: DataConfig):
+        self._num_proc = data_config.num_proc
+
+        self._dataset_dct = _load_datasets(
+            data_dir,
+            data_config.data_format,
+            data_config.data_files,
+            self._num_proc,
+        )
+
+    def _get_dataset(self, split: NamedSplit) -> Optional[Dataset]:
+        return self._dataset_dct.get(split, None)
+
+    def get_dataset(
+            self,
+            split: NamedSplit,
+            process_fn: Callable[[dict[str, Any]], dict[str, Any]],
+            batched: bool = True,
+            remove_orig_columns: bool = True,
+    ) -> Optional[Dataset]:
+        orig_dataset = self._get_dataset(split)
+        if orig_dataset is None:
+            return
+
+        if remove_orig_columns:
+            remove_columns = orig_dataset.column_names
+        else:
+            remove_columns = None
+        return orig_dataset.map(
+            process_fn,
+            batched=batched,
+            remove_columns=remove_columns,
+            num_proc=self._num_proc,
+        )
+
+
+def process_message(message):
+    if 'tools' in message and message['role'] == 'system':
+        for tool in message['tools']:
+            parameters = tool['function']['parameters']['properties']
+            tool['function']['parameters']['properties'] = \
+                {k: v for k, v in parameters.items() if
+                 v is not None}
+    elif 'tools' in message:
+        del message['tools']
+    return message
+
+
+def process_batch(
+        batch: Mapping[str, Sequence],
+        tokenizer: PreTrainedTokenizer,
+        max_input_length: int,
+        max_output_length: int,
+) -> dict[str, list]:
+    batched_conv = batch['messages']
+    batched_input_ids = []
+    batched_labels = []
+
+    for conv in batched_conv:
+        input_ids = [151331, 151333]
+        loss_masks = [False, False]
+        for message in conv:
+            message = process_message(message)
+            loss_mask_val = False if message['role'] in ('system', 'user', 'observation') else True
+            new_input_ids = tokenizer.apply_chat_template([message], tokenize=True, return_dict=False)[0][2:]
+            new_loss_masks = [loss_mask_val] * len(new_input_ids)
+            input_ids += new_input_ids
+            loss_masks += new_loss_masks
+        input_ids.append(tokenizer.eos_token_id)
+        loss_masks = [False, *loss_masks]
+        labels = []
+        for input_id, mask in zip(input_ids, loss_masks):
+            if mask:
+                labels.append(input_id)
+            else:
+                labels.append(-100)
+        max_length = max_input_length + max_output_length + 1
+        batched_input_ids.append(input_ids[:max_length])
+        batched_labels.append(labels[:max_length])
+    return {'input_ids': batched_input_ids, 'labels': batched_labels}
+
+
+def process_batch_eval(
+        batch: Mapping[str, Sequence],
+        tokenizer: PreTrainedTokenizer,
+        max_input_length: int,
+        max_output_length: int,
+) -> dict[str, list]:
+    batched_conv = batch['messages']
+    batched_input_ids = []
+    batched_output_ids = []
+
+    for conv in batched_conv:
+
+        input_ids = [151331, 151333]
+        for message in conv:
+            if len(input_ids) >= max_input_length:
+                break
+            else:
+                message = process_message(message)
+                new_input_ids = tokenizer.apply_chat_template([message], tokenize=True, return_dict=False)[0][2:]
+                if message['role'] == 'assistant':
+                    output_prompt, output_ids = (
+                        new_input_ids[:1],
+                        new_input_ids[1:],
+                    )
+                    output_ids.append(tokenizer.eos_token_id)
+                    batched_input_ids.append(
+                        input_ids[:max_input_length] + output_prompt[:1]
+                    )
+                    batched_output_ids.append(output_ids[:max_output_length])
+                input_ids += new_input_ids
+    return {'input_ids': batched_input_ids, 'output_ids': batched_output_ids}
+
+
+def load_tokenizer_and_model(
+        model_dir: str,
+        peft_config: Optional[PeftConfig] = None,
+):
+    tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
+    if peft_config is not None:
+        model = AutoModelForCausalLM.from_pretrained(
+            model_dir,
+            trust_remote_code=True,
+            empty_init=False,
+            use_cache=False,
+            torch_dtype=torch.bfloat16  # Must use BFloat 16
+        )
+        model = get_peft_model(model, peft_config)
+        model.print_trainable_parameters()
+    else:
+        model = AutoModelForCausalLM.from_pretrained(
+            model_dir,
+            trust_remote_code=True,
+            empty_init=False,
+            use_cache=False,
+            torch_dtype=torch.bfloat16
+        )
+    return tokenizer, model
+
+
+def compute_metrics(eval_preds: EvalPrediction, tokenizer):
+    batched_pred_ids, batched_label_ids = eval_preds
+    metrics_dct = {'rouge-1': [], 'rouge-2': [], 'rouge-l': [], 'bleu-4': []}
+    for pred_ids, label_ids in zip(batched_pred_ids, batched_label_ids):
+        pred_txt = tokenizer.decode(pred_ids).strip()
+        label_txt = tokenizer.decode(label_ids).strip()
+        pred_tokens = list(jieba.cut(pred_txt))
+        label_tokens = list(jieba.cut(label_txt))
+        rouge = Rouge()
+        scores = rouge.get_scores(' '.join(pred_tokens), ' '.join(label_tokens))
+        for k, v in scores[0].items():
+            metrics_dct[k].append(round(v['f'] * 100, 4))
+        metrics_dct['bleu-4'].append(
+            sentence_bleu([label_tokens], pred_tokens, smoothing_function=SmoothingFunction().method3))
+    return {k: np.mean(v) for k, v in metrics_dct.items()}
+
+
+@app.command()
+def main(
+        data_dir: Annotated[str, typer.Argument(help='')],
+        model_dir: Annotated[
+            str,
+            typer.Argument(
+                help='A string that specifies the model id of a pretrained model configuration hosted on huggingface.co, or a path to a directory containing a model configuration file.'
+            ),
+        ],
+        config_file: Annotated[str, typer.Argument(help='')],
+        auto_resume_from_checkpoint: str = typer.Argument(
+            default='',
+            help='If entered as yes, automatically use the latest save checkpoint. If it is a numerical example 12 15, use the corresponding save checkpoint. If the input is no, restart training'
+        ),
+
+):
+    ft_config = FinetuningConfig.from_file(config_file)
+    tokenizer, model = load_tokenizer_and_model(model_dir, peft_config=ft_config.peft_config)
+    data_manager = DataManager(data_dir, ft_config.data_config)
+
+    train_dataset = data_manager.get_dataset(
+        Split.TRAIN,
+        functools.partial(
+            process_batch,
+            tokenizer=tokenizer,
+            max_input_length=ft_config.max_input_length,
+            max_output_length=ft_config.max_output_length,
+        ),
+        batched=True,
+    )
+    print('train_dataset:', train_dataset)
+    val_dataset = data_manager.get_dataset(
+        Split.VALIDATION,
+        functools.partial(
+            process_batch_eval,
+            tokenizer=tokenizer,
+            max_input_length=ft_config.max_input_length,
+            max_output_length=ft_config.max_output_length,
+        ),
+        batched=True,
+    )
+    if val_dataset is not None:
+        print('val_dataset:', val_dataset)
+    test_dataset = data_manager.get_dataset(
+        Split.TEST,
+        functools.partial(
+            process_batch_eval,
+            tokenizer=tokenizer,
+            max_input_length=ft_config.max_input_length,
+            max_output_length=ft_config.max_output_length,
+        ),
+        batched=True,
+    )
+    if test_dataset is not None:
+        print('test_dataset:', test_dataset)
+
+    model.gradient_checkpointing_enable()
+    model.enable_input_require_grads()
+
+    trainer = Seq2SeqTrainer(
+        model=model,
+        args=ft_config.training_args,
+        data_collator=DataCollatorForSeq2Seq(
+            tokenizer=tokenizer,
+            padding='longest',
+            return_tensors='pt',
+        ),
+        train_dataset=train_dataset,
+        eval_dataset=val_dataset.select(list(range(50))),
+        compute_metrics=functools.partial(compute_metrics, tokenizer=tokenizer),
+    )
+
+    if auto_resume_from_checkpoint.upper() == "" or auto_resume_from_checkpoint is None:
+        trainer.train()
+    else:
+        output_dir = ft_config.training_args.output_dir
+        dirlist = os.listdir(output_dir)
+        checkpoint_sn = 0
+        for checkpoint_str in dirlist:
+            if checkpoint_str.find("eckpoint") > 0 and checkpoint_str.find("tmp") == -1:
+                checkpoint = int(checkpoint_str.replace("checkpoint-", ""))
+                if checkpoint > checkpoint_sn:
+                    checkpoint_sn = checkpoint
+        if auto_resume_from_checkpoint.upper() == "YES":
+            if checkpoint_sn > 0:
+                model.gradient_checkpointing_enable()
+                model.enable_input_require_grads()
+                checkpoint_directory = os.path.join(output_dir, "checkpoint-" + str(checkpoint_sn))
+                print("resume checkpoint from  checkpoint-" + str(checkpoint_sn))
+                trainer.train(resume_from_checkpoint=checkpoint_directory)
+            else:
+                trainer.train()
+        else:
+            if auto_resume_from_checkpoint.isdigit():
+                if int(auto_resume_from_checkpoint) > 0:
+                    checkpoint_sn = int(auto_resume_from_checkpoint)
+                    model.gradient_checkpointing_enable()
+                    model.enable_input_require_grads()
+                    checkpoint_directory = os.path.join(output_dir, "checkpoint-" + str(checkpoint_sn))
+                    print("resume checkpoint from  checkpoint-" + str(checkpoint_sn))
+                    trainer.train(resume_from_checkpoint=checkpoint_directory)
+            else:
+                print(auto_resume_from_checkpoint,
+                      "The specified checkpoint sn(" + auto_resume_from_checkpoint + ") has not been saved. Please search for the correct checkpoint in the model output directory")
+
+    if test_dataset is not None:
+        trainer.predict(test_dataset)
+
+
+if __name__ == '__main__':
+    app()

finetune_demo/inference.py

+from pathlib import Path
+from typing import Annotated, Union
+
+import typer
+from peft import AutoPeftModelForCausalLM, PeftModelForCausalLM
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    PreTrainedModel,
+    PreTrainedTokenizer,
+    PreTrainedTokenizerFast
+)
+
+ModelType = Union[PreTrainedModel, PeftModelForCausalLM]
+TokenizerType = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
+
+app = typer.Typer(pretty_exceptions_show_locals=False)
+
+
+def load_model_and_tokenizer(
+        model_dir: Union[str, Path], trust_remote_code: bool = True
+) -> tuple[ModelType, TokenizerType]:
+    model_dir = Path(model_dir).expanduser().resolve()
+    if (model_dir / 'adapter_config.json').exists():
+        model = AutoPeftModelForCausalLM.from_pretrained(
+            model_dir, trust_remote_code=trust_remote_code, device_map='auto'
+        )
+        tokenizer_dir = model.peft_config['default'].base_model_name_or_path
+    else:
+        model = AutoModelForCausalLM.from_pretrained(
+            model_dir, trust_remote_code=trust_remote_code, device_map='auto'
+        )
+        tokenizer_dir = model_dir
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_dir, trust_remote_code=trust_remote_code, encode_special_tokens=True, use_fast=False
+    )
+    return model, tokenizer
+
+
+@app.command()
+def main(
+        model_dir: Annotated[str, typer.Argument(help='')],
+):
+    messages = [
+        {
+            "role": "system", "content": "",
+            "tools":
+                [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "create_calendar_event",
+                            "description": "Create a new calendar event",
+                            "parameters": {
+                                "type": "object",
+                                "properties": {
+                                    "title": {
+                                        "type": "string",
+                                        "description": "The title of the event"
+                                    },
+                                    "start_time": {
+                                        "type": "string",
+                                        "description": "The start time of the event in the format YYYY-MM-DD HH:MM"
+                                    },
+                                    "end_time": {
+                                        "type": "string",
+                                        "description": "The end time of the event in the format YYYY-MM-DD HH:MM"
+                                    }
+                                },
+                                "required": [
+                                    "title",
+                                    "start_time",
+                                    "end_time"
+                                ]
+                            }
+                        }
+                    }
+                ]
+
+        },
+        {
+            "role": "user",
+            "content": "Can you help me create a calendar event for my meeting tomorrow? The title is \"Team Meeting\". It starts at 10:00 AM and ends at 11:00 AM."
+        },
+    ]
+    model, tokenizer = load_model_and_tokenizer(model_dir)
+    inputs = tokenizer.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        tokenize=True,
+        return_tensors="pt"
+    ).to(model.device)
+    generate_kwargs = {
+        "input_ids": inputs,
+        "max_new_tokens": 1024,
+        "do_sample": True,
+        "top_p": 0.8,
+        "temperature": 0.8,
+        "repetition_penalty": 1.2,
+        "eos_token_id": model.config.eos_token_id,
+    }
+    outputs = model.generate(**generate_kwargs)
+    response = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True).strip()
+    print("=========")
+    print(response)
+
+
+if __name__ == '__main__':
+    app()

finetune_demo/requirements.txt

+jieba>=0.42.1
+datasets>=2.19.1
+peft>=0.11.0
+deepspeed>=0.13.3
+nltk==3.8.1
\ No newline at end of file

glm-4v_inference.py

+import torch
+from PIL import Image
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import os
+import time
+
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "6"
+
+device = "cuda"
+
+tokenizer = AutoTokenizer.from_pretrained("/home/wanglch/projects/GLM-4V/glm-4v-b", trust_remote_code=True)
+
+query = 'OCR这张图片的文字内容'
+image = Image.open("/home/wanglch/projects/GLM-4V/images/pic3.jpg").convert('RGB')
+inputs = tokenizer.apply_chat_template([{"role": "user", "image": image, "content": query}],
+                                       add_generation_prompt=True, tokenize=True, return_tensors="pt",
+                                       return_dict=True)  # chat mode
+
+inputs = inputs.to(device)
+model = AutoModelForCausalLM.from_pretrained(
+    "/home/wanglch/projects/GLM-4V/glm-4v-b",
+    torch_dtype=torch.float16,
+    low_cpu_mem_usage=True,
+    trust_remote_code=True
+).to(device).eval()
+
+gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
+with torch.no_grad():
+    outputs = model.generate(**inputs, **gen_kwargs)
+    outputs = outputs[:, inputs['input_ids'].shape[1]:]
+    print(tokenizer.decode(outputs[0]))
\ No newline at end of file

lora_finetune_multi.sh

+NPROC_PER_NODE=2 \
+CUDA_VISIBLE_DEVICES=5,6 swift sft \
+    --model_id_or_path /home/wanglch/projects/GLM-4V/glm-4v-b \
+    --model_type glm4v-9b-chat \
+    --dataset coco-en-2-mini \
+    --output_dir /home/wanglch/projects/saves/glm-4v/train_dtk \
+    --num_train_epochs 1 \
+    --eval_steps 200 \
+    --sft_type lora \
+    
\ No newline at end of file

lora_finetune_single.sh

+CUDA_VISIBLE_DEVICES=0 swift sft \
+    --model_id_or_path /home/wanglch/projects/GLM-4V/glm-4v-b \
+    --model_type glm4v-9b-chat \
+    --dataset coco-mini-en-2 \
+    --output_dir /home/wanglch/projects/saves/glm-4v/train_cuda \
+    --num_train_epochs 1 \
+    --eval_steps 200 \
+    --sft_type lora \
\ No newline at end of file

model.properties

+# 模型唯一标识
+modelCode = 691
+# 模型名称
+modelName=minicpm-v_pytorch                    
+# 模型描述
+modelDescription=多模态OCR大模型，端侧可用
+# 应用场景
+appScenario=推理,OCR,金融,教育,政府,科研,交通,广媒
+# 框架类型
+frameType=pytorch

requirements.txt

+# use vllm
+# vllm>=0.4.3
+
+torch
+torchvision
+transformers==4.40.0
+huggingface-hub>=0.23.1
+sentencepiece>=0.2.0
+pydantic>=2.7.1
+timm>=0.9.16
+tiktoken>=0.7.0
+accelerate>=0.30.1
+sentence_transformers>=2.7.0
+
+# web demo
+gradio>=4.33.0
+
+# openai demo
+openai>=1.31.1
+einops>=0.7.0
+sse-starlette>=2.1.0
+
+# INT4
+bitsandbytes>=0.43.1
+
+# PEFT model, not need if you don't use PEFT finetune model.
+peft>=0.11.0
\ No newline at end of file

resources/WECHAT.md

+<div align="center">
+<img src=wechat.jpg width="60%"/>
+
+<p> 扫码关注公众号，加入「GLM-4交流群」 </p>
+<p> Scan the QR code to follow the official account and join the "ChatGLM Discussion Group" </p>
+</div>
+

swift-main/.dev_scripts/build_docs.sh

+pip install -r requirements/docs.txt
+cd docs
+rm -rf build
+
+# update api rst
+#rm -rf source/api/
+#sphinx-apidoc --module-first -o source/api/ ../modelscope/
+make html

swift-main/.dev_scripts/ci_container_test.sh

+if [ "$MODELSCOPE_SDK_DEBUG" == "True" ]; then
+    # pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
+    pip install -r requirements/tests.txt -i https://mirrors.aliyun.com/pypi/simple/
+    git config --global --add safe.directory /swift
+    git config --global user.email tmp
+    git config --global user.name tmp.com
+
+    # linter test
+    # use internal project for pre-commit due to the network problem
+    if [ `git remote -v | grep alibaba  | wc -l` -gt 1 ]; then
+        pre-commit run -c .pre-commit-config_local.yaml --all-files
+        if [ $? -ne 0 ]; then
+            echo "linter test failed, please run 'pre-commit run --all-files' to check"
+            echo "From the repository folder"
+            echo "Run 'pip install -r requirements/tests.txt' install test dependencies."
+            echo "Run 'pre-commit install' install pre-commit hooks."
+            echo "Finally run linter with command: 'pre-commit run --all-files' to check."
+            echo "Ensure there is no failure!!!!!!!!"
+            exit -1
+        fi
+    fi
+
+    pip install -r requirements/framework.txt -U -i https://mirrors.aliyun.com/pypi/simple/
+    pip install -r requirements/llm.txt -U -i https://mirrors.aliyun.com/pypi/simple/
+    pip install diffusers decord einops -U -i https://mirrors.aliyun.com/pypi/simple/
+
+    # test with install
+    pip install .
+    pip install auto_gptq bitsandbytes deepspeed -U -i https://mirrors.aliyun.com/pypi/simple/
+else
+    echo "Running case in release image, run case directly!"
+fi
+# remove torch_extensions folder to avoid ci hang.
+rm -rf ~/.cache/torch_extensions
+if [ $# -eq 0 ]; then
+    ci_command="python tests/run.py --subprocess"
+else
+    ci_command="$@"
+fi
+echo "Running case with command: $ci_command"
+$ci_command