|
快速上手 |
|
======== |
|
|
|
本节中,我们将演示如何使用 XTuner 微调模型,帮助您快速上手 XTuner。 |
|
|
|
在成功安装 XTuner |
|
后,便可以开始进行模型的微调。在本节中,我们将演示如何使用 XTuner,应用 |
|
QLoRA 算法在 Colorist 数据集上微调 InternLM2-Chat-7B。 |
|
|
|
Colorist 数据集(\ `HuggingFace |
|
链接 <https://huggingface.co/datasets/burkelibbey/colors>`__\ ;\ `ModelScope |
|
链接 <https://www.modelscope.cn/datasets/fanqiNO1/colors/summary>`__\ )是一个根据颜色描述提供颜色选择与建议的数据集,经过该数据集微调的模型可以做到根据用户对于颜色的描述,从而给出16进制下的颜色编码,如用户输入“宁静而又相当明亮的浅天蓝色,介于天蓝色和婴儿蓝之间,因其亮度而带有一丝轻微的荧光感。”,模型输出 |
|
|image1|\ ,该颜色很符合用户的描述。以下是该数据集的几条样例数据: |
|
|
|
+-----------------------+-----------------------+-------------------+ |
|
| 英文描述 | 中文描述 | 颜色 | |
|
+=======================+=======================+===================+ |
|
| Light Sky Blue: A | 浅天蓝色 | #66ccff: |image8| | |
|
| calming, fairly | :一种介于天蓝和婴儿 | | |
|
| bright color that | 蓝之间的平和、相当明 | | |
|
| falls between sky | 亮的颜色,由于明亮而 | | |
|
| blue and baby blue, | 带有一丝轻微的荧光。 | | |
|
| with a hint of slight | | | |
|
| fluorescence due to | | | |
|
| its brightness. | | | |
|
+-----------------------+-----------------------+-------------------+ |
|
| Bright red: This is a | 鲜红色: | #ee0000: |image9| | |
|
| very vibrant, | 这是一种非常鲜 | | |
|
| saturated and vivid | 艳、饱和、生动的红色 | | |
|
| shade of red, | ,类似成熟苹果或新鲜 | | |
|
| resembling the color | 血液的颜色。它是标准 | | |
|
| of ripe apples or | RGB | | |
|
| fresh blood. It is as | 调色板上的红色,不含 | | |
|
| red as you can get on | 任何蓝色或绿色元素。 | | |
|
| a standard RGB color | | | |
|
| palette, with no | | | |
|
| elements of either | | | |
|
| blue or green. | | | |
|
+-----------------------+-----------------------+-------------------+ |
|
| Bright Turquoise: | 明亮的绿松石 | #00ffcc: | |
|
| This color mixes the | 色:这种颜色融合了鲜 | |image10| | |
|
| freshness of bright | 绿色的清新和淡蓝色的 | | |
|
| green with the | 宁静,呈现出一种充满 | | |
|
| tranquility of light | 活力的绿松石色调。它 | | |
|
| blue, leading to a | 让人联想到热带水域。 | | |
|
| vibrant shade of | | | |
|
| turquoise. It is | | | |
|
| reminiscent of | | | |
|
| tropical waters. | | | |
|
+-----------------------+-----------------------+-------------------+ |
|
|
|
准备模型权重 |
|
------------ |
|
|
|
在微调模型前,首先要准备待微调模型的权重。 |
|
|
|
.. _从-huggingface-下载-1: |
|
|
|
从 HuggingFace 下载 |
|
~~~~~~~~~~~~~~~~~~~ |
|
|
|
.. code:: bash |
|
|
|
pip install -U huggingface_hub |
|
|
|
|
|
huggingface-cli download internlm/internlm2-chat-7b \ |
|
--local-dir Shanghai_AI_Laboratory/internlm2-chat-7b \ |
|
--local-dir-use-symlinks False \ |
|
--resume-download |
|
|
|
.. _从-modelscope-下载-1: |
|
|
|
从 ModelScope 下载 |
|
~~~~~~~~~~~~~~~~~~ |
|
|
|
由于从 HuggingFace |
|
拉取模型权重,可能存在下载过程不稳定、下载速度过慢等问题。因此在下载过程遇到网络问题时,我们则可以选择从 |
|
ModelScope 下载 InternLM2-Chat-7B 的权重。 |
|
|
|
.. code:: bash |
|
|
|
pip install -U modelscope |
|
|
|
|
|
python -c "from modelscope import snapshot_download; snapshot_download('Shanghai_AI_Laboratory/internlm2-chat-7b', cache_dir='.')" |
|
|
|
在完成下载后,便可以开始准备微调数据集了。 |
|
|
|
此处附上 HuggingFace 链接与 ModelScope 链接: |
|
|
|
- HuggingFace |
|
链接位于:\ https://huggingface.co/internlm/internlm2-chat-7b |
|
|
|
- ModelScope |
|
链接位于:\ https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-chat-7b/summary |
|
|
|
准备微调数据集 |
|
-------------- |
|
|
|
接下来,我们需要准备微调数据集。 |
|
|
|
.. _从-huggingface-下载-2: |
|
|
|
从 HuggingFace 下载 |
|
~~~~~~~~~~~~~~~~~~~ |
|
|
|
.. code:: bash |
|
|
|
git clone https://huggingface.co/datasets/burkelibbey/colors |
|
|
|
.. _从-modelscope-下载-2: |
|
|
|
从 ModelScope 下载 |
|
~~~~~~~~~~~~~~~~~~ |
|
|
|
由于相同的问题,因此我们可以选择从 ModelScope 下载所需要的微调数据集。 |
|
|
|
.. code:: bash |
|
|
|
git clone https://www.modelscope.cn/datasets/fanqiNO1/colors.git |
|
|
|
此处附上 HuggingFace 链接与 ModelScope 链接: |
|
|
|
- HuggingFace |
|
链接位于:\ https://huggingface.co/datasets/burkelibbey/colors |
|
|
|
- ModelScope 链接位于:\ https://modelscope.cn/datasets/fanqiNO1/colors |
|
|
|
准备配置文件 |
|
------------ |
|
|
|
XTuner 提供了多个开箱即用的配置文件,可以通过 ``xtuner list-cfg`` |
|
查看。我们执行如下指令,以复制一个配置文件到当前目录。 |
|
|
|
.. code:: bash |
|
|
|
xtuner copy-cfg internlm2_7b_qlora_colorist_e5 . |
|
|
|
配置文件名的解释: |
|
|
|
======== ============================== |
|
配置文件 internlm2_7b_qlora_colorist_e5 |
|
======== ============================== |
|
模型名 internlm2_7b |
|
使用算法 qlora |
|
数据集 colorist |
|
训练时长 5 epochs |
|
======== ============================== |
|
|
|
此时该目录文件结构应如下所示: |
|
|
|
.. code:: bash |
|
|
|
. |
|
├── colors |
|
│ ├── colors.json |
|
│ ├── dataset_infos.json |
|
│ ├── README.md |
|
│ └── train.jsonl |
|
├── internlm2_7b_qlora_colorist_e5_copy.py |
|
└── Shanghai_AI_Laboratory |
|
└── internlm2-chat-7b |
|
├── config.json |
|
├── configuration_internlm2.py |
|
├── configuration.json |
|
├── generation_config.json |
|
├── modeling_internlm2.py |
|
├── pytorch_model-00001-of-00008.bin |
|
├── pytorch_model-00002-of-00008.bin |
|
├── pytorch_model-00003-of-00008.bin |
|
├── pytorch_model-00004-of-00008.bin |
|
├── pytorch_model-00005-of-00008.bin |
|
├── pytorch_model-00006-of-00008.bin |
|
├── pytorch_model-00007-of-00008.bin |
|
├── pytorch_model-00008-of-00008.bin |
|
├── pytorch_model.bin.index.json |
|
├── README.md |
|
├── special_tokens_map.json |
|
├── tokenization_internlm2_fast.py |
|
├── tokenization_internlm2.py |
|
├── tokenizer_config.json |
|
└── tokenizer.model |
|
|
|
修改配置文件 |
|
------------ |
|
|
|
| 在这一步中,我们需要修改待微调模型路径和数据路径为本地路径,并且修改数据集加载方式。 |
|
| 此外,由于复制得到的配置文件是基于基座(Base)模型的,所以还需要修改 |
|
``prompt_template`` 以适配对话(Chat)模型。 |
|
|
|
.. code:: diff |
|
|
|
####################################################################### |
|
# PART 1 Settings # |
|
####################################################################### |
|
# Model |
|
- pretrained_model_name_or_path = 'internlm/internlm2-7b' |
|
+ pretrained_model_name_or_path = './Shanghai_AI_Laboratory/internlm2-chat-7b' |
|
|
|
# Data |
|
- data_path = 'burkelibbey/colors' |
|
+ data_path = './colors/train.jsonl' |
|
- prompt_template = PROMPT_TEMPLATE.default |
|
+ prompt_template = PROMPT_TEMPLATE.internlm2_chat |
|
|
|
... |
|
####################################################################### |
|
# PART 3 Dataset & Dataloader # |
|
####################################################################### |
|
train_dataset = dict( |
|
type=process_hf_dataset, |
|
- dataset=dict(type=load_dataset, path=data_path), |
|
+ dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_path)), |
|
tokenizer=tokenizer, |
|
max_length=max_length, |
|
dataset_map_fn=colors_map_fn, |
|
template_map_fn=dict( |
|
type=template_map_fn_factory, template=prompt_template), |
|
remove_unused_columns=True, |
|
shuffle_before_pack=True, |
|
pack_to_max_length=pack_to_max_length) |
|
|
|
因此在这一步中,修改了 |
|
``pretrained_model_name_or_path``\ 、\ ``data_path``\ 、\ ``prompt_template`` |
|
以及 ``train_dataset`` 中的 ``dataset`` 字段。 |
|
|
|
启动微调 |
|
-------- |
|
|
|
在完成上述操作后,便可以使用下面的指令启动微调任务了。 |
|
|
|
.. code:: bash |
|
|
|
# 单机单卡 |
|
xtuner train ./internlm2_7b_qlora_colorist_e5_copy.py |
|
# 单机多卡 |
|
NPROC_PER_NODE=${GPU_NUM} xtuner train ./internlm2_7b_qlora_colorist_e5_copy.py |
|
# slurm 情况 |
|
srun ${SRUN_ARGS} xtuner train ./internlm2_7b_qlora_colorist_e5_copy.py --launcher slurm |
|
|
|
正确输出的训练日志应类似如下所示: |
|
|
|
.. code:: text |
|
|
|
01/29 21:35:34 - mmengine - INFO - Iter(train) [ 10/720] lr: 9.0001e-05 eta: 0:31:46 time: 2.6851 data_time: 0.0077 memory: 12762 loss: 2.6900 |
|
01/29 21:36:02 - mmengine - INFO - Iter(train) [ 20/720] lr: 1.9000e-04 eta: 0:32:01 time: 2.8037 data_time: 0.0071 memory: 13969 loss: 2.6049 grad_norm: 0.9361 |
|
01/29 21:36:29 - mmengine - INFO - Iter(train) [ 30/720] lr: 1.9994e-04 eta: 0:31:24 time: 2.7031 data_time: 0.0070 memory: 13969 loss: 2.5795 grad_norm: 0.9361 |
|
01/29 21:36:57 - mmengine - INFO - Iter(train) [ 40/720] lr: 1.9969e-04 eta: 0:30:55 time: 2.7247 data_time: 0.0069 memory: 13969 loss: 2.3352 grad_norm: 0.8482 |
|
01/29 21:37:24 - mmengine - INFO - Iter(train) [ 50/720] lr: 1.9925e-04 eta: 0:30:28 time: 2.7286 data_time: 0.0068 memory: 13969 loss: 2.2816 grad_norm: 0.8184 |
|
01/29 21:37:51 - mmengine - INFO - Iter(train) [ 60/720] lr: 1.9863e-04 eta: 0:29:58 time: 2.7048 data_time: 0.0069 memory: 13969 loss: 2.2040 grad_norm: 0.8184 |
|
01/29 21:38:18 - mmengine - INFO - Iter(train) [ 70/720] lr: 1.9781e-04 eta: 0:29:31 time: 2.7302 data_time: 0.0068 memory: 13969 loss: 2.1912 grad_norm: 0.8460 |
|
01/29 21:38:46 - mmengine - INFO - Iter(train) [ 80/720] lr: 1.9681e-04 eta: 0:29:05 time: 2.7338 data_time: 0.0069 memory: 13969 loss: 2.1512 grad_norm: 0.8686 |
|
01/29 21:39:13 - mmengine - INFO - Iter(train) [ 90/720] lr: 1.9563e-04 eta: 0:28:36 time: 2.7047 data_time: 0.0068 memory: 13969 loss: 2.0653 grad_norm: 0.8686 |
|
01/29 21:39:40 - mmengine - INFO - Iter(train) [100/720] lr: 1.9426e-04 eta: 0:28:09 time: 2.7383 data_time: 0.0070 memory: 13969 loss: 1.9819 grad_norm: 0.9127 |
|
|
|
在训练开始前,可以看到模型的输出如下所示: |
|
|
|
.. code:: text |
|
|
|
2024/01/29 21:34:58 - mmengine - INFO - before_train in EvaluateChatHook. |
|
2024/01/29 21:35:03 - mmengine - INFO - Sample output: |
|
<s><|im_start|>system |
|
You are a professional color designer. Please provide the corresponding colors based on the description of Human. |
|
<|im_end|> |
|
<|im_start|>user |
|
请给我一个像天空一样清澈透明的蓝色。<|im_end|> |
|
<|im_start|>assistant |
|
为了匹配您所描述的“像天空一样清澈透明的蓝色”,我建议您选择一种名为“天蓝”(Cerulean)的颜色。这种颜色通常被用来代表天空、海洋和清澈的水域,它具有明亮 |
|
|
|
2024/01/29 21:35:07 - mmengine - INFO - Sample output: |
|
<s><|im_start|>system |
|
You are a professional color designer. Please provide the corresponding colors based on the description of Human. |
|
<|im_end|> |
|
<|im_start|>user |
|
Please give me a clear blue like the sky.<|im_end|> |
|
<|im_start|>assistant |
|
The color you're describing is a clear, bright blue that resembles the color of the sky on a sunny day. This color is often referred to as "sky blue" or "cerulean blue." Here are some specific color codes and names that |
|
|
|
而在训练完全完成后,可以观察到模型的输出如下所示: |
|
|
|
.. code:: text |
|
|
|
2024/01/29 22:08:07 - mmengine - INFO - Sample output: |
|
<s><|im_start|>system |
|
You are a professional color designer. Please provide the corresponding colors based on the description of Human. |
|
<|im_end|> |
|
<|im_start|>user |
|
请给我一个像天空一样清澈透明的蓝色。<|im_end|> |
|
<|im_start|>assistant |
|
#1099ee<|im_end|> |
|
|
|
2024/01/29 22:08:08 - mmengine - INFO - Sample output: |
|
<s><|im_start|>system |
|
You are a professional color designer. Please provide the corresponding colors based on the description of Human. |
|
<|im_end|> |
|
<|im_start|>user |
|
Please give me a clear blue like the sky.<|im_end|> |
|
<|im_start|>assistant |
|
#0066dd<|im_end|> |
|
|
|
模型输出的颜色如下所示: |
|
|
|
- 天空一样清澈透明的蓝色:\ |image11| |
|
|
|
- A clear blue like the sky: |image12| |
|
|
|
不难发现,模型在经过训练后,其输出已经完全与数据集内容所对齐了。 |
|
|
|
.. _模型转换--lora-合并: |
|
|
|
模型转换 + LoRA 合并 |
|
-------------------- |
|
|
|
在训练完成后,我们会得到几个 ``.pth`` 文件,这些文件存储了 QLoRA |
|
算法训练过程所更新的参数,而\ **不是**\ 模型的全部参数。因此我们需要将这些 |
|
``.pth`` 文件转换为 HuggingFace 格式,并合并入原始的语言模型权重中。 |
|
|
|
模型转换 |
|
~~~~~~~~ |
|
|
|
XTuner 已经集成好了将模型转换为 HuggingFace 格式的工具,我们只需要执行 |
|
|
|
.. code:: bash |
|
|
|
|
|
mkdir work_dirs/internlm2_7b_qlora_colorist_e5_copy/iter_720_hf |
|
|
|
|
|
xtuner convert pth_to_hf internlm2_7b_qlora_colorist_e5_copy.py \ |
|
work_dirs/internlm2_7b_qlora_colorist_e5_copy/iter_720.pth \ |
|
work_dirs/internlm2_7b_qlora_colorist_e5_copy/iter_720_hf |
|
|
|
该条转换命令将会根据配置文件 ``internlm2_7b_qlora_colorist_e5_copy.py`` |
|
的内容,将 |
|
``work_dirs/internlm2_7b_qlora_colorist_e5_copy/iter_720.pth`` 转换为 hf |
|
格式,并保存在 |
|
``work_dirs/internlm2_7b_qlora_colorist_e5_copy/iter_720_hf`` 位置。 |
|
|
|
LoRA 合并 |
|
~~~~~~~~~ |
|
|
|
XTuner 也已经集成好了合并 LoRA 权重的工具,我们只需执行如下指令: |
|
|
|
.. code:: bash |
|
|
|
|
|
mkdir work_dirs/internlm2_7b_qlora_colorist_e5_copy/merged |
|
|
|
|
|
xtuner convert merge Shanghai_AI_Laboratory/internlm2-chat-7b \ |
|
work_dirs/internlm2_7b_qlora_colorist_e5_copy/iter_720_hf \ |
|
work_dirs/internlm2_7b_qlora_colorist_e5_copy/merged \ |
|
--max-shard-size 2GB |
|
|
|
与转换命令类似,该条合并参数命令会读取原始参数路径 |
|
``Shanghai_AI_Laboratory/internlm2-chat-7b`` 以及转换为 hf |
|
格式的部分参数路径 |
|
``work_dirs/internlm2_7b_qlora_colorist_e5_copy/iter_720_hf``\ ,将两部分参数合并后保存于 |
|
``work_dirs/internlm2_7b_qlora_colorist_e5_copy/merged``\ ,其中每个参数切片的最大文件大小为 |
|
2GB。 |
|
|
|
与模型对话 |
|
---------- |
|
|
|
在合并权重后,为了更好地体会到模型的能力,XTuner |
|
也集成了与模型对话的工具。通过如下命令,便可以启动一个与模型对话的简易 |
|
Demo。 |
|
|
|
.. code:: bash |
|
|
|
xtuner chat work_dirs/internlm2_7b_qlora_colorist_e5_copy/merged \ |
|
--prompt-template internlm2_chat \ |
|
--system-template colorist |
|
|
|
当然,我们也可以选择不合并权重,而是直接与 LLM + LoRA Adapter |
|
进行对话,我们只需要执行如下指令: |
|
|
|
.. code:: bash |
|
|
|
xtuner chat Shanghai_AI_Laboratory/internlm2-chat-7b |
|
--adapter work_dirs/internlm2_7b_qlora_colorist_e5_copy/iter_720_hf \ |
|
--prompt-template internlm2_chat \ |
|
--system-template colorist |
|
|
|
其中 ``work_dirs/internlm2_7b_qlora_colorist_e5_copy/merged`` |
|
是合并后的权重路径,\ ``--prompt-template internlm2_chat`` |
|
指定了对话模板为 InternLM2-Chat,\ ``--system-template colorist`` |
|
则是指定了与模型对话时的 System Prompt 为 Colorist 数据集所要求的模板。 |
|
|
|
以下是一个例子: |
|
|
|
.. code:: text |
|
|
|
double enter to end input (EXIT: exit chat, RESET: reset history) >>> 宁静而又相当明亮的浅天蓝色,介于天蓝色和婴儿蓝之间,因其亮度而带有一丝轻微的荧光感。 |
|
|
|
|
|
|
|
其颜色如下所示: |
|
|
|
宁静而又相当明亮的浅天蓝色,介于天蓝色和婴儿蓝之间,因其亮度而带有一丝轻微的荧光感。:\ |image13| |
|
|
|
.. |image1| image:: https://img.shields.io/badge/%2366ccff-66CCFF |
|
.. |image2| image:: https://img.shields.io/badge/%2366ccff-66CCFF |
|
.. |image3| image:: https://img.shields.io/badge/%23ee0000-EE0000 |
|
.. |image4| image:: https://img.shields.io/badge/%2300ffcc-00FFCC |
|
.. |image5| image:: https://img.shields.io/badge/%2366ccff-66CCFF |
|
.. |image6| image:: https://img.shields.io/badge/%23ee0000-EE0000 |
|
.. |image7| image:: https://img.shields.io/badge/%2300ffcc-00FFCC |
|
.. |image8| image:: https://img.shields.io/badge/%2366ccff-66CCFF |
|
.. |image9| image:: https://img.shields.io/badge/%23ee0000-EE0000 |
|
.. |image10| image:: https://img.shields.io/badge/%2300ffcc-00FFCC |
|
.. |image11| image:: https://img.shields.io/badge/天空一样清澈透明的蓝色-1099EE |
|
.. |image12| image:: https://img.shields.io/badge/A_clear_blue_like_the_sky-0066DD |
|
.. |image13| image:: https://img.shields.io/badge/宁静而又相当明亮的浅天蓝色,介于天蓝色和婴儿蓝之间,因其亮度而带有一丝轻微的荧光感。-66CCFF |
|
|