add Incremental Pre-training Guide

README.md

@@ -98,6 +98,7 @@
 </table>
 
 ### 🎇最近更新
+- 【2024.5.7】[增量预训练指南](xtuner_config/pt/README.md)
 - 【2024.4.20】[LLAMA3微调指南](xtuner_config/README_llama3_8b_instruct_qlora_alpaca_e3_M.md)及基于[LLaMA3_8b_instruct的艾薇](https://openxlab.org.cn/models/detail/ajupyter/EmoLLM-LLaMA3_8b_instruct_aiwei)开源
 - 【2023.4.14】新增[快速开始](docs/quick_start.md)和保姆级教程[BabyEmoLLM](Baby_EmoLLM.ipynb)
 - 【2024.4.2】在 Huggingface 上传[老母亲心理咨询师](https://huggingface.co/brycewang2018/EmoLLM-mother/tree/main)

README_EN.md

@@ -101,6 +101,7 @@ The Model aims to fully understand and promote the mental health of individuals,
 </table>
 
 ### Recent Updates
+- [2024.5.7][Incremental Pre-training Guide](xtuner_config/pt/README.md)
  - [2024.4.20] [LLAMA3 fine-tuning guide](xtuner_config/README_llama3_8b_instruct_qlora_alpaca_e3_M.md) and based on [LLaMA3_8b_instruct's aiwei](https://openxlab.org.cn/models/detail/ajupyter/EmoLLM-LLaMA3_8b_instruct_aiwei) open source
 - [2023.4.14] Added [Quick Start](docs/quick_start_EN.md) and Nanny level tutorial [BabyEmoLLM](Baby_EmoLLM.ipynb)
 - [2024.4.2] Uploaded at Huggingface [Old Mother Counsellor](https://huggingface.co/brycewang2018/EmoLLM-mother/tree/main)

xtuner_config/pt/README.md

@@ -0,0 +1,36 @@
+# 增量预训练教程
+
+# 增量预训练简介
+增量预训练旨在提升模型在特定领域或任务的能力。
+
+
+# 预训练流程
+- Step1 处理数据
+- Step2 配置config（全量、Lora、Qlora）
+- Step3 启动训练（单卡、多卡、是否使用deepspeed）
+- Step4 模型合成
+- Step5 模型测试
+- Step6 模型上传
+
+# EmoLLM增量预训练教程
+基于微调中的数据集[datasets](../../datasets)修改而来
+
+- Step1 修改`ft2pt.py`中的文件路径
+这里以[output2.json](../../datasets/processed/output2.json)为例，运行脚本生成[pt.json](../../datasets/pt/pt.json)
+
+- Step2 [config](./internlm2_chat_1_8b_qlora_e3_pt.py)
+注意：本config采用了**变长注意力 (Variable Length Attention)**
+需要安装flash_attn
+`MAX_JOBS=4 pip install flash-attn --no-build-isolation`
+
+
+- Step3 训练：
+```
+# On a single GPU
+xtuner train internlm2_chat_1_8b_qlora_e3_pt.py --deepspeed deepspeed_zero2
+# On multiple GPUs
+(DIST) NPROC_PER_NODE=${GPU_NUM} xtuner train internlm2_chat_1_8b_qlora_e3_pt.py --deepspeed deepspeed_zero2
+(SLURM) srun ${SRUN_ARGS} xtuner train internlm2_chat_1_8b_qlora_e3_pt.py --launcher slurm --deepspeed deepspeed_zero2
+```
+
+- 其余流程请参考[微调教程](../../xtuner_config/README.md)

xtuner_config/pt/ft2pt.py

@@ -0,0 +1,48 @@
+# 将微调的数据格式转为预训练的格式
+import json
+
+
+def convert(data_path:str, target_path:str):
+    # 假设原始JSON数据存储在名为'data.json'的文件中
+    filename = data_path
+
+    # 读取文件内容
+    with open(filename, 'rt', encoding='utf-8') as file:
+        original_json = file.read()
+
+    # 将原始JSON字符串解析为Python对象
+    data = json.loads(original_json)
+
+    # 遍历每个对话
+    converted_data = []
+
+    # 遍历原始数据中的每个对话对象
+    for conversation_group in data:
+        # 遍历每个对话
+        for dialog in conversation_group["conversation"]:
+            # 创建一个新的对话对象，用于存储转换后的对话
+            new_conversation_group = {
+                "conversation": []
+            }
+            # 创建一个新的对话，其中输出被替换为"xxx"
+            new_dialog = {
+                "input": '',
+                "output": f'问题:{dialog["input"]}\n答案:{dialog["output"]}',
+            }
+            # 将新的对话添加到新对话对象的列表中
+            new_conversation_group["conversation"].append(new_dialog)
+            
+            # 将新对话对象添加到转换后的数据列表中
+            converted_data.append(new_conversation_group)
+
+
+    # 将更新后的数据转换回JSON字符串，并格式化输出
+    updated_json = json.dumps(converted_data, indent=4, ensure_ascii=False)
+
+
+    # 将更新后的JSON数据写入到新的文件中
+    with open(f'{target_path}', 'wt', encoding='utf-8') as file:
+        file.write(updated_json)
+
+if __name__ == '__main__':
+    convert(data_path='./output2.json', target_path='pt.json')

xtuner_config/pt/internlm2_chat_1_8b_qlora_e3_pt.py

@@ -0,0 +1,202 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = '/root/share/new_models/Shanghai_AI_Laboratory/internlm2-chat-1_8b'
+use_varlen_attn = True # True
+
+# Data
+data_path = '/root/wxz/work/pt/pt.json'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 1
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '我们经常因为一些小事争吵，他总是忽略我的感受。我感到很孤独，'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_en = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_path)),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=None,
+    template_map_fn=None,
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_en,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)