duzx16
commited on
Commit
•
a5172e3
1
Parent(s):
554c591
Init commit
Browse files- MODEL_LICENSE +33 -0
- README.md +88 -0
- config.json +40 -0
- configuration_chatglm.py +57 -0
- modeling_chatglm.py +1131 -0
- quantization.py +188 -0
- tokenization_chatglm.py +235 -0
- tokenizer_config.json +12 -0
MODEL_LICENSE
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
The ChatGLM2-6B License
|
2 |
+
|
3 |
+
1. Definitions
|
4 |
+
|
5 |
+
“Licensor” means the ChatGLM2-6B Model Team that distributes its Software.
|
6 |
+
|
7 |
+
“Software” means the ChatGLM2-6B model parameters made available under this license.
|
8 |
+
|
9 |
+
2. License Grant
|
10 |
+
|
11 |
+
Subject to the terms and conditions of this License, the Licensor hereby grants to you a non-exclusive, worldwide, non-transferable, non-sublicensable, revocable, royalty-free copyright license to use the Software solely for your non-commercial research purposes.
|
12 |
+
|
13 |
+
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
|
14 |
+
|
15 |
+
3. Restriction
|
16 |
+
|
17 |
+
You will not use, copy, modify, merge, publish, distribute, reproduce, or create derivative works of the Software, in whole or in part, for any commercial, military, or illegal purposes.
|
18 |
+
|
19 |
+
You will not use the Software for any act that may undermine China's national security and national unity, harm the public interest of society, or infringe upon the rights and interests of human beings.
|
20 |
+
|
21 |
+
4. Disclaimer
|
22 |
+
|
23 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
|
24 |
+
|
25 |
+
5. Limitation of Liability
|
26 |
+
|
27 |
+
EXCEPT TO THE EXTENT PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL THEORY, WHETHER BASED IN TORT, NEGLIGENCE, CONTRACT, LIABILITY, OR OTHERWISE WILL ANY LICENSOR BE LIABLE TO YOU FOR ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES, OR ANY OTHER COMMERCIAL LOSSES, EVEN IF THE LICENSOR HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
|
28 |
+
|
29 |
+
6. Dispute Resolution
|
30 |
+
|
31 |
+
This license shall be governed and construed in accordance with the laws of People’s Republic of China. Any dispute arising from or in connection with this License shall be submitted to Haidian District People's Court in Beijing.
|
32 |
+
|
33 |
+
Note that the license is subject to update to a more comprehensive version. For any questions related to the license and copyright, please contact us at glm-130b@googlegroups.com.
|
README.md
ADDED
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- zh
|
4 |
+
- en
|
5 |
+
tags:
|
6 |
+
- glm
|
7 |
+
- chatglm
|
8 |
+
- thudm
|
9 |
+
---
|
10 |
+
# ChatGLM-6B
|
11 |
+
<p align="center">
|
12 |
+
🌐 <a href="https://chatglm.cn/blog" target="_blank">Blog</a> • 💻 <a href="https://github.com/THUDM/ChatGLM-6B" target="_blank">Github Repo</a> • 🐦 <a href="https://twitter.com/thukeg" target="_blank">Twitter</a> • 📃 <a href="https://arxiv.org/abs/2103.10360" target="_blank">[GLM@ACL 22]</a> <a href="https://github.com/THUDM/GLM" target="_blank">[GitHub]</a> • 📃 <a href="https://arxiv.org/abs/2210.02414" target="_blank">[GLM-130B@ICLR 23]</a> <a href="https://github.com/THUDM/GLM-130B" target="_blank">[GitHub]</a> <br>
|
13 |
+
</p>
|
14 |
+
|
15 |
+
<p align="center">
|
16 |
+
👋 Join our <a href="https://join.slack.com/t/chatglm/shared_invite/zt-1th2q5u69-7tURzFuOPanmuHy9hsZnKA" target="_blank">Slack</a> and <a href="https://github.com/THUDM/ChatGLM-6B/blob/main/resources/WECHAT.md" target="_blank">WeChat</a>
|
17 |
+
</p>
|
18 |
+
|
19 |
+
## 介绍
|
20 |
+
ChatGLM-6B 是一个开源的、支持中英双语问答的对话语言模型,基于 [General Language Model (GLM)](https://github.com/THUDM/GLM) 架构,具有 62 亿参数。结合模型量化技术,用户可以在消费级的显卡上进行本地部署(INT4 量化级别下最低只需 6GB 显存)。ChatGLM-6B 使用了和 [ChatGLM](https://chatglm.cn) 相同的技术,针对中文问答和对话进行了优化。经过约 1T 标识符的中英双语训练,辅以监督微调、反馈自助、人类反馈强化学习等技术的加持,62 亿参数的 ChatGLM-6B 已经能生成相当符合人类偏好的回答。
|
21 |
+
|
22 |
+
ChatGLM-6B is an open bilingual language model based on [General Language Model (GLM)](https://github.com/THUDM/GLM) framework, with 6.2 billion parameters. With the quantization technique, users can deploy locally on consumer-grade graphics cards (only 6GB of GPU memory is required at the INT4 quantization level). ChatGLM-6B uses technology similar to ChatGPT, optimized for Chinese QA and dialogue. The model is trained for about 1T tokens of Chinese and English corpus, supplemented by supervised fine-tuning, feedback bootstrap, and reinforcement learning wit human feedback. With only about 6.2 billion parameters, the model is able to generate answers that are in line with human preference.
|
23 |
+
|
24 |
+
## 软件依赖
|
25 |
+
|
26 |
+
```shell
|
27 |
+
pip install protobuf==3.20.0 transformers==4.27.1 icetk cpm_kernels
|
28 |
+
```
|
29 |
+
|
30 |
+
## 代码调用
|
31 |
+
|
32 |
+
可以通过如下代码调用 ChatGLM-6B 模型来生成对话:
|
33 |
+
|
34 |
+
```ipython
|
35 |
+
>>> from transformers import AutoTokenizer, AutoModel
|
36 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
|
37 |
+
>>> model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
|
38 |
+
>>> response, history = model.chat(tokenizer, "你好", history=[])
|
39 |
+
>>> print(response)
|
40 |
+
你好👋!我是人工智能助手 ChatGLM-6B,很高兴见到你,欢迎问我任何问题。
|
41 |
+
>>> response, history = model.chat(tokenizer, "晚上睡不着应该怎么办", history=history)
|
42 |
+
>>> print(response)
|
43 |
+
晚上睡不着可能会让你感到焦虑或不舒服,但以下是一些可以帮助你入睡的方法:
|
44 |
+
|
45 |
+
1. 制定规律的睡眠时间表:保持规律的睡眠时间表可以帮助你建立健康的睡眠习惯,使你更容易入睡。尽量在每天的相同时间上床,并在同一时间起床。
|
46 |
+
2. 创造一个舒适的睡眠环境:确保睡眠环境舒适,安静,黑暗且温度适宜。可以使用舒适的床上用品,并保持房间通风。
|
47 |
+
3. 放松身心:在睡前做些放松的活动,例如泡个热水澡,听些轻柔的音乐,阅读一些有趣的书籍等,有助于缓解紧张和焦虑,使你更容易入睡。
|
48 |
+
4. 避免饮用含有咖啡因的饮料:咖啡因是一种刺激性物质,会影响你的睡眠质量。尽量避免在睡前饮用含有咖啡因的饮料,例如咖啡,茶和可乐。
|
49 |
+
5. 避免在床上做与睡眠无关的事情:在床上做些与睡眠无关的事情,例如看电影,玩游戏或工作等,可能会干扰你的睡眠。
|
50 |
+
6. 尝试呼吸技巧:深呼吸是一种放松技巧,可以帮助你缓解紧张和焦虑,使你更容易入睡。试着慢慢吸气,保持几秒钟,然后缓慢呼气。
|
51 |
+
|
52 |
+
如果这些方法无法帮助你入睡,你可以考虑咨询医生或睡眠专家,寻求进一步的建议。
|
53 |
+
```
|
54 |
+
|
55 |
+
关于更多的使用说明,包括如何运行命令行和网页版本的 DEMO,以及使用模型量化以节省显存,请参考我们的 [Github Repo](https://github.com/THUDM/ChatGLM-6B)。
|
56 |
+
|
57 |
+
For more instructions, including how to run CLI and web demos, and model quantization, please refer to our [Github Repo](https://github.com/THUDM/ChatGLM-6B).
|
58 |
+
|
59 |
+
## Change Log
|
60 |
+
* v0.1.0 ([f83182](https://huggingface.co/THUDM/chatglm-6b/commit/f83182484538e663a03d3f73647f10f89878f438))
|
61 |
+
|
62 |
+
## 协议
|
63 |
+
|
64 |
+
本仓库的代码依照 [Apache-2.0](LICENSE) 协议开源,ChatGLM-6B 模型的权重的使用则需要遵循 [Model License](MODEL_LICENSE)。
|
65 |
+
|
66 |
+
## 引用
|
67 |
+
|
68 |
+
如果你觉得我们的工作有帮助的话,请考虑引用下列论文:
|
69 |
+
|
70 |
+
```
|
71 |
+
@inproceedings{
|
72 |
+
zeng2023glm-130b,
|
73 |
+
title={{GLM}-130B: An Open Bilingual Pre-trained Model},
|
74 |
+
author={Aohan Zeng and Xiao Liu and Zhengxiao Du and Zihan Wang and Hanyu Lai and Ming Ding and Zhuoyi Yang and Yifan Xu and Wendi Zheng and Xiao Xia and Weng Lam Tam and Zixuan Ma and Yufei Xue and Jidong Zhai and Wenguang Chen and Zhiyuan Liu and Peng Zhang and Yuxiao Dong and Jie Tang},
|
75 |
+
booktitle={The Eleventh International Conference on Learning Representations (ICLR)},
|
76 |
+
year={2023},
|
77 |
+
url={https://openreview.net/forum?id=-Aw0rrrPUF}
|
78 |
+
}
|
79 |
+
```
|
80 |
+
```
|
81 |
+
@inproceedings{du2022glm,
|
82 |
+
title={GLM: General Language Model Pretraining with Autoregressive Blank Infilling},
|
83 |
+
author={Du, Zhengxiao and Qian, Yujie and Liu, Xiao and Ding, Ming and Qiu, Jiezhong and Yang, Zhilin and Tang, Jie},
|
84 |
+
booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
|
85 |
+
pages={320--335},
|
86 |
+
year={2022}
|
87 |
+
}
|
88 |
+
```
|
config.json
ADDED
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "THUDM/chatglm2-6b",
|
3 |
+
"architectures": [
|
4 |
+
"ChatGLMModel"
|
5 |
+
],
|
6 |
+
"auto_map": {
|
7 |
+
"AutoConfig": "configuration_chatglm.ChatGLMConfig",
|
8 |
+
"AutoModel": "modeling_chatglm.ChatGLMForConditionalGeneration",
|
9 |
+
"AutoModelForSeq2SeqLM": "modeling_chatglm.ChatGLMForConditionalGeneration"
|
10 |
+
},
|
11 |
+
"add_bias_linear": false,
|
12 |
+
"add_qkv_bias": true,
|
13 |
+
"apply_query_key_layer_scaling": true,
|
14 |
+
"apply_residual_connection_post_layernorm": false,
|
15 |
+
"attention_dropout": 0.0,
|
16 |
+
"attention_softmax_in_fp32": true,
|
17 |
+
"bias_dropout_fusion": true,
|
18 |
+
"ffn_hidden_size": 13696,
|
19 |
+
"fp32_residual_connection": false,
|
20 |
+
"hidden_dropout": 0.0,
|
21 |
+
"hidden_size": 4096,
|
22 |
+
"interleaved_qkv": false,
|
23 |
+
"kv_channels": 128,
|
24 |
+
"layernorm_epsilon": 1e-05,
|
25 |
+
"multi_query_attention": true,
|
26 |
+
"multi_query_group_num": 2,
|
27 |
+
"num_attention_heads": 32,
|
28 |
+
"num_layers": 28,
|
29 |
+
"original_rope": true,
|
30 |
+
"padded_vocab_size": 65024,
|
31 |
+
"post_layer_norm": true,
|
32 |
+
"rmsnorm": true,
|
33 |
+
"rotary_percent": 0.5,
|
34 |
+
"seq_length": 32768,
|
35 |
+
"use_cache": true,
|
36 |
+
"torch_dtype": "float16",
|
37 |
+
"transformers_version": "4.27.1",
|
38 |
+
"tie_word_embeddings": false,
|
39 |
+
"eos_token_id": 2
|
40 |
+
}
|
configuration_chatglm.py
ADDED
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import PretrainedConfig
|
2 |
+
|
3 |
+
|
4 |
+
class ChatGLMConfig(PretrainedConfig):
|
5 |
+
def __init__(
|
6 |
+
self,
|
7 |
+
num_layers=28,
|
8 |
+
padded_vocab_size=65024,
|
9 |
+
hidden_size=4096,
|
10 |
+
ffn_hidden_size=13696,
|
11 |
+
kv_channels=128,
|
12 |
+
num_attention_heads=32,
|
13 |
+
seq_length=2048,
|
14 |
+
hidden_dropout=0.0,
|
15 |
+
attention_dropout=0.0,
|
16 |
+
layernorm_epsilon=1e-5,
|
17 |
+
rmsnorm=True,
|
18 |
+
apply_residual_connection_post_layernorm=False,
|
19 |
+
post_layer_norm=True,
|
20 |
+
add_bias_linear=False,
|
21 |
+
add_qkv_bias=False,
|
22 |
+
interleaved_qkv=False,
|
23 |
+
bias_dropout_fusion=True,
|
24 |
+
rotary_percent=1.0,
|
25 |
+
multi_query_attention=False,
|
26 |
+
multi_query_group_num=1,
|
27 |
+
apply_query_key_layer_scaling=True,
|
28 |
+
attention_softmax_in_fp32=True,
|
29 |
+
fp32_residual_connection=False,
|
30 |
+
quantization_bit=0,
|
31 |
+
**kwargs
|
32 |
+
):
|
33 |
+
self.num_layers = num_layers
|
34 |
+
self.padded_vocab_size = padded_vocab_size
|
35 |
+
self.hidden_size = hidden_size
|
36 |
+
self.ffn_hidden_size = ffn_hidden_size
|
37 |
+
self.kv_channels = kv_channels
|
38 |
+
self.num_attention_heads = num_attention_heads
|
39 |
+
self.seq_length = seq_length
|
40 |
+
self.hidden_dropout = hidden_dropout
|
41 |
+
self.attention_dropout = attention_dropout
|
42 |
+
self.layernorm_epsilon = layernorm_epsilon
|
43 |
+
self.rmsnorm = rmsnorm
|
44 |
+
self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
|
45 |
+
self.post_layer_norm = post_layer_norm
|
46 |
+
self.add_bias_linear = add_bias_linear
|
47 |
+
self.add_qkv_bias = add_qkv_bias
|
48 |
+
self.interleaved_qkv = interleaved_qkv
|
49 |
+
self.bias_dropout_fusion = bias_dropout_fusion
|
50 |
+
self.rotary_percent = rotary_percent
|
51 |
+
self.multi_query_attention = multi_query_attention
|
52 |
+
self.multi_query_group_num = multi_query_group_num
|
53 |
+
self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
|
54 |
+
self.attention_softmax_in_fp32 = attention_softmax_in_fp32
|
55 |
+
self.fp32_residual_connection = fp32_residual_connection
|
56 |
+
self.quantization_bit = quantization_bit
|
57 |
+
super().__init__(**kwargs)
|
modeling_chatglm.py
ADDED
@@ -0,0 +1,1131 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
""" PyTorch ChatGLM model. """
|
2 |
+
|
3 |
+
import math
|
4 |
+
import copy
|
5 |
+
import warnings
|
6 |
+
import re
|
7 |
+
import sys
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.utils.checkpoint
|
11 |
+
import torch.nn.functional as F
|
12 |
+
from torch import nn
|
13 |
+
from torch.nn import CrossEntropyLoss, LayerNorm
|
14 |
+
from torch.nn.utils import skip_init
|
15 |
+
from typing import Optional, Tuple, Union, List, Callable, Dict, Any
|
16 |
+
|
17 |
+
from transformers.modeling_outputs import (
|
18 |
+
BaseModelOutputWithPast,
|
19 |
+
CausalLMOutputWithPast,
|
20 |
+
)
|
21 |
+
from transformers.modeling_utils import PreTrainedModel
|
22 |
+
from transformers.utils import logging
|
23 |
+
from transformers.generation.logits_process import LogitsProcessor
|
24 |
+
from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList, GenerationConfig, ModelOutput
|
25 |
+
|
26 |
+
from .configuration_chatglm import ChatGLMConfig
|
27 |
+
|
28 |
+
# flags required to enable jit fusion kernels
|
29 |
+
|
30 |
+
if sys.platform != 'darwin':
|
31 |
+
torch._C._jit_set_profiling_mode(False)
|
32 |
+
torch._C._jit_set_profiling_executor(False)
|
33 |
+
torch._C._jit_override_can_fuse_on_cpu(True)
|
34 |
+
torch._C._jit_override_can_fuse_on_gpu(True)
|
35 |
+
|
36 |
+
logger = logging.get_logger(__name__)
|
37 |
+
|
38 |
+
_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM-6B"
|
39 |
+
_CONFIG_FOR_DOC = "ChatGLM6BConfig"
|
40 |
+
|
41 |
+
CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
42 |
+
"THUDM/chatglm-6b",
|
43 |
+
# See all ChatGLM-6B models at https://huggingface.co/models?filter=chatglm
|
44 |
+
]
|
45 |
+
|
46 |
+
|
47 |
+
def default_init(cls, *args, **kwargs):
|
48 |
+
return cls(*args, **kwargs)
|
49 |
+
|
50 |
+
|
51 |
+
class InvalidScoreLogitsProcessor(LogitsProcessor):
|
52 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
|
53 |
+
if torch.isnan(scores).any() or torch.isinf(scores).any():
|
54 |
+
scores.zero_()
|
55 |
+
scores[..., 5] = 5e4
|
56 |
+
return scores
|
57 |
+
|
58 |
+
|
59 |
+
def split_tensor_along_last_dim(
|
60 |
+
tensor: torch.Tensor,
|
61 |
+
num_partitions: int,
|
62 |
+
contiguous_split_chunks: bool = False,
|
63 |
+
) -> List[torch.Tensor]:
|
64 |
+
"""Split a tensor along its last dimension.
|
65 |
+
|
66 |
+
Arguments:
|
67 |
+
tensor: input tensor.
|
68 |
+
num_partitions: number of partitions to split the tensor
|
69 |
+
contiguous_split_chunks: If True, make each chunk contiguous
|
70 |
+
in memory.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
A list of Tensors
|
74 |
+
"""
|
75 |
+
# Get the size and dimension.
|
76 |
+
last_dim = tensor.dim() - 1
|
77 |
+
last_dim_size = tensor.size()[last_dim] // num_partitions
|
78 |
+
# Split.
|
79 |
+
tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
|
80 |
+
# Note: torch.split does not create contiguous tensors by default.
|
81 |
+
if contiguous_split_chunks:
|
82 |
+
return tuple(chunk.contiguous() for chunk in tensor_list)
|
83 |
+
|
84 |
+
return tensor_list
|
85 |
+
|
86 |
+
|
87 |
+
class RotaryEmbedding(nn.Module):
|
88 |
+
def __init__(self, dim, original_impl=False, device=None, dtype=None):
|
89 |
+
super().__init__()
|
90 |
+
inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, device=device, dtype=dtype) / dim))
|
91 |
+
self.register_buffer("inv_freq", inv_freq)
|
92 |
+
self.dim = dim
|
93 |
+
self.original_impl = original_impl
|
94 |
+
|
95 |
+
def forward_original_impl(
|
96 |
+
self, seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000
|
97 |
+
):
|
98 |
+
"""Enhanced Transformer with Rotary Position Embedding.
|
99 |
+
|
100 |
+
Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
|
101 |
+
transformers/rope/__init__.py. MIT License:
|
102 |
+
https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license.
|
103 |
+
"""
|
104 |
+
# $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$
|
105 |
+
theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, dtype=dtype, device=device) / n_elem))
|
106 |
+
|
107 |
+
# Create position indexes `[0, 1, ..., seq_len - 1]`
|
108 |
+
seq_idx = torch.arange(seq_len, dtype=dtype, device=device)
|
109 |
+
|
110 |
+
# Calculate the product of position index and $\theta_i$
|
111 |
+
idx_theta = torch.outer(seq_idx, theta).float()
|
112 |
+
|
113 |
+
cache = torch.stack([torch.cos(idx_theta), torch.sin(idx_theta)], dim=-1)
|
114 |
+
|
115 |
+
# this is to mimic the behaviour of complex32, else we will get different results
|
116 |
+
if dtype in (torch.float16, torch.bfloat16, torch.int8):
|
117 |
+
cache = cache.bfloat16() if dtype == torch.bfloat16 else cache.half()
|
118 |
+
return cache
|
119 |
+
|
120 |
+
def forward(self, max_seq_len, offset=0):
|
121 |
+
if self.original_impl:
|
122 |
+
return self.forward_original_impl(
|
123 |
+
max_seq_len, self.dim, dtype=self.inv_freq.dtype, device=self.inv_freq.device
|
124 |
+
)
|
125 |
+
|
126 |
+
|
127 |
+
@torch.jit.script
|
128 |
+
def apply_rotary_pos_emb_original(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
|
129 |
+
# x: [sq, b, np, hn]
|
130 |
+
sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
|
131 |
+
rot_dim = rope_cache.shape[-2] * 2
|
132 |
+
x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
|
133 |
+
# truncate to support variable sizes
|
134 |
+
rope_cache = rope_cache[:sq]
|
135 |
+
xshaped = x.reshape(sq, -1, np, rot_dim // 2, 2)
|
136 |
+
rope_cache = rope_cache.view(sq, -1, 1, xshaped.size(3), 2)
|
137 |
+
x_out2 = torch.stack(
|
138 |
+
[
|
139 |
+
xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
|
140 |
+
xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
|
141 |
+
],
|
142 |
+
-1,
|
143 |
+
)
|
144 |
+
x_out2 = x_out2.flatten(3)
|
145 |
+
return torch.cat((x_out2, x_pass), dim=-1)
|
146 |
+
|
147 |
+
|
148 |
+
class RMSNorm(torch.nn.Module):
|
149 |
+
def __init__(self, normalized_shape, eps=1e-5, device=None, dtype=None, **kwargs):
|
150 |
+
super().__init__()
|
151 |
+
self.weight = torch.nn.Parameter(torch.empty(normalized_shape, device=device, dtype=dtype))
|
152 |
+
self.eps = eps
|
153 |
+
|
154 |
+
def forward(self, hidden_states: torch.Tensor):
|
155 |
+
input_dtype = hidden_states.dtype
|
156 |
+
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
|
157 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
|
158 |
+
|
159 |
+
return (self.weight * hidden_states).to(input_dtype)
|
160 |
+
|
161 |
+
|
162 |
+
class CoreAttention(torch.nn.Module):
|
163 |
+
def __init__(self, config: ChatGLMConfig, layer_number):
|
164 |
+
super(CoreAttention, self).__init__()
|
165 |
+
|
166 |
+
self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling
|
167 |
+
self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
|
168 |
+
if self.apply_query_key_layer_scaling:
|
169 |
+
self.attention_softmax_in_fp32 = True
|
170 |
+
self.layer_number = max(1, layer_number)
|
171 |
+
|
172 |
+
projection_size = config.kv_channels * config.num_attention_heads
|
173 |
+
|
174 |
+
# Per attention head and per partition values.
|
175 |
+
self.hidden_size_per_partition = projection_size
|
176 |
+
self.hidden_size_per_attention_head = projection_size // config.num_attention_heads
|
177 |
+
self.num_attention_heads_per_partition = config.num_attention_heads
|
178 |
+
|
179 |
+
coeff = None
|
180 |
+
self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
|
181 |
+
if self.apply_query_key_layer_scaling:
|
182 |
+
coeff = self.layer_number
|
183 |
+
self.norm_factor *= coeff
|
184 |
+
self.coeff = coeff
|
185 |
+
|
186 |
+
self.attention_dropout = torch.nn.Dropout(config.attention_dropout)
|
187 |
+
|
188 |
+
def forward(self, query_layer, key_layer, value_layer, attention_mask):
|
189 |
+
pytorch_major_version = int(torch.__version__.split('.')[0])
|
190 |
+
if pytorch_major_version >= 2:
|
191 |
+
query_layer, key_layer, value_layer = [k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer]]
|
192 |
+
if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
|
193 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
|
194 |
+
is_causal=True)
|
195 |
+
else:
|
196 |
+
if attention_mask is not None:
|
197 |
+
attention_mask = ~attention_mask
|
198 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
|
199 |
+
attention_mask)
|
200 |
+
context_layer = context_layer.permute(2, 0, 1, 3)
|
201 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
202 |
+
context_layer = context_layer.reshape(*new_context_layer_shape)
|
203 |
+
else:
|
204 |
+
# Raw attention scores
|
205 |
+
|
206 |
+
# [b, np, sq, sk]
|
207 |
+
output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
|
208 |
+
|
209 |
+
# [sq, b, np, hn] -> [sq, b * np, hn]
|
210 |
+
query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
|
211 |
+
# [sk, b, np, hn] -> [sk, b * np, hn]
|
212 |
+
key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
|
213 |
+
|
214 |
+
# preallocting input tensor: [b * np, sq, sk]
|
215 |
+
matmul_input_buffer = torch.empty(
|
216 |
+
output_size[0] * output_size[1], output_size[2], output_size[3], dtype=query_layer.dtype,
|
217 |
+
device=query_layer.device
|
218 |
+
)
|
219 |
+
|
220 |
+
# Raw attention scores. [b * np, sq, sk]
|
221 |
+
matmul_result = torch.baddbmm(
|
222 |
+
matmul_input_buffer,
|
223 |
+
query_layer.transpose(0, 1), # [b * np, sq, hn]
|
224 |
+
key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk]
|
225 |
+
beta=0.0,
|
226 |
+
alpha=(1.0 / self.norm_factor),
|
227 |
+
)
|
228 |
+
|
229 |
+
# change view to [b, np, sq, sk]
|
230 |
+
attention_scores = matmul_result.view(*output_size)
|
231 |
+
|
232 |
+
# ===========================
|
233 |
+
# Attention probs and dropout
|
234 |
+
# ===========================
|
235 |
+
|
236 |
+
# attention scores and attention mask [b, np, sq, sk]
|
237 |
+
if self.attention_softmax_in_fp32:
|
238 |
+
attention_scores = attention_scores.float()
|
239 |
+
if self.coeff is not None:
|
240 |
+
attention_scores = attention_scores * self.coeff
|
241 |
+
if attention_mask is None and attention_scores.shape[2] == attention_scores.shape[3]:
|
242 |
+
attention_mask = torch.ones(output_size[0], 1, output_size[2], output_size[3],
|
243 |
+
device=attention_scores.device, dtype=torch.bool)
|
244 |
+
attention_mask.tril_()
|
245 |
+
attention_mask = ~attention_mask
|
246 |
+
if attention_mask is not None:
|
247 |
+
attention_scores = attention_scores.masked_fill(attention_mask, float("-inf"))
|
248 |
+
attention_probs = F.softmax(attention_scores, dim=-1)
|
249 |
+
attention_probs = attention_probs.type_as(value_layer)
|
250 |
+
|
251 |
+
# This is actually dropping out entire tokens to attend to, which might
|
252 |
+
# seem a bit unusual, but is taken from the original Transformer paper.
|
253 |
+
attention_probs = self.attention_dropout(attention_probs)
|
254 |
+
# =========================
|
255 |
+
# Context layer. [sq, b, hp]
|
256 |
+
# =========================
|
257 |
+
|
258 |
+
# value_layer -> context layer.
|
259 |
+
# [sk, b, np, hn] --> [b, np, sq, hn]
|
260 |
+
|
261 |
+
# context layer shape: [b, np, sq, hn]
|
262 |
+
output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))
|
263 |
+
# change view [sk, b * np, hn]
|
264 |
+
value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)
|
265 |
+
# change view [b * np, sq, sk]
|
266 |
+
attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
|
267 |
+
# matmul: [b * np, sq, hn]
|
268 |
+
context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
|
269 |
+
# change view [b, np, sq, hn]
|
270 |
+
context_layer = context_layer.view(*output_size)
|
271 |
+
# [b, np, sq, hn] --> [sq, b, np, hn]
|
272 |
+
context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
|
273 |
+
# [sq, b, np, hn] --> [sq, b, hp]
|
274 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
275 |
+
context_layer = context_layer.view(*new_context_layer_shape)
|
276 |
+
|
277 |
+
return context_layer
|
278 |
+
|
279 |
+
|
280 |
+
class SelfAttention(torch.nn.Module):
|
281 |
+
"""Parallel self-attention layer abstract class.
|
282 |
+
|
283 |
+
Self-attention layer takes input with size [s, b, h]
|
284 |
+
and returns output of the same size.
|
285 |
+
"""
|
286 |
+
|
287 |
+
def __init__(self, config: ChatGLMConfig, layer_number, device=None):
|
288 |
+
super(SelfAttention, self).__init__()
|
289 |
+
self.layer_number = max(1, layer_number)
|
290 |
+
|
291 |
+
self.projection_size = config.kv_channels * config.num_attention_heads
|
292 |
+
|
293 |
+
# Per attention head and per partition values.
|
294 |
+
self.hidden_size_per_attention_head = self.projection_size // config.num_attention_heads
|
295 |
+
self.num_attention_heads_per_partition = config.num_attention_heads
|
296 |
+
|
297 |
+
self.multi_query_attention = config.multi_query_attention
|
298 |
+
self.qkv_hidden_size = 3 * self.projection_size
|
299 |
+
if self.multi_query_attention:
|
300 |
+
self.num_multi_query_groups_per_partition = config.multi_query_group_num
|
301 |
+
self.qkv_hidden_size = (
|
302 |
+
self.projection_size + 2 * self.hidden_size_per_attention_head * config.multi_query_group_num
|
303 |
+
)
|
304 |
+
self.query_key_value = nn.Linear(config.hidden_size, self.qkv_hidden_size,
|
305 |
+
bias=config.add_bias_linear or config.add_qkv_bias,
|
306 |
+
device=device, **_config_to_kwargs(config)
|
307 |
+
)
|
308 |
+
|
309 |
+
self.core_attention = CoreAttention(config, self.layer_number)
|
310 |
+
|
311 |
+
# Output.
|
312 |
+
self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear,
|
313 |
+
device=device, **_config_to_kwargs(config)
|
314 |
+
)
|
315 |
+
|
316 |
+
self.interleaved_qkv = config.interleaved_qkv
|
317 |
+
|
318 |
+
def _allocate_memory(self, inference_max_sequence_len, batch_size, device=None, dtype=None):
|
319 |
+
if self.multi_query_attention:
|
320 |
+
num_attention_heads = self.num_multi_query_groups_per_partition
|
321 |
+
else:
|
322 |
+
num_attention_heads = self.num_attention_heads_per_partition
|
323 |
+
return torch.empty(
|
324 |
+
inference_max_sequence_len,
|
325 |
+
batch_size,
|
326 |
+
num_attention_heads,
|
327 |
+
self.hidden_size_per_attention_head,
|
328 |
+
dtype=dtype,
|
329 |
+
device=device,
|
330 |
+
)
|
331 |
+
|
332 |
+
def forward(
|
333 |
+
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
|
334 |
+
):
|
335 |
+
# hidden_states: [sq, b, h]
|
336 |
+
|
337 |
+
# =================================================
|
338 |
+
# Pre-allocate memory for key-values for inference.
|
339 |
+
# =================================================
|
340 |
+
# =====================
|
341 |
+
# Query, Key, and Value
|
342 |
+
# =====================
|
343 |
+
|
344 |
+
# Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
|
345 |
+
mixed_x_layer = self.query_key_value(hidden_states)
|
346 |
+
|
347 |
+
if self.multi_query_attention:
|
348 |
+
(query_layer, key_layer, value_layer) = mixed_x_layer.split(
|
349 |
+
[
|
350 |
+
self.num_attention_heads_per_partition * self.hidden_size_per_attention_head,
|
351 |
+
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
|
352 |
+
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
|
353 |
+
],
|
354 |
+
dim=-1,
|
355 |
+
)
|
356 |
+
query_layer = query_layer.view(
|
357 |
+
query_layer.size()[:-1] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
358 |
+
)
|
359 |
+
key_layer = key_layer.view(
|
360 |
+
key_layer.size()[:-1] + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
|
361 |
+
)
|
362 |
+
value_layer = value_layer.view(
|
363 |
+
value_layer.size()[:-1]
|
364 |
+
+ (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
|
365 |
+
)
|
366 |
+
else:
|
367 |
+
if self.interleaved_qkv:
|
368 |
+
new_tensor_shape = mixed_x_layer.size()[:-1] + \
|
369 |
+
(self.num_attention_heads_per_partition,
|
370 |
+
3 * self.hidden_size_per_attention_head)
|
371 |
+
mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
|
372 |
+
|
373 |
+
# [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
|
374 |
+
(query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
|
375 |
+
|
376 |
+
if not self.interleaved_qkv:
|
377 |
+
query_layer = query_layer.view(
|
378 |
+
query_layer.size()[:-1] + (
|
379 |
+
self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
380 |
+
).contiguous()
|
381 |
+
key_layer = key_layer.view(
|
382 |
+
key_layer.size()[:-1] + (
|
383 |
+
self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
384 |
+
).contiguous()
|
385 |
+
value_layer = value_layer.view(
|
386 |
+
value_layer.size()[:-1] + (
|
387 |
+
self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
388 |
+
).contiguous()
|
389 |
+
|
390 |
+
# apply relative positional encoding (rotary embedding)
|
391 |
+
if rotary_pos_emb is not None:
|
392 |
+
query_layer = apply_rotary_pos_emb_original(query_layer, rotary_pos_emb)
|
393 |
+
key_layer = apply_rotary_pos_emb_original(key_layer, rotary_pos_emb)
|
394 |
+
|
395 |
+
# adjust key and value for inference
|
396 |
+
if use_cache:
|
397 |
+
if kv_cache is not None:
|
398 |
+
cache_k, cache_v = kv_cache
|
399 |
+
key_layer = torch.cat((cache_k, key_layer), dim=0)
|
400 |
+
value_layer = torch.cat((cache_v, value_layer), dim=0)
|
401 |
+
kv_cache = (key_layer, value_layer)
|
402 |
+
else:
|
403 |
+
kv_cache = None
|
404 |
+
|
405 |
+
if self.multi_query_attention:
|
406 |
+
key_layer = key_layer.unsqueeze(-2)
|
407 |
+
key_layer = key_layer.expand(
|
408 |
+
-1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1
|
409 |
+
)
|
410 |
+
key_layer = key_layer.contiguous().view(
|
411 |
+
key_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
412 |
+
)
|
413 |
+
value_layer = value_layer.unsqueeze(-2)
|
414 |
+
value_layer = value_layer.expand(
|
415 |
+
-1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1
|
416 |
+
)
|
417 |
+
value_layer = value_layer.contiguous().view(
|
418 |
+
value_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
419 |
+
)
|
420 |
+
|
421 |
+
# ==================================
|
422 |
+
# core attention computation
|
423 |
+
# ==================================
|
424 |
+
|
425 |
+
context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask)
|
426 |
+
|
427 |
+
# =================
|
428 |
+
# Output. [sq, b, h]
|
429 |
+
# =================
|
430 |
+
|
431 |
+
output = self.dense(context_layer)
|
432 |
+
|
433 |
+
return output, kv_cache
|
434 |
+
|
435 |
+
|
436 |
+
def _config_to_kwargs(args):
|
437 |
+
common_kwargs = {
|
438 |
+
"dtype": args.torch_dtype,
|
439 |
+
}
|
440 |
+
return common_kwargs
|
441 |
+
|
442 |
+
|
443 |
+
class MLP(torch.nn.Module):
|
444 |
+
"""MLP.
|
445 |
+
|
446 |
+
MLP will take the input with h hidden state, project it to 4*h
|
447 |
+
hidden dimension, perform nonlinear transformation, and project the
|
448 |
+
state back into h hidden dimension.
|
449 |
+
"""
|
450 |
+
|
451 |
+
def __init__(self, config: ChatGLMConfig, device=None):
|
452 |
+
super(MLP, self).__init__()
|
453 |
+
|
454 |
+
self.add_bias = config.add_bias_linear
|
455 |
+
|
456 |
+
# Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
|
457 |
+
self.dense_h_to_4h = nn.Linear(
|
458 |
+
config.hidden_size,
|
459 |
+
config.ffn_hidden_size * 2,
|
460 |
+
bias=self.add_bias,
|
461 |
+
device=device,
|
462 |
+
**_config_to_kwargs(config)
|
463 |
+
)
|
464 |
+
|
465 |
+
def swiglu(x):
|
466 |
+
x = torch.chunk(x, 2, dim=-1)
|
467 |
+
return F.silu(x[0]) * x[1]
|
468 |
+
|
469 |
+
self.activation_func = swiglu
|
470 |
+
|
471 |
+
# Project back to h.
|
472 |
+
self.dense_4h_to_h = nn.Linear(
|
473 |
+
config.ffn_hidden_size,
|
474 |
+
config.hidden_size,
|
475 |
+
bias=self.add_bias,
|
476 |
+
device=device,
|
477 |
+
**_config_to_kwargs(config)
|
478 |
+
)
|
479 |
+
|
480 |
+
def forward(self, hidden_states):
|
481 |
+
# [s, b, 4hp]
|
482 |
+
intermediate_parallel = self.dense_h_to_4h(hidden_states)
|
483 |
+
intermediate_parallel = self.activation_func(intermediate_parallel)
|
484 |
+
# [s, b, h]
|
485 |
+
output = self.dense_4h_to_h(intermediate_parallel)
|
486 |
+
return output
|
487 |
+
|
488 |
+
|
489 |
+
class GLMBlock(torch.nn.Module):
|
490 |
+
"""A single transformer layer.
|
491 |
+
|
492 |
+
Transformer layer takes input with size [s, b, h] and returns an
|
493 |
+
output of the same size.
|
494 |
+
"""
|
495 |
+
|
496 |
+
def __init__(self, config: ChatGLMConfig, layer_number, device=None):
|
497 |
+
super(GLMBlock, self).__init__()
|
498 |
+
self.layer_number = layer_number
|
499 |
+
|
500 |
+
self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
|
501 |
+
|
502 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
503 |
+
|
504 |
+
LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
|
505 |
+
# Layernorm on the input data.
|
506 |
+
self.input_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
|
507 |
+
dtype=config.torch_dtype)
|
508 |
+
|
509 |
+
# Self attention.
|
510 |
+
self.self_attention = SelfAttention(config, layer_number, device=device)
|
511 |
+
self.hidden_dropout = config.hidden_dropout
|
512 |
+
|
513 |
+
# Layernorm on the attention output
|
514 |
+
self.post_attention_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
|
515 |
+
dtype=config.torch_dtype)
|
516 |
+
|
517 |
+
# MLP
|
518 |
+
self.mlp = MLP(config, device=device)
|
519 |
+
|
520 |
+
def forward(
|
521 |
+
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True,
|
522 |
+
):
|
523 |
+
# hidden_states: [s, b, h]
|
524 |
+
|
525 |
+
# Layer norm at the beginning of the transformer layer.
|
526 |
+
layernorm_output = self.input_layernorm(hidden_states)
|
527 |
+
# Self attention.
|
528 |
+
attention_output, kv_cache = self.self_attention(
|
529 |
+
layernorm_output,
|
530 |
+
attention_mask,
|
531 |
+
rotary_pos_emb,
|
532 |
+
kv_cache=kv_cache,
|
533 |
+
use_cache=use_cache
|
534 |
+
)
|
535 |
+
|
536 |
+
# Residual connection.
|
537 |
+
if self.apply_residual_connection_post_layernorm:
|
538 |
+
residual = layernorm_output
|
539 |
+
else:
|
540 |
+
residual = hidden_states
|
541 |
+
|
542 |
+
layernorm_input = torch.nn.functional.dropout(attention_output, p=self.hidden_dropout, training=self.training)
|
543 |
+
layernorm_input = residual + layernorm_input
|
544 |
+
|
545 |
+
# Layer norm post the self attention.
|
546 |
+
layernorm_output = self.post_attention_layernorm(layernorm_input)
|
547 |
+
|
548 |
+
# MLP.
|
549 |
+
mlp_output = self.mlp(layernorm_output)
|
550 |
+
|
551 |
+
# Second residual connection.
|
552 |
+
if self.apply_residual_connection_post_layernorm:
|
553 |
+
residual = layernorm_output
|
554 |
+
else:
|
555 |
+
residual = layernorm_input
|
556 |
+
|
557 |
+
output = torch.nn.functional.dropout(mlp_output, p=self.hidden_dropout, training=self.training)
|
558 |
+
output = residual + output
|
559 |
+
|
560 |
+
return output, kv_cache
|
561 |
+
|
562 |
+
|
563 |
+
class GLMTransformer(torch.nn.Module):
|
564 |
+
"""Transformer class."""
|
565 |
+
|
566 |
+
def __init__(self, config: ChatGLMConfig, device=None):
|
567 |
+
super(GLMTransformer, self).__init__()
|
568 |
+
|
569 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
570 |
+
self.post_layer_norm = config.post_layer_norm
|
571 |
+
|
572 |
+
# Number of layers.
|
573 |
+
self.num_layers = config.num_layers
|
574 |
+
|
575 |
+
# Transformer layers.
|
576 |
+
def build_layer(layer_number):
|
577 |
+
return GLMBlock(config, layer_number, device=device)
|
578 |
+
|
579 |
+
self.layers = torch.nn.ModuleList([build_layer(i + 1) for i in range(self.num_layers)])
|
580 |
+
|
581 |
+
if self.post_layer_norm:
|
582 |
+
LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
|
583 |
+
# Final layer norm before output.
|
584 |
+
self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
|
585 |
+
dtype=config.torch_dtype)
|
586 |
+
|
587 |
+
def _get_layer(self, layer_number):
|
588 |
+
return self.layers[layer_number]
|
589 |
+
|
590 |
+
def forward(
|
591 |
+
self, hidden_states, attention_mask, rotary_pos_emb, kv_caches=None,
|
592 |
+
use_cache: Optional[bool] = True,
|
593 |
+
output_hidden_states: Optional[bool] = False,
|
594 |
+
):
|
595 |
+
if not kv_caches:
|
596 |
+
kv_caches = [None for _ in range(self.num_layers)]
|
597 |
+
presents = () if use_cache else None
|
598 |
+
all_self_attentions = None
|
599 |
+
all_hidden_states = () if output_hidden_states else None
|
600 |
+
for index in range(self.num_layers):
|
601 |
+
if output_hidden_states:
|
602 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
603 |
+
|
604 |
+
layer = self._get_layer(index)
|
605 |
+
|
606 |
+
hidden_states, kv_cache = layer(
|
607 |
+
hidden_states,
|
608 |
+
attention_mask,
|
609 |
+
rotary_pos_emb,
|
610 |
+
kv_cache=kv_caches[index],
|
611 |
+
use_cache=use_cache
|
612 |
+
)
|
613 |
+
if use_cache:
|
614 |
+
presents = presents + (kv_cache,)
|
615 |
+
|
616 |
+
if output_hidden_states:
|
617 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
618 |
+
|
619 |
+
# Final layer norm.
|
620 |
+
if self.post_layer_norm:
|
621 |
+
hidden_states = self.final_layernorm(hidden_states)
|
622 |
+
|
623 |
+
return hidden_states, presents, all_hidden_states, all_self_attentions
|
624 |
+
|
625 |
+
|
626 |
+
class ChatGLMPreTrainedModel(PreTrainedModel):
|
627 |
+
"""
|
628 |
+
An abstract class to handle weights initialization and
|
629 |
+
a simple interface for downloading and loading pretrained models.
|
630 |
+
"""
|
631 |
+
|
632 |
+
is_parallelizable = False
|
633 |
+
supports_gradient_checkpointing = True
|
634 |
+
config_class = ChatGLMConfig
|
635 |
+
base_model_prefix = "transformer"
|
636 |
+
_no_split_modules = ["GLMBlock"]
|
637 |
+
|
638 |
+
def _init_weights(self, module: nn.Module):
|
639 |
+
"""Initialize the weights."""
|
640 |
+
return
|
641 |
+
|
642 |
+
def get_masks(self, input_ids, past_key_values, padding_mask=None):
|
643 |
+
batch_size, seq_length = input_ids.shape
|
644 |
+
full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device)
|
645 |
+
full_attention_mask.tril_()
|
646 |
+
past_length = 0
|
647 |
+
if past_key_values:
|
648 |
+
past_length = past_key_values[0][0].shape[0]
|
649 |
+
if past_length:
|
650 |
+
full_attention_mask = torch.cat((torch.ones(batch_size, seq_length, past_length,
|
651 |
+
device=input_ids.device), full_attention_mask), dim=-1)
|
652 |
+
if padding_mask is not None:
|
653 |
+
full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1)
|
654 |
+
if not past_length and padding_mask is not None:
|
655 |
+
full_attention_mask -= padding_mask.unsqueeze(-1) - 1
|
656 |
+
full_attention_mask = (full_attention_mask < 0.5).bool()
|
657 |
+
full_attention_mask.unsqueeze_(1)
|
658 |
+
return full_attention_mask
|
659 |
+
|
660 |
+
def get_position_ids(self, input_ids, device):
|
661 |
+
batch_size, seq_length = input_ids.shape
|
662 |
+
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
|
663 |
+
return position_ids
|
664 |
+
|
665 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
666 |
+
if isinstance(module, ChatGLMModel):
|
667 |
+
module.gradient_checkpointing = value
|
668 |
+
|
669 |
+
|
670 |
+
class Embedding(torch.nn.Module):
|
671 |
+
"""Language model embeddings."""
|
672 |
+
|
673 |
+
def __init__(self, config: ChatGLMConfig, device=None):
|
674 |
+
super(Embedding, self).__init__()
|
675 |
+
|
676 |
+
self.hidden_size = config.hidden_size
|
677 |
+
# Word embeddings (parallel).
|
678 |
+
self.word_embeddings = nn.Embedding(
|
679 |
+
config.padded_vocab_size,
|
680 |
+
self.hidden_size,
|
681 |
+
dtype=config.torch_dtype,
|
682 |
+
device=device
|
683 |
+
)
|
684 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
685 |
+
|
686 |
+
def forward(self, input_ids):
|
687 |
+
# Embeddings.
|
688 |
+
words_embeddings = self.word_embeddings(input_ids)
|
689 |
+
embeddings = words_embeddings
|
690 |
+
# Data format change to avoid explicit tranposes : [b s h] --> [s b h].
|
691 |
+
embeddings = embeddings.transpose(0, 1).contiguous()
|
692 |
+
# If the input flag for fp32 residual connection is set, convert for float.
|
693 |
+
if self.fp32_residual_connection:
|
694 |
+
embeddings = embeddings.float()
|
695 |
+
return embeddings
|
696 |
+
|
697 |
+
|
698 |
+
class ChatGLMModel(ChatGLMPreTrainedModel):
|
699 |
+
def __init__(self, config: ChatGLMConfig, device=None, empty_init=True):
|
700 |
+
super().__init__(config)
|
701 |
+
if empty_init:
|
702 |
+
init_method = skip_init
|
703 |
+
else:
|
704 |
+
init_method = default_init
|
705 |
+
init_kwargs = {}
|
706 |
+
if device is not None:
|
707 |
+
init_kwargs["device"] = device
|
708 |
+
self.embedding = init_method(Embedding, config, **init_kwargs)
|
709 |
+
|
710 |
+
# Rotary positional embeddings
|
711 |
+
self.seq_length = config.seq_length
|
712 |
+
rotary_dim = (
|
713 |
+
config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
|
714 |
+
)
|
715 |
+
|
716 |
+
if config.rotary_percent < 1.0:
|
717 |
+
rotary_dim = int(rotary_dim * config.rotary_percent)
|
718 |
+
|
719 |
+
# partial rotary embeddings, which is better than full rotary
|
720 |
+
# Wang and Komatsuzaki et al
|
721 |
+
# https://github.com/kingoflolz/mesh-transformer-jax/
|
722 |
+
self.rotary_pos_emb = RotaryEmbedding(rotary_dim, original_impl=config.original_rope, device=device,
|
723 |
+
dtype=config.torch_dtype)
|
724 |
+
self.encoder = init_method(GLMTransformer, config, **init_kwargs)
|
725 |
+
self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
|
726 |
+
dtype=config.torch_dtype, **init_kwargs)
|
727 |
+
self.gradient_checkpointing = False
|
728 |
+
|
729 |
+
def forward(
|
730 |
+
self,
|
731 |
+
input_ids,
|
732 |
+
position_ids: Optional[torch.Tensor] = None,
|
733 |
+
attention_mask: Optional[torch.BoolTensor] = None,
|
734 |
+
full_attention_mask: Optional[torch.BoolTensor] = None,
|
735 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
|
736 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
737 |
+
use_cache: Optional[bool] = None,
|
738 |
+
output_hidden_states: Optional[bool] = None,
|
739 |
+
return_dict: Optional[bool] = None,
|
740 |
+
):
|
741 |
+
output_hidden_states = (
|
742 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
743 |
+
)
|
744 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
745 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
746 |
+
|
747 |
+
batch_size, seq_length = input_ids.shape
|
748 |
+
|
749 |
+
if inputs_embeds is None:
|
750 |
+
inputs_embeds = self.embedding(input_ids)
|
751 |
+
|
752 |
+
if full_attention_mask is None:
|
753 |
+
if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1):
|
754 |
+
full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)
|
755 |
+
|
756 |
+
# Rotary positional embeddings
|
757 |
+
rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
|
758 |
+
if position_ids is not None:
|
759 |
+
rotary_pos_emb = rotary_pos_emb[position_ids]
|
760 |
+
else:
|
761 |
+
rotary_pos_emb = rotary_pos_emb[None, :seq_length]
|
762 |
+
rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
|
763 |
+
|
764 |
+
# Run encoder.
|
765 |
+
hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
|
766 |
+
inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb,
|
767 |
+
kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
|
768 |
+
)
|
769 |
+
|
770 |
+
if not return_dict:
|
771 |
+
return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
|
772 |
+
|
773 |
+
return BaseModelOutputWithPast(
|
774 |
+
last_hidden_state=hidden_states,
|
775 |
+
past_key_values=presents,
|
776 |
+
hidden_states=all_hidden_states,
|
777 |
+
attentions=all_self_attentions,
|
778 |
+
)
|
779 |
+
|
780 |
+
def quantize(self, weight_bit_width: int):
|
781 |
+
from .quantization import quantize
|
782 |
+
quantize(self.encoder, weight_bit_width)
|
783 |
+
return self
|
784 |
+
|
785 |
+
|
786 |
+
class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
787 |
+
def __init__(self, config: ChatGLMConfig, empty_init=True, device=None):
|
788 |
+
super().__init__(config)
|
789 |
+
|
790 |
+
self.max_sequence_length = config.max_length
|
791 |
+
self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
|
792 |
+
self.config = config
|
793 |
+
self.quantized = False
|
794 |
+
|
795 |
+
if self.config.quantization_bit:
|
796 |
+
self.quantize(self.config.quantization_bit, empty_init=True)
|
797 |
+
|
798 |
+
def _update_model_kwargs_for_generation(
|
799 |
+
self,
|
800 |
+
outputs: ModelOutput,
|
801 |
+
model_kwargs: Dict[str, Any],
|
802 |
+
is_encoder_decoder: bool = False,
|
803 |
+
standardize_cache_format: bool = False,
|
804 |
+
) -> Dict[str, Any]:
|
805 |
+
# update past_key_values
|
806 |
+
model_kwargs["past_key_values"] = self._extract_past_from_model_output(
|
807 |
+
outputs, standardize_cache_format=standardize_cache_format
|
808 |
+
)
|
809 |
+
|
810 |
+
# update attention mask
|
811 |
+
if "attention_mask" in model_kwargs:
|
812 |
+
attention_mask = model_kwargs["attention_mask"]
|
813 |
+
model_kwargs["attention_mask"] = torch.cat(
|
814 |
+
[attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
|
815 |
+
)
|
816 |
+
|
817 |
+
# update position ids
|
818 |
+
if "position_ids" in model_kwargs:
|
819 |
+
position_ids = model_kwargs["position_ids"]
|
820 |
+
new_position_id = position_ids[..., -1:].clone()
|
821 |
+
new_position_id += 1
|
822 |
+
model_kwargs["position_ids"] = torch.cat(
|
823 |
+
[position_ids, new_position_id], dim=-1
|
824 |
+
)
|
825 |
+
|
826 |
+
model_kwargs["is_first_forward"] = False
|
827 |
+
return model_kwargs
|
828 |
+
|
829 |
+
def prepare_inputs_for_generation(
|
830 |
+
self,
|
831 |
+
input_ids: torch.LongTensor,
|
832 |
+
past_key_values: Optional[torch.Tensor] = None,
|
833 |
+
attention_mask: Optional[torch.Tensor] = None,
|
834 |
+
position_ids: Optional[torch.Tensor] = None,
|
835 |
+
is_first_forward: bool = True,
|
836 |
+
**kwargs
|
837 |
+
) -> dict:
|
838 |
+
# only last token for input_ids if past is not None
|
839 |
+
if position_ids is None:
|
840 |
+
position_ids = self.get_position_ids(input_ids, device=input_ids.device)
|
841 |
+
if not is_first_forward:
|
842 |
+
position_ids = position_ids[..., -1:]
|
843 |
+
input_ids = input_ids[:, -1:]
|
844 |
+
return {
|
845 |
+
"input_ids": input_ids,
|
846 |
+
"past_key_values": past_key_values,
|
847 |
+
"position_ids": position_ids,
|
848 |
+
"attention_mask": attention_mask,
|
849 |
+
"return_last_logit": True
|
850 |
+
}
|
851 |
+
|
852 |
+
def forward(
|
853 |
+
self,
|
854 |
+
input_ids: Optional[torch.Tensor] = None,
|
855 |
+
position_ids: Optional[torch.Tensor] = None,
|
856 |
+
attention_mask: Optional[torch.Tensor] = None,
|
857 |
+
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
|
858 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
859 |
+
labels: Optional[torch.Tensor] = None,
|
860 |
+
use_cache: Optional[bool] = None,
|
861 |
+
output_attentions: Optional[bool] = None,
|
862 |
+
output_hidden_states: Optional[bool] = None,
|
863 |
+
return_dict: Optional[bool] = None,
|
864 |
+
return_last_logit: Optional[bool] = False,
|
865 |
+
):
|
866 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
867 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
868 |
+
|
869 |
+
transformer_outputs = self.transformer(
|
870 |
+
input_ids=input_ids,
|
871 |
+
position_ids=position_ids,
|
872 |
+
attention_mask=attention_mask,
|
873 |
+
past_key_values=past_key_values,
|
874 |
+
inputs_embeds=inputs_embeds,
|
875 |
+
use_cache=use_cache,
|
876 |
+
output_hidden_states=output_hidden_states,
|
877 |
+
return_dict=return_dict,
|
878 |
+
)
|
879 |
+
|
880 |
+
hidden_states = transformer_outputs[0]
|
881 |
+
if return_last_logit:
|
882 |
+
hidden_states = hidden_states[-1:]
|
883 |
+
lm_logits = self.transformer.output_layer(hidden_states)
|
884 |
+
lm_logits = lm_logits.transpose(0, 1).contiguous()
|
885 |
+
|
886 |
+
loss = None
|
887 |
+
if labels is not None:
|
888 |
+
lm_logits = lm_logits.to(torch.float32)
|
889 |
+
|
890 |
+
# Shift so that tokens < n predict n
|
891 |
+
shift_logits = lm_logits[..., :-1, :].contiguous()
|
892 |
+
shift_labels = labels[..., 1:].contiguous()
|
893 |
+
# Flatten the tokens
|
894 |
+
loss_fct = CrossEntropyLoss(ignore_index=-100)
|
895 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
896 |
+
|
897 |
+
lm_logits = lm_logits.to(hidden_states.dtype)
|
898 |
+
loss = loss.to(hidden_states.dtype)
|
899 |
+
|
900 |
+
if not return_dict:
|
901 |
+
output = (lm_logits,) + transformer_outputs[1:]
|
902 |
+
return ((loss,) + output) if loss is not None else output
|
903 |
+
|
904 |
+
return CausalLMOutputWithPast(
|
905 |
+
loss=loss,
|
906 |
+
logits=lm_logits,
|
907 |
+
past_key_values=transformer_outputs.past_key_values,
|
908 |
+
hidden_states=transformer_outputs.hidden_states,
|
909 |
+
attentions=transformer_outputs.attentions,
|
910 |
+
)
|
911 |
+
|
912 |
+
@staticmethod
|
913 |
+
def _reorder_cache(
|
914 |
+
past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
|
915 |
+
) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
|
916 |
+
"""
|
917 |
+
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
918 |
+
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
919 |
+
beam_idx at every generation step.
|
920 |
+
|
921 |
+
Output shares the same memory storage as `past`.
|
922 |
+
"""
|
923 |
+
return tuple(
|
924 |
+
(
|
925 |
+
layer_past[0].index_select(1, beam_idx.to(layer_past[0].device)),
|
926 |
+
layer_past[1].index_select(1, beam_idx.to(layer_past[1].device)),
|
927 |
+
)
|
928 |
+
for layer_past in past
|
929 |
+
)
|
930 |
+
|
931 |
+
def process_response(self, response):
|
932 |
+
response = response.strip()
|
933 |
+
response = response.replace("[[训练时间]]", "2023年")
|
934 |
+
return response
|
935 |
+
|
936 |
+
def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = None):
|
937 |
+
prompt = ""
|
938 |
+
for i, (old_query, response) in enumerate(history):
|
939 |
+
prompt += "[Round {}]\n\n问:{}\n\n答:{}\n\n".format(i + 1, old_query, response)
|
940 |
+
prompt += "[Round {}]\n\n问:{}\n\n答:".format(len(history) + 1, query)
|
941 |
+
inputs = tokenizer([prompt], return_tensors="pt")
|
942 |
+
inputs = inputs.to(self.device)
|
943 |
+
return inputs
|
944 |
+
|
945 |
+
def build_stream_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = None):
|
946 |
+
if history:
|
947 |
+
prompt = "\n\n[Round {}]\n\n问:{}\n\n答:".format(len(history) + 1, query)
|
948 |
+
input_ids = tokenizer.encode(prompt, add_special_tokens=False)
|
949 |
+
input_ids = input_ids[1:]
|
950 |
+
inputs = tokenizer.batch_encode_plus([(input_ids, None)], return_tensors="pt", add_special_tokens=False)
|
951 |
+
else:
|
952 |
+
prompt = "[Round {}]\n\n问:{}\n\n答:".format(len(history) + 1, query)
|
953 |
+
inputs = tokenizer([prompt], return_tensors="pt")
|
954 |
+
inputs = inputs.to(self.device)
|
955 |
+
return inputs
|
956 |
+
|
957 |
+
|
958 |
+
@torch.no_grad()
|
959 |
+
def chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048, num_beams=1,
|
960 |
+
do_sample=True, top_p=0.8, temperature=0.8, logits_processor=None, **kwargs):
|
961 |
+
if history is None:
|
962 |
+
history = []
|
963 |
+
if logits_processor is None:
|
964 |
+
logits_processor = LogitsProcessorList()
|
965 |
+
logits_processor.append(InvalidScoreLogitsProcessor())
|
966 |
+
gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
|
967 |
+
"temperature": temperature, "logits_processor": logits_processor, **kwargs}
|
968 |
+
inputs = self.build_inputs(tokenizer, query, history=history)
|
969 |
+
outputs = self.generate(**inputs, **gen_kwargs)
|
970 |
+
outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
|
971 |
+
response = tokenizer.decode(outputs)
|
972 |
+
response = self.process_response(response)
|
973 |
+
history = history + [(query, response)]
|
974 |
+
return response, history
|
975 |
+
|
976 |
+
@torch.no_grad()
|
977 |
+
def stream_chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, past_key_values=None,
|
978 |
+
max_length: int = 2048, do_sample=True, top_p=0.8, temperature=0.8, logits_processor=None,
|
979 |
+
return_past_key_values=False, **kwargs):
|
980 |
+
if history is None:
|
981 |
+
history = []
|
982 |
+
if logits_processor is None:
|
983 |
+
logits_processor = LogitsProcessorList()
|
984 |
+
logits_processor.append(InvalidScoreLogitsProcessor())
|
985 |
+
gen_kwargs = {"max_length": max_length, "do_sample": do_sample, "top_p": top_p,
|
986 |
+
"temperature": temperature, "logits_processor": logits_processor, **kwargs}
|
987 |
+
if past_key_values is None and not return_past_key_values:
|
988 |
+
inputs = self.build_inputs(tokenizer, query, history=history)
|
989 |
+
else:
|
990 |
+
inputs = self.build_stream_inputs(tokenizer, query, history=history)
|
991 |
+
if past_key_values is not None:
|
992 |
+
past_length = past_key_values[0][0].shape[0]
|
993 |
+
inputs.position_ids += past_length
|
994 |
+
attention_mask = inputs.attention_mask
|
995 |
+
attention_mask = torch.cat((attention_mask.new_ones(1, past_length), attention_mask), dim=1)
|
996 |
+
inputs['attention_mask'] = attention_mask
|
997 |
+
for outputs in self.stream_generate(**inputs, past_key_values=past_key_values,
|
998 |
+
return_past_key_values=return_past_key_values, **gen_kwargs):
|
999 |
+
if return_past_key_values:
|
1000 |
+
outputs, past_key_values = outputs
|
1001 |
+
outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
|
1002 |
+
response = tokenizer.decode(outputs)
|
1003 |
+
response = self.process_response(response)
|
1004 |
+
new_history = history + [(query, response)]
|
1005 |
+
if return_past_key_values:
|
1006 |
+
yield response, new_history, past_key_values
|
1007 |
+
else:
|
1008 |
+
yield response, new_history
|
1009 |
+
|
1010 |
+
@torch.no_grad()
|
1011 |
+
def stream_generate(
|
1012 |
+
self,
|
1013 |
+
input_ids,
|
1014 |
+
generation_config: Optional[GenerationConfig] = None,
|
1015 |
+
logits_processor: Optional[LogitsProcessorList] = None,
|
1016 |
+
stopping_criteria: Optional[StoppingCriteriaList] = None,
|
1017 |
+
prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
|
1018 |
+
return_past_key_values=False,
|
1019 |
+
**kwargs,
|
1020 |
+
):
|
1021 |
+
batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]
|
1022 |
+
|
1023 |
+
if generation_config is None:
|
1024 |
+
generation_config = self.generation_config
|
1025 |
+
generation_config = copy.deepcopy(generation_config)
|
1026 |
+
model_kwargs = generation_config.update(**kwargs)
|
1027 |
+
bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id
|
1028 |
+
|
1029 |
+
if isinstance(eos_token_id, int):
|
1030 |
+
eos_token_id = [eos_token_id]
|
1031 |
+
|
1032 |
+
has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
|
1033 |
+
if has_default_max_length and generation_config.max_new_tokens is None:
|
1034 |
+
warnings.warn(
|
1035 |
+
f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. "
|
1036 |
+
"This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we"
|
1037 |
+
" recommend using `max_new_tokens` to control the maximum length of the generation.",
|
1038 |
+
UserWarning,
|
1039 |
+
)
|
1040 |
+
elif generation_config.max_new_tokens is not None:
|
1041 |
+
generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
|
1042 |
+
if not has_default_max_length:
|
1043 |
+
logger.warn(
|
1044 |
+
f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
|
1045 |
+
f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
|
1046 |
+
"Please refer to the documentation for more information. "
|
1047 |
+
"(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)",
|
1048 |
+
UserWarning,
|
1049 |
+
)
|
1050 |
+
|
1051 |
+
if input_ids_seq_length >= generation_config.max_length:
|
1052 |
+
input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
|
1053 |
+
logger.warning(
|
1054 |
+
f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
|
1055 |
+
f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
|
1056 |
+
" increasing `max_new_tokens`."
|
1057 |
+
)
|
1058 |
+
|
1059 |
+
# 2. Set generation parameters if not already defined
|
1060 |
+
logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
|
1061 |
+
stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
|
1062 |
+
|
1063 |
+
logits_processor = self._get_logits_processor(
|
1064 |
+
generation_config=generation_config,
|
1065 |
+
input_ids_seq_length=input_ids_seq_length,
|
1066 |
+
encoder_input_ids=input_ids,
|
1067 |
+
prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
|
1068 |
+
logits_processor=logits_processor,
|
1069 |
+
)
|
1070 |
+
|
1071 |
+
stopping_criteria = self._get_stopping_criteria(
|
1072 |
+
generation_config=generation_config, stopping_criteria=stopping_criteria
|
1073 |
+
)
|
1074 |
+
logits_warper = self._get_logits_warper(generation_config)
|
1075 |
+
|
1076 |
+
unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
|
1077 |
+
scores = None
|
1078 |
+
while True:
|
1079 |
+
model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
|
1080 |
+
# forward pass to get next token
|
1081 |
+
outputs = self(
|
1082 |
+
**model_inputs,
|
1083 |
+
return_dict=True,
|
1084 |
+
output_attentions=False,
|
1085 |
+
output_hidden_states=False,
|
1086 |
+
)
|
1087 |
+
|
1088 |
+
next_token_logits = outputs.logits[:, -1, :]
|
1089 |
+
|
1090 |
+
# pre-process distribution
|
1091 |
+
next_token_scores = logits_processor(input_ids, next_token_logits)
|
1092 |
+
next_token_scores = logits_warper(input_ids, next_token_scores)
|
1093 |
+
|
1094 |
+
# sample
|
1095 |
+
probs = nn.functional.softmax(next_token_scores, dim=-1)
|
1096 |
+
if generation_config.do_sample:
|
1097 |
+
next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
|
1098 |
+
else:
|
1099 |
+
next_tokens = torch.argmax(probs, dim=-1)
|
1100 |
+
|
1101 |
+
# update generated ids, model inputs, and length for next step
|
1102 |
+
input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
|
1103 |
+
model_kwargs = self._update_model_kwargs_for_generation(
|
1104 |
+
outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
|
1105 |
+
)
|
1106 |
+
unfinished_sequences = unfinished_sequences.mul((sum(next_tokens != i for i in eos_token_id)).long())
|
1107 |
+
if return_past_key_values:
|
1108 |
+
yield input_ids, outputs.past_key_values
|
1109 |
+
else:
|
1110 |
+
yield input_ids
|
1111 |
+
# stop when each sentence is finished, or if we exceed the maximum length
|
1112 |
+
if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
|
1113 |
+
break
|
1114 |
+
|
1115 |
+
def quantize(self, bits: int, empty_init=False, device=None, **kwargs):
|
1116 |
+
if bits == 0:
|
1117 |
+
return
|
1118 |
+
|
1119 |
+
from .quantization import quantize
|
1120 |
+
|
1121 |
+
if self.quantized:
|
1122 |
+
logger.info("Already quantized.")
|
1123 |
+
return self
|
1124 |
+
|
1125 |
+
self.quantized = True
|
1126 |
+
|
1127 |
+
self.config.quantization_bit = bits
|
1128 |
+
|
1129 |
+
self.transformer.encoder = quantize(self.transformer.encoder, bits, empty_init=empty_init, device=device,
|
1130 |
+
**kwargs)
|
1131 |
+
return self
|
quantization.py
ADDED
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|