ian commited on
Commit
5955d06
1 Parent(s): d59f08a

add model2

Browse files
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 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - zh
4
+ - en
5
+ tags:
6
+ - glm
7
+ - chatglm
8
+ - thudm
9
+ ---
10
+ # ChatGLM2-6B
11
+ <p align="center">
12
+ 💻 <a href="https://github.com/THUDM/ChatGLM2-6B" 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-1y7pqoloy-9b1g6T6JjA8J0KxvUjbwJw" 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**2**-6B 是开源中英双语对话模型 [ChatGLM-6B](https://github.com/THUDM/ChatGLM-6B) 的第二代版本,在保留了初代模型对话流畅、部署门槛较低等众多优秀特性的基础之上,ChatGLM**2**-6B 引入了如下新特性:
21
+
22
+ 1. **更强大的性能**:基于 ChatGLM 初代模型的开发经验,我们全面升级了 ChatGLM2-6B 的基座模型。ChatGLM2-6B 使用了 [GLM](https://github.com/THUDM/GLM) 的混合目标函数,经过了 1.4T 中英标识符的预训练与人类偏好对齐训练,[评测结果](#评测结果)显示,相比于初代模型,ChatGLM2-6B 在 MMLU(+23%)、CEval(+33%)、GSM8K(+571%) 、BBH(+60%)等数据集上的性能取得了大幅度的提升,在同尺寸开源模型中具有较强的竞争力。
23
+ 2. **更长的上下文**:基于 [FlashAttention](https://github.com/HazyResearch/flash-attention) 技术,我们将基座模型的上下文长度(Context Length)由 ChatGLM-6B 的 2K 扩展到了 32K,并在对话阶段使用 8K 的上下文长度训练,允许更多轮次的对话。但当前版本的 ChatGLM2-6B 对单轮超长文档的理解能力有限,我们会在后续迭代升级中着重进行优化。
24
+ 3. **更高效的推理**:基于 [Multi-Query Attention](http://arxiv.org/abs/1911.02150) 技术,ChatGLM2-6B 有更高效的推理速度和更低的显存占用:在官方的模型实现下,推理速度相比初代提升了 42%,INT4 量化下,6G 显存支持的对话长度由 1K 提升到了 8K。
25
+
26
+ ChatGLM**2**-6B is the second-generation version of the open-source bilingual (Chinese-English) chat model [ChatGLM-6B](https://github.com/THUDM/ChatGLM-6B). It retains the smooth conversation flow and low deployment threshold of the first-generation model, while introducing the following new features:
27
+
28
+ 1. **Stronger Performance**: Based on the development experience of the first-generation ChatGLM model, we have fully upgraded the base model of ChatGLM2-6B. ChatGLM2-6B uses the hybrid objective function of [GLM](https://github.com/THUDM/GLM), and has undergone pre-training with 1.4T bilingual tokens and human preference alignment training. The [evaluation results](README.md#evaluation-results) show that, compared to the first-generation model, ChatGLM2-6B has achieved substantial improvements in performance on datasets like MMLU (+23%), CEval (+33%), GSM8K (+571%), BBH (+60%), showing strong competitiveness among models of the same size.
29
+ 2. **Longer Context**: Based on [FlashAttention](https://github.com/HazyResearch/flash-attention) technique, we have extended the context length of the base model from 2K in ChatGLM-6B to 32K, and trained with a context length of 8K during the dialogue alignment, allowing for more rounds of dialogue. However, the current version of ChatGLM2-6B has limited understanding of single-round ultra-long documents, which we will focus on optimizing in future iterations.
30
+ 3. **More Efficient Inference**: Based on [Multi-Query Attention](http://arxiv.org/abs/1911.02150) technique, ChatGLM2-6B has more efficient inference speed and lower GPU memory usage: under the official implementation, the inference speed has increased by 42% compared to the first generation; under INT4 quantization, the dialogue length supported by 6G GPU memory has increased from 1K to 8K.
31
+
32
+ ## 软件依赖
33
+
34
+ ```shell
35
+ pip install protobuf transformers==4.30.2 cpm_kernels torch>=2.0 gradio mdtex2html sentencepiece accelerate
36
+ ```
37
+
38
+ ## 代码调用
39
+
40
+ 可以通过如下代码调用 ChatGLM-6B 模型来生成对话:
41
+
42
+ ```ipython
43
+ >>> from transformers import AutoTokenizer, AutoModel
44
+ >>> tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm2-6b-int4", trust_remote_code=True)
45
+ >>> model = AutoModel.from_pretrained("THUDM/chatglm2-6b-int4", trust_remote_code=True).half().cuda()
46
+ >>> model = model.eval()
47
+ >>> response, history = model.chat(tokenizer, "你好", history=[])
48
+ >>> print(response)
49
+ 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。
50
+ >>> response, history = model.chat(tokenizer, "晚上睡不着应该怎么办", history=history)
51
+ >>> print(response)
52
+ 晚上睡不着可能会让你���到焦虑或不舒服,但以下是一些可以帮助你入睡的方法:
53
+
54
+ 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。
55
+ 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。
56
+ 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。
57
+ 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。
58
+ 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。
59
+ 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。
60
+
61
+ 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。
62
+ ```
63
+
64
+ 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/ChatGLM2-6B)。
65
+
66
+ For more instructions, including how to run CLI and web demos, and model quantization, please refer to our [Github Repo](https://github.com/THUDM/ChatGLM2-6B).
67
+
68
+ ## Change Log
69
+ * v1.0
70
+
71
+ ## 协议
72
+
73
+ 本仓库的代码依照 [Apache-2.0](LICENSE) 协议开源,ChatGLM2-6B 模型的权重的使用则需要遵循 [Model License](MODEL_LICENSE)。
74
+
75
+ ## 引用
76
+
77
+ 如果你觉得我们的工作有帮助的话,请考虑引用下列论文,ChatGLM2-6B 的论文会在近期公布,尽情期待~
78
+
79
+ ```
80
+ @article{zeng2022glm,
81
+ title={Glm-130b: An open bilingual pre-trained model},
82
+ author={Zeng, Aohan and Liu, Xiao and Du, Zhengxiao and Wang, Zihan and Lai, Hanyu and Ding, Ming and Yang, Zhuoyi and Xu, Yifan and Zheng, Wendi and Xia, Xiao and others},
83
+ journal={arXiv preprint arXiv:2210.02414},
84
+ year={2022}
85
+ }
86
+ ```
87
+ ```
88
+ @inproceedings{du2022glm,
89
+ title={GLM: General Language Model Pretraining with Autoregressive Blank Infilling},
90
+ author={Du, Zhengxiao and Qian, Yujie and Liu, Xiao and Ding, Ming and Qiu, Jiezhong and Yang, Zhilin and Tang, Jie},
91
+ booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
92
+ pages={320--335},
93
+ year={2022}
94
+ }
95
+ ```
config.json DELETED
@@ -1 +0,0 @@
1
- ../../blobs/dbb1da64cd4778e837b6f5a184293fc1d2b69097
 
 
config.json ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "THUDM/chatglm2-6b",
3
+ "model_type": "chatglm",
4
+ "architectures": [
5
+ "ChatGLMModel"
6
+ ],
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_chatglm.ChatGLMConfig",
9
+ "AutoModel": "modeling_chatglm.ChatGLMForConditionalGeneration",
10
+ "AutoModelForSeq2SeqLM": "modeling_chatglm.ChatGLMForConditionalGeneration"
11
+ },
12
+ "add_bias_linear": false,
13
+ "add_qkv_bias": true,
14
+ "apply_query_key_layer_scaling": true,
15
+ "apply_residual_connection_post_layernorm": false,
16
+ "attention_dropout": 0.0,
17
+ "attention_softmax_in_fp32": true,
18
+ "bias_dropout_fusion": true,
19
+ "ffn_hidden_size": 13696,
20
+ "fp32_residual_connection": false,
21
+ "hidden_dropout": 0.0,
22
+ "hidden_size": 4096,
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
+ "quantization_bit": 4,
33
+ "rmsnorm": true,
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
+ "pad_token_id": 0
41
+ }
configuration_chatglm.py DELETED
@@ -1 +0,0 @@
1
- ../../blobs/3730b4e34d8688587e88cbb75384542244a40662
 
 
configuration_chatglm.py ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import PretrainedConfig
2
+
3
+
4
+ class ChatGLMConfig(PretrainedConfig):
5
+ model_type = "chatglm"
6
+ def __init__(
7
+ self,
8
+ num_layers=28,
9
+ padded_vocab_size=65024,
10
+ hidden_size=4096,
11
+ ffn_hidden_size=13696,
12
+ kv_channels=128,
13
+ num_attention_heads=32,
14
+ seq_length=2048,
15
+ hidden_dropout=0.0,
16
+ attention_dropout=0.0,
17
+ layernorm_epsilon=1e-5,
18
+ rmsnorm=True,
19
+ apply_residual_connection_post_layernorm=False,
20
+ post_layer_norm=True,
21
+ add_bias_linear=False,
22
+ add_qkv_bias=False,
23
+ bias_dropout_fusion=True,
24
+ multi_query_attention=False,
25
+ multi_query_group_num=1,
26
+ apply_query_key_layer_scaling=True,
27
+ attention_softmax_in_fp32=True,
28
+ fp32_residual_connection=False,
29
+ quantization_bit=0,
30
+ pre_seq_len=None,
31
+ prefix_projection=False,
32
+ **kwargs
33
+ ):
34
+ self.num_layers = num_layers
35
+ self.vocab_size = padded_vocab_size
36
+ self.padded_vocab_size = padded_vocab_size
37
+ self.hidden_size = hidden_size
38
+ self.ffn_hidden_size = ffn_hidden_size
39
+ self.kv_channels = kv_channels
40
+ self.num_attention_heads = num_attention_heads
41
+ self.seq_length = seq_length
42
+ self.hidden_dropout = hidden_dropout
43
+ self.attention_dropout = attention_dropout
44
+ self.layernorm_epsilon = layernorm_epsilon
45
+ self.rmsnorm = rmsnorm
46
+ self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
47
+ self.post_layer_norm = post_layer_norm
48
+ self.add_bias_linear = add_bias_linear
49
+ self.add_qkv_bias = add_qkv_bias
50
+ self.bias_dropout_fusion = bias_dropout_fusion
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
+ self.pre_seq_len = pre_seq_len
58
+ self.prefix_projection = prefix_projection
59
+ super().__init__(**kwargs)
modeling_chatglm.py DELETED
@@ -1 +0,0 @@
1
- ../../blobs/230f4c3b7c6cdc56d6ecec1043eeccfc7b0a7844
 
 
modeling_chatglm.py ADDED
@@ -0,0 +1,1193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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/ChatGLM2-6B"
39
+ _CONFIG_FOR_DOC = "ChatGLM6BConfig"
40
+
41
+ CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [
42
+ "THUDM/chatglm2-6b",
43
+ # See all ChatGLM 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
+ class PrefixEncoder(torch.nn.Module):
60
+ """
61
+ The torch.nn model to encode the prefix
62
+ Input shape: (batch-size, prefix-length)
63
+ Output shape: (batch-size, prefix-length, 2*layers*hidden)
64
+ """
65
+
66
+ def __init__(self, config: ChatGLMConfig):
67
+ super().__init__()
68
+ self.prefix_projection = config.prefix_projection
69
+ if self.prefix_projection:
70
+ # Use a two-layer MLP to encode the prefix
71
+ kv_size = config.num_layers * config.kv_channels * config.multi_query_group_num * 2
72
+ self.embedding = torch.nn.Embedding(config.pre_seq_len, kv_size)
73
+ self.trans = torch.nn.Sequential(
74
+ torch.nn.Linear(kv_size, config.hidden_size),
75
+ torch.nn.Tanh(),
76
+ torch.nn.Linear(config.hidden_size, kv_size)
77
+ )
78
+ else:
79
+ self.embedding = torch.nn.Embedding(config.pre_seq_len,
80
+ config.num_layers * config.kv_channels * config.multi_query_group_num * 2)
81
+
82
+ def forward(self, prefix: torch.Tensor):
83
+ if self.prefix_projection:
84
+ prefix_tokens = self.embedding(prefix)
85
+ past_key_values = self.trans(prefix_tokens)
86
+ else:
87
+ past_key_values = self.embedding(prefix)
88
+ return past_key_values
89
+
90
+
91
+ def split_tensor_along_last_dim(
92
+ tensor: torch.Tensor,
93
+ num_partitions: int,
94
+ contiguous_split_chunks: bool = False,
95
+ ) -> List[torch.Tensor]:
96
+ """Split a tensor along its last dimension.
97
+
98
+ Arguments:
99
+ tensor: input tensor.
100
+ num_partitions: number of partitions to split the tensor
101
+ contiguous_split_chunks: If True, make each chunk contiguous
102
+ in memory.
103
+
104
+ Returns:
105
+ A list of Tensors
106
+ """
107
+ # Get the size and dimension.
108
+ last_dim = tensor.dim() - 1
109
+ last_dim_size = tensor.size()[last_dim] // num_partitions
110
+ # Split.
111
+ tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
112
+ # Note: torch.split does not create contiguous tensors by default.
113
+ if contiguous_split_chunks:
114
+ return tuple(chunk.contiguous() for chunk in tensor_list)
115
+
116
+ return tensor_list
117
+
118
+
119
+ class RotaryEmbedding(nn.Module):
120
+ def __init__(self, dim, original_impl=False, device=None, dtype=None):
121
+ super().__init__()
122
+ inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, device=device).to(dtype=dtype) / dim))
123
+ self.register_buffer("inv_freq", inv_freq)
124
+ self.dim = dim
125
+ self.original_impl = original_impl
126
+
127
+ def forward_impl(
128
+ self, seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000
129
+ ):
130
+ """Enhanced Transformer with Rotary Position Embedding.
131
+
132
+ Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
133
+ transformers/rope/__init__.py. MIT License:
134
+ https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license.
135
+ """
136
+ # $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$
137
+ theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, dtype=dtype, device=device) / n_elem))
138
+
139
+ # Create position indexes `[0, 1, ..., seq_len - 1]`
140
+ seq_idx = torch.arange(seq_len, dtype=dtype, device=device)
141
+
142
+ # Calculate the product of position index and $\theta_i$
143
+ idx_theta = torch.outer(seq_idx, theta).float()
144
+
145
+ cache = torch.stack([torch.cos(idx_theta), torch.sin(idx_theta)], dim=-1)
146
+
147
+ # this is to mimic the behaviour of complex32, else we will get different results
148
+ if dtype in (torch.float16, torch.bfloat16, torch.int8):
149
+ cache = cache.bfloat16() if dtype == torch.bfloat16 else cache.half()
150
+ return cache
151
+
152
+ def forward(self, max_seq_len, offset=0):
153
+ return self.forward_impl(
154
+ max_seq_len, self.dim, dtype=self.inv_freq.dtype, device=self.inv_freq.device
155
+ )
156
+
157
+
158
+ @torch.jit.script
159
+ def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
160
+ # x: [sq, b, np, hn]
161
+ sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
162
+ rot_dim = rope_cache.shape[-2] * 2
163
+ x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
164
+ # truncate to support variable sizes
165
+ rope_cache = rope_cache[:sq]
166
+ xshaped = x.reshape(sq, -1, np, rot_dim // 2, 2)
167
+ rope_cache = rope_cache.view(sq, -1, 1, xshaped.size(3), 2)
168
+ x_out2 = torch.stack(
169
+ [
170
+ xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
171
+ xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
172
+ ],
173
+ -1,
174
+ )
175
+ x_out2 = x_out2.flatten(3)
176
+ return torch.cat((x_out2, x_pass), dim=-1)
177
+
178
+
179
+ class RMSNorm(torch.nn.Module):
180
+ def __init__(self, normalized_shape, eps=1e-5, device=None, dtype=None, **kwargs):
181
+ super().__init__()
182
+ self.weight = torch.nn.Parameter(torch.empty(normalized_shape, device=device, dtype=dtype))
183
+ self.eps = eps
184
+
185
+ def forward(self, hidden_states: torch.Tensor):
186
+ input_dtype = hidden_states.dtype
187
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
188
+ hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
189
+
190
+ return (self.weight * hidden_states).to(input_dtype)
191
+
192
+
193
+ class CoreAttention(torch.nn.Module):
194
+ def __init__(self, config: ChatGLMConfig, layer_number):
195
+ super(CoreAttention, self).__init__()
196
+
197
+ self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling
198
+ self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
199
+ if self.apply_query_key_layer_scaling:
200
+ self.attention_softmax_in_fp32 = True
201
+ self.layer_number = max(1, layer_number)
202
+
203
+ projection_size = config.kv_channels * config.num_attention_heads
204
+
205
+ # Per attention head and per partition values.
206
+ self.hidden_size_per_partition = projection_size
207
+ self.hidden_size_per_attention_head = projection_size // config.num_attention_heads
208
+ self.num_attention_heads_per_partition = config.num_attention_heads
209
+
210
+ coeff = None
211
+ self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
212
+ if self.apply_query_key_layer_scaling:
213
+ coeff = self.layer_number
214
+ self.norm_factor *= coeff
215
+ self.coeff = coeff
216
+
217
+ self.attention_dropout = torch.nn.Dropout(config.attention_dropout)
218
+
219
+ def forward(self, query_layer, key_layer, value_layer, attention_mask):
220
+ pytorch_major_version = int(torch.__version__.split('.')[0])
221
+ if pytorch_major_version >= 2:
222
+ query_layer, key_layer, value_layer = [k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer]]
223
+ if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
224
+ context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
225
+ is_causal=True)
226
+ else:
227
+ if attention_mask is not None:
228
+ attention_mask = ~attention_mask
229
+ context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
230
+ attention_mask)
231
+ context_layer = context_layer.permute(2, 0, 1, 3)
232
+ new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
233
+ context_layer = context_layer.reshape(*new_context_layer_shape)
234
+ else:
235
+ # Raw attention scores
236
+
237
+ # [b, np, sq, sk]
238
+ output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
239
+
240
+ # [sq, b, np, hn] -> [sq, b * np, hn]
241
+ query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
242
+ # [sk, b, np, hn] -> [sk, b * np, hn]
243
+ key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
244
+
245
+ # preallocting input tensor: [b * np, sq, sk]
246
+ matmul_input_buffer = torch.empty(
247
+ output_size[0] * output_size[1], output_size[2], output_size[3], dtype=query_layer.dtype,
248
+ device=query_layer.device
249
+ )
250
+
251
+ # Raw attention scores. [b * np, sq, sk]
252
+ matmul_result = torch.baddbmm(
253
+ matmul_input_buffer,
254
+ query_layer.transpose(0, 1), # [b * np, sq, hn]
255
+ key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk]
256
+ beta=0.0,
257
+ alpha=(1.0 / self.norm_factor),
258
+ )
259
+
260
+ # change view to [b, np, sq, sk]
261
+ attention_scores = matmul_result.view(*output_size)
262
+
263
+ # ===========================
264
+ # Attention probs and dropout
265
+ # ===========================
266
+
267
+ # attention scores and attention mask [b, np, sq, sk]
268
+ if self.attention_softmax_in_fp32:
269
+ attention_scores = attention_scores.float()
270
+ if self.coeff is not None:
271
+ attention_scores = attention_scores * self.coeff
272
+ if attention_mask is None and attention_scores.shape[2] == attention_scores.shape[3]:
273
+ attention_mask = torch.ones(output_size[0], 1, output_size[2], output_size[3],
274
+ device=attention_scores.device, dtype=torch.bool)
275
+ attention_mask.tril_()
276
+ attention_mask = ~attention_mask
277
+ if attention_mask is not None:
278
+ attention_scores = attention_scores.masked_fill(attention_mask, float("-inf"))
279
+ attention_probs = F.softmax(attention_scores, dim=-1)
280
+ attention_probs = attention_probs.type_as(value_layer)
281
+
282
+ # This is actually dropping out entire tokens to attend to, which might
283
+ # seem a bit unusual, but is taken from the original Transformer paper.
284
+ attention_probs = self.attention_dropout(attention_probs)
285
+ # =========================
286
+ # Context layer. [sq, b, hp]
287
+ # =========================
288
+
289
+ # value_layer -> context layer.
290
+ # [sk, b, np, hn] --> [b, np, sq, hn]
291
+
292
+ # context layer shape: [b, np, sq, hn]
293
+ output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))
294
+ # change view [sk, b * np, hn]
295
+ value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)
296
+ # change view [b * np, sq, sk]
297
+ attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
298
+ # matmul: [b * np, sq, hn]
299
+ context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
300
+ # change view [b, np, sq, hn]
301
+ context_layer = context_layer.view(*output_size)
302
+ # [b, np, sq, hn] --> [sq, b, np, hn]
303
+ context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
304
+ # [sq, b, np, hn] --> [sq, b, hp]
305
+ new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
306
+ context_layer = context_layer.view(*new_context_layer_shape)
307
+
308
+ return context_layer
309
+
310
+
311
+ class SelfAttention(torch.nn.Module):
312
+ """Parallel self-attention layer abstract class.
313
+
314
+ Self-attention layer takes input with size [s, b, h]
315
+ and returns output of the same size.
316
+ """
317
+
318
+ def __init__(self, config: ChatGLMConfig, layer_number, device=None):
319
+ super(SelfAttention, self).__init__()
320
+ self.layer_number = max(1, layer_number)
321
+
322
+ self.projection_size = config.kv_channels * config.num_attention_heads
323
+
324
+ # Per attention head and per partition values.
325
+ self.hidden_size_per_attention_head = self.projection_size // config.num_attention_heads
326
+ self.num_attention_heads_per_partition = config.num_attention_heads
327
+
328
+ self.multi_query_attention = config.multi_query_attention
329
+ self.qkv_hidden_size = 3 * self.projection_size
330
+ if self.multi_query_attention:
331
+ self.num_multi_query_groups_per_partition = config.multi_query_group_num
332
+ self.qkv_hidden_size = (
333
+ self.projection_size + 2 * self.hidden_size_per_attention_head * config.multi_query_group_num
334
+ )
335
+ self.query_key_value = nn.Linear(config.hidden_size, self.qkv_hidden_size,
336
+ bias=config.add_bias_linear or config.add_qkv_bias,
337
+ device=device, **_config_to_kwargs(config)
338
+ )
339
+
340
+ self.core_attention = CoreAttention(config, self.layer_number)
341
+
342
+ # Output.
343
+ self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear,
344
+ device=device, **_config_to_kwargs(config)
345
+ )
346
+
347
+ def _allocate_memory(self, inference_max_sequence_len, batch_size, device=None, dtype=None):
348
+ if self.multi_query_attention:
349
+ num_attention_heads = self.num_multi_query_groups_per_partition
350
+ else:
351
+ num_attention_heads = self.num_attention_heads_per_partition
352
+ return torch.empty(
353
+ inference_max_sequence_len,
354
+ batch_size,
355
+ num_attention_heads,
356
+ self.hidden_size_per_attention_head,
357
+ dtype=dtype,
358
+ device=device,
359
+ )
360
+
361
+ def forward(
362
+ self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
363
+ ):
364
+ # hidden_states: [sq, b, h]
365
+
366
+ # =================================================
367
+ # Pre-allocate memory for key-values for inference.
368
+ # =================================================
369
+ # =====================
370
+ # Query, Key, and Value
371
+ # =====================
372
+
373
+ # Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
374
+ mixed_x_layer = self.query_key_value(hidden_states)
375
+
376
+ if self.multi_query_attention:
377
+ (query_layer, key_layer, value_layer) = mixed_x_layer.split(
378
+ [
379
+ self.num_attention_heads_per_partition * self.hidden_size_per_attention_head,
380
+ self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
381
+ self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
382
+ ],
383
+ dim=-1,
384
+ )
385
+ query_layer = query_layer.view(
386
+ query_layer.size()[:-1] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
387
+ )
388
+ key_layer = key_layer.view(
389
+ key_layer.size()[:-1] + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
390
+ )
391
+ value_layer = value_layer.view(
392
+ value_layer.size()[:-1]
393
+ + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
394
+ )
395
+ else:
396
+ new_tensor_shape = mixed_x_layer.size()[:-1] + \
397
+ (self.num_attention_heads_per_partition,
398
+ 3 * self.hidden_size_per_attention_head)
399
+ mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
400
+
401
+ # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
402
+ (query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
403
+
404
+ # apply relative positional encoding (rotary embedding)
405
+ if rotary_pos_emb is not None:
406
+ query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb)
407
+ key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb)
408
+
409
+ # adjust key and value for inference
410
+ if kv_cache is not None:
411
+ cache_k, cache_v = kv_cache
412
+ key_layer = torch.cat((cache_k, key_layer), dim=0)
413
+ value_layer = torch.cat((cache_v, value_layer), dim=0)
414
+ if use_cache:
415
+ kv_cache = (key_layer, value_layer)
416
+ else:
417
+ kv_cache = None
418
+
419
+ if self.multi_query_attention:
420
+ key_layer = key_layer.unsqueeze(-2)
421
+ key_layer = key_layer.expand(
422
+ -1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1
423
+ )
424
+ key_layer = key_layer.contiguous().view(
425
+ key_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
426
+ )
427
+ value_layer = value_layer.unsqueeze(-2)
428
+ value_layer = value_layer.expand(
429
+ -1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1
430
+ )
431
+ value_layer = value_layer.contiguous().view(
432
+ value_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
433
+ )
434
+
435
+ # ==================================
436
+ # core attention computation
437
+ # ==================================
438
+
439
+ context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask)
440
+
441
+ # =================
442
+ # Output. [sq, b, h]
443
+ # =================
444
+
445
+ output = self.dense(context_layer)
446
+
447
+ return output, kv_cache
448
+
449
+
450
+ def _config_to_kwargs(args):
451
+ common_kwargs = {
452
+ "dtype": args.torch_dtype,
453
+ }
454
+ return common_kwargs
455
+
456
+
457
+ class MLP(torch.nn.Module):
458
+ """MLP.
459
+
460
+ MLP will take the input with h hidden state, project it to 4*h
461
+ hidden dimension, perform nonlinear transformation, and project the
462
+ state back into h hidden dimension.
463
+ """
464
+
465
+ def __init__(self, config: ChatGLMConfig, device=None):
466
+ super(MLP, self).__init__()
467
+
468
+ self.add_bias = config.add_bias_linear
469
+
470
+ # Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
471
+ self.dense_h_to_4h = nn.Linear(
472
+ config.hidden_size,
473
+ config.ffn_hidden_size * 2,
474
+ bias=self.add_bias,
475
+ device=device,
476
+ **_config_to_kwargs(config)
477
+ )
478
+
479
+ def swiglu(x):
480
+ x = torch.chunk(x, 2, dim=-1)
481
+ return F.silu(x[0]) * x[1]
482
+
483
+ self.activation_func = swiglu
484
+
485
+ # Project back to h.
486
+ self.dense_4h_to_h = nn.Linear(
487
+ config.ffn_hidden_size,
488
+ config.hidden_size,
489
+ bias=self.add_bias,
490
+ device=device,
491
+ **_config_to_kwargs(config)
492
+ )
493
+
494
+ def forward(self, hidden_states):
495
+ # [s, b, 4hp]
496
+ intermediate_parallel = self.dense_h_to_4h(hidden_states)
497
+ intermediate_parallel = self.activation_func(intermediate_parallel)
498
+ # [s, b, h]
499
+ output = self.dense_4h_to_h(intermediate_parallel)
500
+ return output
501
+
502
+
503
+ class GLMBlock(torch.nn.Module):
504
+ """A single transformer layer.
505
+
506
+ Transformer layer takes input with size [s, b, h] and returns an
507
+ output of the same size.
508
+ """
509
+
510
+ def __init__(self, config: ChatGLMConfig, layer_number, device=None):
511
+ super(GLMBlock, self).__init__()
512
+ self.layer_number = layer_number
513
+
514
+ self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
515
+
516
+ self.fp32_residual_connection = config.fp32_residual_connection
517
+
518
+ LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
519
+ # Layernorm on the input data.
520
+ self.input_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
521
+ dtype=config.torch_dtype)
522
+
523
+ # Self attention.
524
+ self.self_attention = SelfAttention(config, layer_number, device=device)
525
+ self.hidden_dropout = config.hidden_dropout
526
+
527
+ # Layernorm on the attention output
528
+ self.post_attention_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
529
+ dtype=config.torch_dtype)
530
+
531
+ # MLP
532
+ self.mlp = MLP(config, device=device)
533
+
534
+ def forward(
535
+ self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True,
536
+ ):
537
+ # hidden_states: [s, b, h]
538
+
539
+ # Layer norm at the beginning of the transformer layer.
540
+ layernorm_output = self.input_layernorm(hidden_states)
541
+ # Self attention.
542
+ attention_output, kv_cache = self.self_attention(
543
+ layernorm_output,
544
+ attention_mask,
545
+ rotary_pos_emb,
546
+ kv_cache=kv_cache,
547
+ use_cache=use_cache
548
+ )
549
+
550
+ # Residual connection.
551
+ if self.apply_residual_connection_post_layernorm:
552
+ residual = layernorm_output
553
+ else:
554
+ residual = hidden_states
555
+
556
+ layernorm_input = torch.nn.functional.dropout(attention_output, p=self.hidden_dropout, training=self.training)
557
+ layernorm_input = residual + layernorm_input
558
+
559
+ # Layer norm post the self attention.
560
+ layernorm_output = self.post_attention_layernorm(layernorm_input)
561
+
562
+ # MLP.
563
+ mlp_output = self.mlp(layernorm_output)
564
+
565
+ # Second residual connection.
566
+ if self.apply_residual_connection_post_layernorm:
567
+ residual = layernorm_output
568
+ else:
569
+ residual = layernorm_input
570
+
571
+ output = torch.nn.functional.dropout(mlp_output, p=self.hidden_dropout, training=self.training)
572
+ output = residual + output
573
+
574
+ return output, kv_cache
575
+
576
+
577
+ class GLMTransformer(torch.nn.Module):
578
+ """Transformer class."""
579
+
580
+ def __init__(self, config: ChatGLMConfig, device=None):
581
+ super(GLMTransformer, self).__init__()
582
+
583
+ self.fp32_residual_connection = config.fp32_residual_connection
584
+ self.post_layer_norm = config.post_layer_norm
585
+
586
+ # Number of layers.
587
+ self.num_layers = config.num_layers
588
+
589
+ # Transformer layers.
590
+ def build_layer(layer_number):
591
+ return GLMBlock(config, layer_number, device=device)
592
+
593
+ self.layers = torch.nn.ModuleList([build_layer(i + 1) for i in range(self.num_layers)])
594
+
595
+ if self.post_layer_norm:
596
+ LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
597
+ # Final layer norm before output.
598
+ self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
599
+ dtype=config.torch_dtype)
600
+
601
+ self.gradient_checkpointing = False
602
+
603
+ def _get_layer(self, layer_number):
604
+ return self.layers[layer_number]
605
+
606
+ def forward(
607
+ self, hidden_states, attention_mask, rotary_pos_emb, kv_caches=None,
608
+ use_cache: Optional[bool] = True,
609
+ output_hidden_states: Optional[bool] = False,
610
+ ):
611
+ if not kv_caches:
612
+ kv_caches = [None for _ in range(self.num_layers)]
613
+ presents = () if use_cache else None
614
+ if self.gradient_checkpointing and self.training:
615
+ if use_cache:
616
+ logger.warning_once(
617
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
618
+ )
619
+ use_cache = False
620
+
621
+ all_self_attentions = None
622
+ all_hidden_states = () if output_hidden_states else None
623
+ for index in range(self.num_layers):
624
+ if output_hidden_states:
625
+ all_hidden_states = all_hidden_states + (hidden_states,)
626
+
627
+ layer = self._get_layer(index)
628
+ if self.gradient_checkpointing and self.training:
629
+ layer_ret = torch.utils.checkpoint.checkpoint(
630
+ layer,
631
+ hidden_states,
632
+ attention_mask,
633
+ rotary_pos_emb,
634
+ kv_caches[index],
635
+ use_cache
636
+ )
637
+ else:
638
+ layer_ret = layer(
639
+ hidden_states,
640
+ attention_mask,
641
+ rotary_pos_emb,
642
+ kv_cache=kv_caches[index],
643
+ use_cache=use_cache
644
+ )
645
+ hidden_states, kv_cache = layer_ret
646
+ if use_cache:
647
+ presents = presents + (kv_cache,)
648
+
649
+ if output_hidden_states:
650
+ all_hidden_states = all_hidden_states + (hidden_states,)
651
+
652
+ # Final layer norm.
653
+ if self.post_layer_norm:
654
+ hidden_states = self.final_layernorm(hidden_states)
655
+
656
+ return hidden_states, presents, all_hidden_states, all_self_attentions
657
+
658
+
659
+ class ChatGLMPreTrainedModel(PreTrainedModel):
660
+ """
661
+ An abstract class to handle weights initialization and
662
+ a simple interface for downloading and loading pretrained models.
663
+ """
664
+
665
+ is_parallelizable = False
666
+ supports_gradient_checkpointing = True
667
+ config_class = ChatGLMConfig
668
+ base_model_prefix = "transformer"
669
+ _no_split_modules = ["GLMBlock"]
670
+
671
+ def _init_weights(self, module: nn.Module):
672
+ """Initialize the weights."""
673
+ return
674
+
675
+ def get_masks(self, input_ids, past_key_values, padding_mask=None):
676
+ batch_size, seq_length = input_ids.shape
677
+ full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device)
678
+ full_attention_mask.tril_()
679
+ past_length = 0
680
+ if past_key_values:
681
+ past_length = past_key_values[0][0].shape[0]
682
+ if past_length:
683
+ full_attention_mask = torch.cat((torch.ones(batch_size, seq_length, past_length,
684
+ device=input_ids.device), full_attention_mask), dim=-1)
685
+ if padding_mask is not None:
686
+ full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1)
687
+ if not past_length and padding_mask is not None:
688
+ full_attention_mask -= padding_mask.unsqueeze(-1) - 1
689
+ full_attention_mask = (full_attention_mask < 0.5).bool()
690
+ full_attention_mask.unsqueeze_(1)
691
+ return full_attention_mask
692
+
693
+ def get_position_ids(self, input_ids, device):
694
+ batch_size, seq_length = input_ids.shape
695
+ position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
696
+ return position_ids
697
+
698
+ def _set_gradient_checkpointing(self, module, value=False):
699
+ if isinstance(module, GLMTransformer):
700
+ module.gradient_checkpointing = value
701
+
702
+
703
+ class Embedding(torch.nn.Module):
704
+ """Language model embeddings."""
705
+
706
+ def __init__(self, config: ChatGLMConfig, device=None):
707
+ super(Embedding, self).__init__()
708
+
709
+ self.hidden_size = config.hidden_size
710
+ # Word embeddings (parallel).
711
+ self.word_embeddings = nn.Embedding(
712
+ config.padded_vocab_size,
713
+ self.hidden_size,
714
+ dtype=config.torch_dtype,
715
+ device=device
716
+ )
717
+ self.fp32_residual_connection = config.fp32_residual_connection
718
+
719
+ def forward(self, input_ids):
720
+ # Embeddings.
721
+ words_embeddings = self.word_embeddings(input_ids)
722
+ embeddings = words_embeddings
723
+ # Data format change to avoid explicit tranposes : [b s h] --> [s b h].
724
+ embeddings = embeddings.transpose(0, 1).contiguous()
725
+ # If the input flag for fp32 residual connection is set, convert for float.
726
+ if self.fp32_residual_connection:
727
+ embeddings = embeddings.float()
728
+ return embeddings
729
+
730
+
731
+ class ChatGLMModel(ChatGLMPreTrainedModel):
732
+ def __init__(self, config: ChatGLMConfig, device=None, empty_init=True):
733
+ super().__init__(config)
734
+ if empty_init:
735
+ init_method = skip_init
736
+ else:
737
+ init_method = default_init
738
+ init_kwargs = {}
739
+ if device is not None:
740
+ init_kwargs["device"] = device
741
+ self.embedding = init_method(Embedding, config, **init_kwargs)
742
+ self.num_layers = config.num_layers
743
+ self.multi_query_group_num = config.multi_query_group_num
744
+ self.kv_channels = config.kv_channels
745
+
746
+ # Rotary positional embeddings
747
+ self.seq_length = config.seq_length
748
+ rotary_dim = (
749
+ config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
750
+ )
751
+
752
+ self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, original_impl=config.original_rope, device=device,
753
+ dtype=config.torch_dtype)
754
+ self.encoder = init_method(GLMTransformer, config, **init_kwargs)
755
+ self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
756
+ dtype=config.torch_dtype, **init_kwargs)
757
+ self.pre_seq_len = config.pre_seq_len
758
+ self.prefix_projection = config.prefix_projection
759
+ if self.pre_seq_len is not None:
760
+ for param in self.parameters():
761
+ param.requires_grad = False
762
+ self.prefix_tokens = torch.arange(self.pre_seq_len).long()
763
+ self.prefix_encoder = PrefixEncoder(config)
764
+ self.dropout = torch.nn.Dropout(0.1)
765
+
766
+ def get_input_embeddings(self):
767
+ return self.embedding.word_embeddings
768
+
769
+ def get_prompt(self, batch_size, device, dtype=torch.half):
770
+ prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device)
771
+ past_key_values = self.prefix_encoder(prefix_tokens).type(dtype)
772
+ past_key_values = past_key_values.view(
773
+ batch_size,
774
+ self.pre_seq_len,
775
+ self.num_layers * 2,
776
+ self.multi_query_group_num,
777
+ self.kv_channels
778
+ )
779
+ # seq_len, b, nh, hidden_size
780
+ past_key_values = self.dropout(past_key_values)
781
+ past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2)
782
+ return past_key_values
783
+
784
+ def forward(
785
+ self,
786
+ input_ids,
787
+ position_ids: Optional[torch.Tensor] = None,
788
+ attention_mask: Optional[torch.BoolTensor] = None,
789
+ full_attention_mask: Optional[torch.BoolTensor] = None,
790
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
791
+ inputs_embeds: Optional[torch.Tensor] = None,
792
+ use_cache: Optional[bool] = None,
793
+ output_hidden_states: Optional[bool] = None,
794
+ return_dict: Optional[bool] = None,
795
+ ):
796
+ output_hidden_states = (
797
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
798
+ )
799
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
800
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
801
+
802
+ batch_size, seq_length = input_ids.shape
803
+
804
+ if inputs_embeds is None:
805
+ inputs_embeds = self.embedding(input_ids)
806
+
807
+ if self.pre_seq_len is not None:
808
+ if past_key_values is None:
809
+ past_key_values = self.get_prompt(batch_size=batch_size, device=input_ids.device,
810
+ dtype=inputs_embeds.dtype)
811
+ if attention_mask is not None:
812
+ attention_mask = torch.cat([attention_mask.new_ones((batch_size, self.pre_seq_len)),
813
+ attention_mask], dim=-1)
814
+
815
+ if full_attention_mask is None:
816
+ if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1):
817
+ full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)
818
+
819
+ # Rotary positional embeddings
820
+ rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
821
+ if position_ids is not None:
822
+ rotary_pos_emb = rotary_pos_emb[position_ids]
823
+ else:
824
+ rotary_pos_emb = rotary_pos_emb[None, :seq_length]
825
+ rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
826
+
827
+ # Run encoder.
828
+ hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
829
+ inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb,
830
+ kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
831
+ )
832
+
833
+ if not return_dict:
834
+ return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
835
+
836
+ return BaseModelOutputWithPast(
837
+ last_hidden_state=hidden_states,
838
+ past_key_values=presents,
839
+ hidden_states=all_hidden_states,
840
+ attentions=all_self_attentions,
841
+ )
842
+
843
+ def quantize(self, weight_bit_width: int):
844
+ from .quantization import quantize
845
+ quantize(self.encoder, weight_bit_width)
846
+ return self
847
+
848
+
849
+ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
850
+ def __init__(self, config: ChatGLMConfig, empty_init=True, device=None):
851
+ super().__init__(config)
852
+
853
+ self.max_sequence_length = config.max_length
854
+ self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
855
+ self.config = config
856
+ self.quantized = False
857
+
858
+ if self.config.quantization_bit:
859
+ self.quantize(self.config.quantization_bit, empty_init=True)
860
+
861
+ def _update_model_kwargs_for_generation(
862
+ self,
863
+ outputs: ModelOutput,
864
+ model_kwargs: Dict[str, Any],
865
+ is_encoder_decoder: bool = False,
866
+ standardize_cache_format: bool = False,
867
+ ) -> Dict[str, Any]:
868
+ # update past_key_values
869
+ model_kwargs["past_key_values"] = self._extract_past_from_model_output(
870
+ outputs, standardize_cache_format=standardize_cache_format
871
+ )
872
+
873
+ # update attention mask
874
+ if "attention_mask" in model_kwargs:
875
+ attention_mask = model_kwargs["attention_mask"]
876
+ model_kwargs["attention_mask"] = torch.cat(
877
+ [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
878
+ )
879
+
880
+ # update position ids
881
+ if "position_ids" in model_kwargs:
882
+ position_ids = model_kwargs["position_ids"]
883
+ new_position_id = position_ids[..., -1:].clone()
884
+ new_position_id += 1
885
+ model_kwargs["position_ids"] = torch.cat(
886
+ [position_ids, new_position_id], dim=-1
887
+ )
888
+
889
+ model_kwargs["is_first_forward"] = False
890
+ return model_kwargs
891
+
892
+ def prepare_inputs_for_generation(
893
+ self,
894
+ input_ids: torch.LongTensor,
895
+ past_key_values: Optional[torch.Tensor] = None,
896
+ attention_mask: Optional[torch.Tensor] = None,
897
+ position_ids: Optional[torch.Tensor] = None,
898
+ is_first_forward: bool = True,
899
+ **kwargs
900
+ ) -> dict:
901
+ # only last token for input_ids if past is not None
902
+ if position_ids is None:
903
+ position_ids = self.get_position_ids(input_ids, device=input_ids.device)
904
+ if not is_first_forward:
905
+ position_ids = position_ids[..., -1:]
906
+ input_ids = input_ids[:, -1:]
907
+ return {
908
+ "input_ids": input_ids,
909
+ "past_key_values": past_key_values,
910
+ "position_ids": position_ids,
911
+ "attention_mask": attention_mask,
912
+ "return_last_logit": True
913
+ }
914
+
915
+ def forward(
916
+ self,
917
+ input_ids: Optional[torch.Tensor] = None,
918
+ position_ids: Optional[torch.Tensor] = None,
919
+ attention_mask: Optional[torch.Tensor] = None,
920
+ past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
921
+ inputs_embeds: Optional[torch.Tensor] = None,
922
+ labels: Optional[torch.Tensor] = None,
923
+ use_cache: Optional[bool] = None,
924
+ output_attentions: Optional[bool] = None,
925
+ output_hidden_states: Optional[bool] = None,
926
+ return_dict: Optional[bool] = None,
927
+ return_last_logit: Optional[bool] = False,
928
+ ):
929
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
930
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
931
+
932
+ transformer_outputs = self.transformer(
933
+ input_ids=input_ids,
934
+ position_ids=position_ids,
935
+ attention_mask=attention_mask,
936
+ past_key_values=past_key_values,
937
+ inputs_embeds=inputs_embeds,
938
+ use_cache=use_cache,
939
+ output_hidden_states=output_hidden_states,
940
+ return_dict=return_dict,
941
+ )
942
+
943
+ hidden_states = transformer_outputs[0]
944
+ if return_last_logit:
945
+ hidden_states = hidden_states[-1:]
946
+ lm_logits = self.transformer.output_layer(hidden_states)
947
+ lm_logits = lm_logits.transpose(0, 1).contiguous()
948
+
949
+ loss = None
950
+ if labels is not None:
951
+ lm_logits = lm_logits.to(torch.float32)
952
+
953
+ # Shift so that tokens < n predict n
954
+ shift_logits = lm_logits[..., :-1, :].contiguous()
955
+ shift_labels = labels[..., 1:].contiguous()
956
+ # Flatten the tokens
957
+ loss_fct = CrossEntropyLoss(ignore_index=-100)
958
+ loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
959
+
960
+ lm_logits = lm_logits.to(hidden_states.dtype)
961
+ loss = loss.to(hidden_states.dtype)
962
+
963
+ if not return_dict:
964
+ output = (lm_logits,) + transformer_outputs[1:]
965
+ return ((loss,) + output) if loss is not None else output
966
+
967
+ return CausalLMOutputWithPast(
968
+ loss=loss,
969
+ logits=lm_logits,
970
+ past_key_values=transformer_outputs.past_key_values,
971
+ hidden_states=transformer_outputs.hidden_states,
972
+ attentions=transformer_outputs.attentions,
973
+ )
974
+
975
+ @staticmethod
976
+ def _reorder_cache(
977
+ past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
978
+ ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
979
+ """
980
+ This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
981
+ [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
982
+ beam_idx at every generation step.
983
+
984
+ Output shares the same memory storage as `past`.
985
+ """
986
+ return tuple(
987
+ (
988
+ layer_past[0].index_select(1, beam_idx.to(layer_past[0].device)),
989
+ layer_past[1].index_select(1, beam_idx.to(layer_past[1].device)),
990
+ )
991
+ for layer_past in past
992
+ )
993
+
994
+ def process_response(self, response):
995
+ response = response.strip()
996
+ response = response.replace("[[训练时间]]", "2023年")
997
+ return response
998
+
999
+ def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = None):
1000
+ prompt = tokenizer.build_prompt(query, history=history)
1001
+ inputs = tokenizer([prompt], return_tensors="pt")
1002
+ inputs = inputs.to(self.device)
1003
+ return inputs
1004
+
1005
+ def build_stream_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = None):
1006
+ if history:
1007
+ prompt = "\n\n[Round {}]\n\n问:{}\n\n答:".format(len(history) + 1, query)
1008
+ input_ids = tokenizer.encode(prompt, add_special_tokens=False)
1009
+ input_ids = input_ids[1:]
1010
+ inputs = tokenizer.batch_encode_plus([(input_ids, None)], return_tensors="pt", add_special_tokens=False)
1011
+ else:
1012
+ prompt = "[Round {}]\n\n问:{}\n\n答:".format(len(history) + 1, query)
1013
+ inputs = tokenizer([prompt], return_tensors="pt")
1014
+ inputs = inputs.to(self.device)
1015
+ return inputs
1016
+
1017
+ @torch.inference_mode()
1018
+ def chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 8192, num_beams=1,
1019
+ do_sample=True, top_p=0.8, temperature=0.8, logits_processor=None, **kwargs):
1020
+ if history is None:
1021
+ history = []
1022
+ if logits_processor is None:
1023
+ logits_processor = LogitsProcessorList()
1024
+ logits_processor.append(InvalidScoreLogitsProcessor())
1025
+ gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
1026
+ "temperature": temperature, "logits_processor": logits_processor, **kwargs}
1027
+ inputs = self.build_inputs(tokenizer, query, history=history)
1028
+ outputs = self.generate(**inputs, **gen_kwargs)
1029
+ outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
1030
+ response = tokenizer.decode(outputs)
1031
+ response = self.process_response(response)
1032
+ history = history + [(query, response)]
1033
+ return response, history
1034
+
1035
+ @torch.inference_mode()
1036
+ def stream_chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, past_key_values=None,
1037
+ max_length: int = 8192, do_sample=True, top_p=0.8, temperature=0.8, logits_processor=None,
1038
+ return_past_key_values=False, **kwargs):
1039
+ if history is None:
1040
+ history = []
1041
+ if logits_processor is None:
1042
+ logits_processor = LogitsProcessorList()
1043
+ logits_processor.append(InvalidScoreLogitsProcessor())
1044
+ gen_kwargs = {"max_length": max_length, "do_sample": do_sample, "top_p": top_p,
1045
+ "temperature": temperature, "logits_processor": logits_processor, **kwargs}
1046
+ if past_key_values is None and not return_past_key_values:
1047
+ inputs = self.build_inputs(tokenizer, query, history=history)
1048
+ else:
1049
+ inputs = self.build_stream_inputs(tokenizer, query, history=history)
1050
+ if past_key_values is not None:
1051
+ past_length = past_key_values[0][0].shape[0]
1052
+ if self.transformer.pre_seq_len is not None:
1053
+ past_length -= self.transformer.pre_seq_len
1054
+ inputs.position_ids += past_length
1055
+ attention_mask = inputs.attention_mask
1056
+ attention_mask = torch.cat((attention_mask.new_ones(1, past_length), attention_mask), dim=1)
1057
+ inputs['attention_mask'] = attention_mask
1058
+ for outputs in self.stream_generate(**inputs, past_key_values=past_key_values,
1059
+ return_past_key_values=return_past_key_values, **gen_kwargs):
1060
+ if return_past_key_values:
1061
+ outputs, past_key_values = outputs
1062
+ outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
1063
+ response = tokenizer.decode(outputs)
1064
+ if response and response[-1] != "�":
1065
+ response = self.process_response(response)
1066
+ new_history = history + [(query, response)]
1067
+ if return_past_key_values:
1068
+ yield response, new_history, past_key_values
1069
+ else:
1070
+ yield response, new_history
1071
+
1072
+ @torch.inference_mode()
1073
+ def stream_generate(
1074
+ self,
1075
+ input_ids,
1076
+ generation_config: Optional[GenerationConfig] = None,
1077
+ logits_processor: Optional[LogitsProcessorList] = None,
1078
+ stopping_criteria: Optional[StoppingCriteriaList] = None,
1079
+ prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
1080
+ return_past_key_values=False,
1081
+ **kwargs,
1082
+ ):
1083
+ batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]
1084
+
1085
+ if generation_config is None:
1086
+ generation_config = self.generation_config
1087
+ generation_config = copy.deepcopy(generation_config)
1088
+ model_kwargs = generation_config.update(**kwargs)
1089
+ bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id
1090
+
1091
+ if isinstance(eos_token_id, int):
1092
+ eos_token_id = [eos_token_id]
1093
+
1094
+ has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
1095
+ if has_default_max_length and generation_config.max_new_tokens is None:
1096
+ warnings.warn(
1097
+ f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. "
1098
+ "This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we"
1099
+ " recommend using `max_new_tokens` to control the maximum length of the generation.",
1100
+ UserWarning,
1101
+ )
1102
+ elif generation_config.max_new_tokens is not None:
1103
+ generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
1104
+ if not has_default_max_length:
1105
+ logger.warn(
1106
+ f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
1107
+ f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
1108
+ "Please refer to the documentation for more information. "
1109
+ "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)",
1110
+ UserWarning,
1111
+ )
1112
+
1113
+ if input_ids_seq_length >= generation_config.max_length:
1114
+ input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
1115
+ logger.warning(
1116
+ f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
1117
+ f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
1118
+ " increasing `max_new_tokens`."
1119
+ )
1120
+
1121
+ # 2. Set generation parameters if not already defined
1122
+ logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
1123
+ stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
1124
+
1125
+ logits_processor = self._get_logits_processor(
1126
+ generation_config=generation_config,
1127
+ input_ids_seq_length=input_ids_seq_length,
1128
+ encoder_input_ids=input_ids,
1129
+ prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
1130
+ logits_processor=logits_processor,
1131
+ )
1132
+
1133
+ stopping_criteria = self._get_stopping_criteria(
1134
+ generation_config=generation_config, stopping_criteria=stopping_criteria
1135
+ )
1136
+ logits_warper = self._get_logits_warper(generation_config)
1137
+
1138
+ unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
1139
+ scores = None
1140
+ while True:
1141
+ model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
1142
+ # forward pass to get next token
1143
+ outputs = self(
1144
+ **model_inputs,
1145
+ return_dict=True,
1146
+ output_attentions=False,
1147
+ output_hidden_states=False,
1148
+ )
1149
+
1150
+ next_token_logits = outputs.logits[:, -1, :]
1151
+
1152
+ # pre-process distribution
1153
+ next_token_scores = logits_processor(input_ids, next_token_logits)
1154
+ next_token_scores = logits_warper(input_ids, next_token_scores)
1155
+
1156
+ # sample
1157
+ probs = nn.functional.softmax(next_token_scores, dim=-1)
1158
+ if generation_config.do_sample:
1159
+ next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
1160
+ else:
1161
+ next_tokens = torch.argmax(probs, dim=-1)
1162
+
1163
+ # update generated ids, model inputs, and length for next step
1164
+ input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
1165
+ model_kwargs = self._update_model_kwargs_for_generation(
1166
+ outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
1167
+ )
1168
+ unfinished_sequences = unfinished_sequences.mul((sum(next_tokens != i for i in eos_token_id)).long())
1169
+ if return_past_key_values:
1170
+ yield input_ids, outputs.past_key_values
1171
+ else:
1172
+ yield input_ids
1173
+ # stop when each sentence is finished, or if we exceed the maximum length
1174
+ if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
1175
+ break
1176
+
1177
+ def quantize(self, bits: int, empty_init=False, device=None, **kwargs):
1178
+ if bits == 0:
1179
+ return
1180
+
1181
+ from .quantization import quantize
1182
+
1183
+ if self.quantized:
1184
+ logger.info("Already quantized.")
1185
+ return self
1186
+
1187
+ self.quantized = True
1188
+
1189
+ self.config.quantization_bit = bits
1190
+
1191
+ self.transformer.encoder = quantize(self.transformer.encoder, bits, empty_init=empty_init, device=device,
1192
+ **kwargs)
1193
+ return self
pytorch_model.bin DELETED
@@ -1 +0,0 @@
1
- ../../blobs/783e9eac82c278f1e57ba93344a96580fd6f2cd222c8703f25998bb16926688e
 
 
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:783e9eac82c278f1e57ba93344a96580fd6f2cd222c8703f25998bb16926688e
3
+ size 3923712365
quantization.py DELETED
@@ -1 +0,0 @@
1
- ../../blobs/401496776e4033da6aa14f05cbf01b480eb17427
 
 
quantization.py ADDED
The diff for this file is too large to render. See raw diff
 
tokenization_chatglm.py DELETED
@@ -1 +0,0 @@
1
- ../../blobs/af6c93b525bba989462106165c8f9d5d01c0e9f1
 
 
tokenization_chatglm.py ADDED
@@ -0,0 +1,257 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import torch
3
+ from typing import List, Optional, Union, Dict
4
+ from sentencepiece import SentencePieceProcessor
5
+ from transformers import PreTrainedTokenizer
6
+ from transformers.utils import logging, PaddingStrategy
7
+ from transformers.tokenization_utils_base import EncodedInput, BatchEncoding
8
+
9
+
10
+ class SPTokenizer:
11
+ def __init__(self, model_path: str):
12
+ # reload tokenizer
13
+ assert os.path.isfile(model_path), model_path
14
+ self.sp_model = SentencePieceProcessor(model_file=model_path)
15
+
16
+ # BOS / EOS token IDs
17
+ self.n_words: int = self.sp_model.vocab_size()
18
+ self.bos_id: int = self.sp_model.bos_id()
19
+ self.eos_id: int = self.sp_model.eos_id()
20
+ self.pad_id: int = self.sp_model.unk_id()
21
+ assert self.sp_model.vocab_size() == self.sp_model.get_piece_size()
22
+
23
+ special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"]
24
+ self.special_tokens = {}
25
+ self.index_special_tokens = {}
26
+ for token in special_tokens:
27
+ self.special_tokens[token] = self.n_words
28
+ self.index_special_tokens[self.n_words] = token
29
+ self.n_words += 1
30
+
31
+ def tokenize(self, s: str):
32
+ return self.sp_model.EncodeAsPieces(s)
33
+
34
+ def encode(self, s: str, bos: bool = False, eos: bool = False) -> List[int]:
35
+ assert type(s) is str
36
+ t = self.sp_model.encode(s)
37
+ if bos:
38
+ t = [self.bos_id] + t
39
+ if eos:
40
+ t = t + [self.eos_id]
41
+ return t
42
+
43
+ def decode(self, t: List[int]) -> str:
44
+ return self.sp_model.decode(t)
45
+
46
+ def decode_tokens(self, tokens: List[str]) -> str:
47
+ text = self.sp_model.DecodePieces(tokens)
48
+ return text
49
+
50
+ def convert_token_to_id(self, token):
51
+ """ Converts a token (str) in an id using the vocab. """
52
+ if token in self.special_tokens:
53
+ return self.special_tokens[token]
54
+ return self.sp_model.PieceToId(token)
55
+
56
+ def convert_id_to_token(self, index):
57
+ """Converts an index (integer) in a token (str) using the vocab."""
58
+ if index in self.index_special_tokens or index in [self.eos_id, self.bos_id, self.pad_id] or index < 0:
59
+ return ""
60
+ return self.sp_model.IdToPiece(index)
61
+
62
+
63
+ class ChatGLMTokenizer(PreTrainedTokenizer):
64
+ vocab_files_names = {"vocab_file": "tokenizer.model"}
65
+
66
+ model_input_names = ["input_ids", "attention_mask", "position_ids"]
67
+
68
+ def __init__(self, vocab_file, padding_side="left", **kwargs):
69
+ super().__init__(padding_side=padding_side, clean_up_tokenization_spaces=False, **kwargs)
70
+ self.name = "GLMTokenizer"
71
+
72
+ self.vocab_file = vocab_file
73
+ self.tokenizer = SPTokenizer(vocab_file)
74
+ self.special_tokens = {
75
+ "<bos>": self.tokenizer.bos_id,
76
+ "<eos>": self.tokenizer.eos_id,
77
+ "<pad>": self.tokenizer.pad_id
78
+ }
79
+
80
+ def get_command(self, token):
81
+ if token in self.special_tokens:
82
+ return self.special_tokens[token]
83
+ assert token in self.tokenizer.special_tokens, f"{token} is not a special token for {self.name}"
84
+ return self.tokenizer.special_tokens[token]
85
+
86
+ @property
87
+ def unk_token(self) -> str:
88
+ return "<unk>"
89
+
90
+ @property
91
+ def pad_token(self) -> str:
92
+ return "<unk>"
93
+
94
+ @property
95
+ def pad_token_id(self):
96
+ return self.get_command("<pad>")
97
+
98
+ @property
99
+ def eos_token(self) -> str:
100
+ return "</s>"
101
+
102
+ @property
103
+ def eos_token_id(self):
104
+ return self.get_command("<eos>")
105
+
106
+ @property
107
+ def vocab_size(self):
108
+ return self.tokenizer.n_words
109
+
110
+ def get_vocab(self):
111
+ """ Returns vocab as a dict """
112
+ vocab = {self._convert_id_to_token(i): i for i in range(self.vocab_size)}
113
+ vocab.update(self.added_tokens_encoder)
114
+ return vocab
115
+
116
+ def _tokenize(self, text, **kwargs):
117
+ return self.tokenizer.tokenize(text)
118
+
119
+ def _convert_token_to_id(self, token):
120
+ """ Converts a token (str) in an id using the vocab. """
121
+ return self.tokenizer.convert_token_to_id(token)
122
+
123
+ def _convert_id_to_token(self, index):
124
+ """Converts an index (integer) in a token (str) using the vocab."""
125
+ return self.tokenizer.convert_id_to_token(index)
126
+
127
+ def convert_tokens_to_string(self, tokens: List[str]) -> str:
128
+ return self.tokenizer.decode_tokens(tokens)
129
+
130
+ def save_vocabulary(self, save_directory, filename_prefix=None):
131
+ """
132
+ Save the vocabulary and special tokens file to a directory.
133
+
134
+ Args:
135
+ save_directory (`str`):
136
+ The directory in which to save the vocabulary.
137
+ filename_prefix (`str`, *optional*):
138
+ An optional prefix to add to the named of the saved files.
139
+
140
+ Returns:
141
+ `Tuple(str)`: Paths to the files saved.
142
+ """
143
+ if os.path.isdir(save_directory):
144
+ vocab_file = os.path.join(
145
+ save_directory, self.vocab_files_names["vocab_file"]
146
+ )
147
+ else:
148
+ vocab_file = save_directory
149
+
150
+ with open(self.vocab_file, 'rb') as fin:
151
+ proto_str = fin.read()
152
+
153
+ with open(vocab_file, "wb") as writer:
154
+ writer.write(proto_str)
155
+
156
+ return (vocab_file,)
157
+
158
+ def get_prefix_tokens(self):
159
+ prefix_tokens = [self.get_command("[gMASK]"), self.get_command("sop")]
160
+ return prefix_tokens
161
+
162
+ def build_prompt(self, query, history=None):
163
+ if history is None:
164
+ history = []
165
+ prompt = ""
166
+ for i, (old_query, response) in enumerate(history):
167
+ prompt += "[Round {}]\n\n问:{}\n\n答:{}\n\n".format(i + 1, old_query, response)
168
+ prompt += "[Round {}]\n\n问:{}\n\n答:".format(len(history) + 1, query)
169
+ return prompt
170
+
171
+ def build_inputs_with_special_tokens(
172
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
173
+ ) -> List[int]:
174
+ """
175
+ Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
176
+ adding special tokens. A BERT sequence has the following format:
177
+
178
+ - single sequence: `[CLS] X [SEP]`
179
+ - pair of sequences: `[CLS] A [SEP] B [SEP]`
180
+
181
+ Args:
182
+ token_ids_0 (`List[int]`):
183
+ List of IDs to which the special tokens will be added.
184
+ token_ids_1 (`List[int]`, *optional*):
185
+ Optional second list of IDs for sequence pairs.
186
+
187
+ Returns:
188
+ `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
189
+ """
190
+ prefix_tokens = self.get_prefix_tokens()
191
+ token_ids_0 = prefix_tokens + token_ids_0
192
+ if token_ids_1 is not None:
193
+ token_ids_0 = token_ids_0 + token_ids_1 + [self.get_command("<eos>")]
194
+ return token_ids_0
195
+
196
+ def _pad(
197
+ self,
198
+ encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding],
199
+ max_length: Optional[int] = None,
200
+ padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
201
+ pad_to_multiple_of: Optional[int] = None,
202
+ return_attention_mask: Optional[bool] = None,
203
+ ) -> dict:
204
+ """
205
+ Pad encoded inputs (on left/right and up to predefined length or max length in the batch)
206
+
207
+ Args:
208
+ encoded_inputs:
209
+ Dictionary of tokenized inputs (`List[int]`) or batch of tokenized inputs (`List[List[int]]`).
210
+ max_length: maximum length of the returned list and optionally padding length (see below).
211
+ Will truncate by taking into account the special tokens.
212
+ padding_strategy: PaddingStrategy to use for padding.
213
+
214
+ - PaddingStrategy.LONGEST Pad to the longest sequence in the batch
215
+ - PaddingStrategy.MAX_LENGTH: Pad to the max length (default)
216
+ - PaddingStrategy.DO_NOT_PAD: Do not pad
217
+ The tokenizer padding sides are defined in self.padding_side:
218
+
219
+ - 'left': pads on the left of the sequences
220
+ - 'right': pads on the right of the sequences
221
+ pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value.
222
+ This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability
223
+ `>= 7.5` (Volta).
224
+ return_attention_mask:
225
+ (optional) Set to False to avoid returning attention mask (default: set to model specifics)
226
+ """
227
+ # Load from model defaults
228
+ assert self.padding_side == "left"
229
+
230
+ required_input = encoded_inputs[self.model_input_names[0]]
231
+ seq_length = len(required_input)
232
+
233
+ if padding_strategy == PaddingStrategy.LONGEST:
234
+ max_length = len(required_input)
235
+
236
+ if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
237
+ max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
238
+
239
+ needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length
240
+
241
+ # Initialize attention mask if not present.
242
+ if "attention_mask" not in encoded_inputs:
243
+ encoded_inputs["attention_mask"] = [1] * seq_length
244
+
245
+ if "position_ids" not in encoded_inputs:
246
+ encoded_inputs["position_ids"] = list(range(seq_length))
247
+
248
+ if needs_to_be_padded:
249
+ difference = max_length - len(required_input)
250
+
251
+ if "attention_mask" in encoded_inputs:
252
+ encoded_inputs["attention_mask"] = [0] * difference + encoded_inputs["attention_mask"]
253
+ if "position_ids" in encoded_inputs:
254
+ encoded_inputs["position_ids"] = [0] * difference + encoded_inputs["position_ids"]
255
+ encoded_inputs[self.model_input_names[0]] = [self.pad_token_id] * difference + required_input
256
+
257
+ return encoded_inputs
tokenizer.model DELETED
@@ -1 +0,0 @@
1
- ../../blobs/e7dc4c393423b76e4373e5157ddc34803a0189ba96b21ddbb40269d31468a6f2
 
 
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e7dc4c393423b76e4373e5157ddc34803a0189ba96b21ddbb40269d31468a6f2
3
+ size 1018370
tokenizer_config.json DELETED
@@ -1 +0,0 @@
1
- ../../blobs/ac27c0b95da639fcf952cbee6f87aaadeb8f159a
 
 
tokenizer_config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name_or_path": "THUDM/chatglm-6b",
3
+ "remove_space": false,
4
+ "do_lower_case": false,
5
+ "tokenizer_class": "ChatGLMTokenizer",
6
+ "auto_map": {
7
+ "AutoTokenizer": [
8
+ "tokenization_chatglm.ChatGLMTokenizer",
9
+ null
10
+ ]
11
+ }
12
+ }