hf
Browse files- README.md +97 -3
- config.json +50 -0
- configuration.json +1 -0
- configuration_yuan.py +44 -0
- generation_config.json +7 -0
- pytorch_model-00001-of-00009.bin +3 -0
- pytorch_model.bin.index.json +0 -0
- special_tokens_map.json +23 -0
- tokenizer.model +3 -0
- tokenizer_config.json +33 -0
- yuan_moe_hf_model.py +1454 -0
README.md
CHANGED
@@ -1,3 +1,97 @@
|
|
1 |
-
---
|
2 |
-
license:
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: other
|
3 |
+
license_name: license-yuan
|
4 |
+
license_link: https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan
|
5 |
+
---
|
6 |
+
|
7 |
+
<div align="center">
|
8 |
+
<h1>
|
9 |
+
Yuan 2
|
10 |
+
</h1>
|
11 |
+
</div>
|
12 |
+
|
13 |
+
<div align="center">
|
14 |
+
<a href="https://github.com/IEIT-Yuan/Yuan-2.0" target="_blank"> 💻GitHub Repo</a> | <a href="http://arxiv.org/pdf/2311.15786.pdf" target="_blank">📃Yuan2.0-paper</a>
|
15 |
+
</div>
|
16 |
+
|
17 |
+
# 目录/Table of Contents
|
18 |
+
|
19 |
+
- [模型介绍/Introduction](#Introduction)
|
20 |
+
- [代码调用/Code Usage](#Usage)
|
21 |
+
- [Benchmark评估/Benchmark Evaluation](#Benchmark)
|
22 |
+
- [声明与协议/Terms and Conditions](#Terms)
|
23 |
+
- [引用/Cite](#Cite)
|
24 |
+
|
25 |
+
|
26 |
+
# <span id="Introduction">模型介绍/Introduction</span>
|
27 |
+
源2.0 是浪潮信息发布的新一代基础语言大模型。我们开源了全部的3个模型源2.0-102B,源2.0-51B和源2.0-2B。并且我们提供了预训练,微调,推理服务的相关脚本,以供研发人员做进一步的开发。源2.0是在源1.0的基础上,利用更多样的高质量预训练数据和指令微调数据集,令模型在语义、数学、推理、代码、知识等不同方面具备更强的理解能力。
|
28 |
+
|
29 |
+
Yuan2.0 is a new generation Fundamental Large Language Model developed by IEIT System. We have published all three models, Yuan 2.0-102B, Yuan 2.0-51B, and Yuan 2.0-2B. And we provide relevant scripts for pretraining, fine-tuning, and inference services for other developers. Yuan2.0 is based on Yuan1.0, utilizing a wider range of high-quality pre training data and instruction fine-tuning datasets to enhance the model's understanding of semantics, mathematics, reasoning, code, knowledge, and other aspects.
|
30 |
+
|
31 |
+
|
32 |
+
# <span id="Usage">代码调用/Code Usage</span>
|
33 |
+
可以通过如下代码调用 Yuan2-2B-MoE 模型来生成文本:
|
34 |
+
|
35 |
+
You can generate text by invoking the Yuan2-2B-MoE model with the following code:
|
36 |
+
|
37 |
+
```python
|
38 |
+
import torch, transformers
|
39 |
+
import sys, os
|
40 |
+
sys.path.append(
|
41 |
+
os.path.abspath(os.path.join(os.path.dirname(__file__), os.path.pardir)))
|
42 |
+
from transformers import AutoModelForCausalLM,AutoTokenizer,LlamaTokenizer
|
43 |
+
|
44 |
+
print("Creat tokenizer...")
|
45 |
+
tokenizer = LlamaTokenizer.from_pretrained('IEITYuan/Yuan2-2B-hf-moe', add_eos_token=False, add_bos_token=False, eos_token='<eod>')
|
46 |
+
tokenizer.add_tokens(['<sep>', '<pad>', '<mask>', '<predict>', '<FIM_SUFFIX>', '<FIM_PREFIX>', '<FIM_MIDDLE>','<commit_before>','<commit_msg>','<commit_after>','<jupyter_start>','<jupyter_text>','<jupyter_code>','<jupyter_output>','<empty_output>'], special_tokens=True)
|
47 |
+
|
48 |
+
print("Creat model...")
|
49 |
+
model = AutoModelForCausalLM.from_pretrained('IEITYuan/Yuan2-2B-hf-moe', device_map='auto', torch_dtype=torch.bfloat16, trust_remote_code=True)
|
50 |
+
|
51 |
+
inputs = tokenizer("请问目前最先进的机器学习算法有哪些?", return_tensors="pt")["input_ids"].to("cuda:0")
|
52 |
+
outputs = model.generate(inputs,do_sample=False,max_length=100)
|
53 |
+
print(tokenizer.decode(outputs[0]))
|
54 |
+
|
55 |
+
```
|
56 |
+
|
57 |
+
# <span id="Benchmark">Benchmark评估/Benchmark Evaluation</span>
|
58 |
+
我们提供了[HumanEval](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/docs/eval_humaneval.md),[AGIEval-GK-Math](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/docs/eval_agieval_math.md),[GSM8K](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/docs/eval_gsm8k.md)和[TruthfulQA](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/docs/eval_TruthfulQA.md)的评估脚本。在4个典型任务上,我们用源2.0不同版本模型上进行了性能测试。
|
59 |
+
|
60 |
+
We have provided evaluation scripts for [HumanEval](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/docs/eval_humaneval.md),[AGIEval-GK-Math](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/docs/eval_agieval_math.md),[GSM8K](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/docs/eval_gsm8k.md) and [TruthfulQA](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/docs/eval_TruthfulQA.md). Performance tests were conducted on different versions of the Yuan2.0 model for four typical tasks.
|
61 |
+
|
62 |
+
|
63 |
+
| Model | GSM8K | AGIEval-GK-Math-QA | AGIEval-GK-Math-Cloze | HumanEval | TurthfulQA |
|
64 |
+
| ----------------- | :----: | :------------: | :---------------: | :-------: | ---------- |
|
65 |
+
| GPT-4 | 92% | 47.0% | 16.1% | 86.6% | 59% |
|
66 |
+
| ChatGPT | 68.6%\* | 36.5% | 7.3% | 66.5%\* | 34%\* |
|
67 |
+
| Llama2 | 56.8% | - | - | 29.9% | - |
|
68 |
+
| 源2.0-102B | 76.6% | 38.7% | 13.5% | 67.1% | 58% |
|
69 |
+
| 源2.0-102B-SC | 86.2% | 45.5% | 15.2% | 77.4% | - |
|
70 |
+
|
71 |
+
\* 使用与源2.0完全相同的输入数据对ChatGPT进行测试,时间2023年11月
|
72 |
+
|
73 |
+
\* Testing ChatGPT using the same input data as Yuan2.0, as of November 2023.
|
74 |
+
|
75 |
+
# <span id="Terms">声明与协议/Terms and Conditions</span>
|
76 |
+
对该模型的原代码仓库使用遵循开源许可协议 Apache 2.0。
|
77 |
+
|
78 |
+
源2.0模型支持商用,不需要申请授权,请您了解并遵循[《源2.0模型许可协议》](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan),勿将开源模型和代码及基于开源项目产生的衍生物用于任何可能给国家和社会带来危害的用途以及用于任何未经过安全评估和备案的服务。
|
79 |
+
|
80 |
+
尽管模型在训练时我们已采取措施尽力确保数据的合规性和准确性,但模型参数量巨大且受概率随机性因素影响,我们无法保证输出内容的准确性,且模型易被输入指令所误导,本项目不承担开源模型和代码导致的数据安全、舆情风险或发生任何模型被误导、滥用、传播、不当利用而产生的风险和责任。**您将对通过使用、复制、分发和修改模型等方式利用该开源项目所产生的风险与后果,独自承担全部责任。**
|
81 |
+
|
82 |
+
The use of the original code repository for this model requires compliance with the open source license agreement Apache 2.0. The Yuan2.0 model supports commercial use and does not require authorization. Please understand and comply with the [《Yuan 2.0 Model License Agreement》](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan). Do not use the open source model and code, as well as derivatives generated from open source projects, for any purposes that may cause harm to the country and society, or for any services that have not undergone security assessment and filing. Although we have taken measures to ensure the compliance and accuracy of the data during training, the model has a huge number of parameters and is affected by probability and randomness factors. We cannot guarantee the accuracy of the output content, and the model is easily misled by input instructions. This project does not assume any data security, public opinion risks, or any model misleading, abusing, spreading caused by open-source models and code Risks and responsibilities arising from improper utilization **You will be solely responsible for the risks and consequences arising from the use, copying, distribution, and modification of the model in this open source project.**
|
83 |
+
|
84 |
+
# <span id="Cite">引用/Cite</span>
|
85 |
+
欢迎阅读我们的技术报告 [YUAN 2.0: A Large Language Model with Localized Filtering-based Attention](http://arxiv.org/pdf/2311.15786.pdf)!
|
86 |
+
|
87 |
+
Welcome to read our technical report [YUAN 2.0: A Large Language Model with Localized Filtering-based Attention](http://arxiv.org/pdf/2311.15786.pdf)!
|
88 |
+
|
89 |
+
```latex
|
90 |
+
@article{Wu2023,
|
91 |
+
title = {{YUAN 2.0: A Large Language Model with Localized Filtering-based Attention}},
|
92 |
+
author = {Wu, Shaohua and Zhao, Xudong and Wang, Shenling and Luo, Jiangang and Li, Lingjun and Chen, Xi and Zhao, Bing and Wang, Wei and Yu, Tong and Zhang, Rongguo and Zhang, Jiahua and Wang, Chao},
|
93 |
+
url = {http://arxiv.org/abs/2311.15786},
|
94 |
+
year = {2023}
|
95 |
+
}
|
96 |
+
|
97 |
+
```
|
config.json
ADDED
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_from_model_config": true,
|
3 |
+
"_name_or_path": "/mnt/beegfs2/sunzeyu/bin_522-7/",
|
4 |
+
"architectures": [
|
5 |
+
"YuanForCausalLM"
|
6 |
+
],
|
7 |
+
"attention_projection_size": 4096,
|
8 |
+
"auto_map": {
|
9 |
+
"AutoConfig": "configuration_yuan.YuanConfig",
|
10 |
+
"AutoModelForCausalLM": "yuan_hf_model.YuanForCausalLM"
|
11 |
+
},
|
12 |
+
"bos_token_id": 77185,
|
13 |
+
"causal_mask": true,
|
14 |
+
"dropout": 0,
|
15 |
+
"eod_token": 77185,
|
16 |
+
"eod_token_id": 77185,
|
17 |
+
"eos_token_id": 77185,
|
18 |
+
"hidden_act": "silu",
|
19 |
+
"hidden_size": 2048,
|
20 |
+
"initializer_range": 0.02,
|
21 |
+
"intermediate_size": 8192,
|
22 |
+
"mask_token_id": 77185,
|
23 |
+
"max_position_embeddings": 4096,
|
24 |
+
"model_max_length": 8192,
|
25 |
+
"model_type": "yuan",
|
26 |
+
"moe_config": {
|
27 |
+
"ffn_hidden_size": 8192,
|
28 |
+
"gated_linear_unit": true,
|
29 |
+
"moe_num_experts": 32,
|
30 |
+
"moe_top_k": 2,
|
31 |
+
"norm_topk_prob": true
|
32 |
+
},
|
33 |
+
"num_attention_heads": 16,
|
34 |
+
"num_hidden_layers": 24,
|
35 |
+
"output_router_logits": true,
|
36 |
+
"pad_token_id": 77185,
|
37 |
+
"reset_attention_mask": false,
|
38 |
+
"reset_position_ids": true,
|
39 |
+
"rms_norm_eps": 1e-06,
|
40 |
+
"sep_token": 77187,
|
41 |
+
"sep_token_id": 77185,
|
42 |
+
"tokenizer_class": "YuanTokenizer",
|
43 |
+
"torch_dtype": "bfloat16",
|
44 |
+
"transformers_version": "4.30.2",
|
45 |
+
"use_cache": true,
|
46 |
+
"use_flash_attention": true,
|
47 |
+
"use_loss_mask": false,
|
48 |
+
"use_moe": true,
|
49 |
+
"vocab_size": 135040
|
50 |
+
}
|
configuration.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"framework":"Pytorch","task":"chatbot"}
|
configuration_yuan.py
ADDED
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
from transformers.configuration_utils import PretrainedConfig
|
3 |
+
|
4 |
+
|
5 |
+
class YuanConfig(PretrainedConfig):
|
6 |
+
model_type = "yuan"
|
7 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
8 |
+
|
9 |
+
def __init__(
|
10 |
+
self,
|
11 |
+
vocab_size=135040,
|
12 |
+
hidden_size=2048,
|
13 |
+
intermediate_size=8192,
|
14 |
+
num_hidden_layers=24,
|
15 |
+
num_attention_heads=32,
|
16 |
+
hidden_act="silu",
|
17 |
+
model_max_length=8192,
|
18 |
+
initializer_range=0.02,
|
19 |
+
rms_norm_eps=1e-6,
|
20 |
+
use_cache=True,
|
21 |
+
pad_token_id=77185,
|
22 |
+
bos_token_id=77185,
|
23 |
+
eos_token_id=77185,
|
24 |
+
tie_word_embeddings=True,
|
25 |
+
**kwargs,
|
26 |
+
):
|
27 |
+
self.vocab_size = vocab_size
|
28 |
+
self.model_max_length = model_max_length
|
29 |
+
self.hidden_size = hidden_size
|
30 |
+
self.intermediate_size = intermediate_size
|
31 |
+
self.num_hidden_layers = num_hidden_layers
|
32 |
+
self.num_attention_heads = num_attention_heads
|
33 |
+
self.hidden_act = hidden_act
|
34 |
+
self.initializer_range = initializer_range
|
35 |
+
self.rms_norm_eps = rms_norm_eps
|
36 |
+
self.use_cache = use_cache
|
37 |
+
super().__init__(
|
38 |
+
pad_token_id=pad_token_id,
|
39 |
+
bos_token_id=bos_token_id,
|
40 |
+
eos_token_id=eos_token_id,
|
41 |
+
tie_word_embeddings=tie_word_embeddings,
|
42 |
+
**kwargs,
|
43 |
+
)
|
44 |
+
|
generation_config.json
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_from_model_config": true,
|
3 |
+
"bos_token_id": 77185,
|
4 |
+
"eos_token_id": 77185,
|
5 |
+
"pad_token_id": 77185,
|
6 |
+
"transformers_version": "4.30.2"
|
7 |
+
}
|
pytorch_model-00001-of-00009.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a3ae777259524efea4b7da15a5dee99d1503a04db6a7604128bafbbd8b0f0d2b
|
3 |
+
size 9966449668
|
pytorch_model.bin.index.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
special_tokens_map.json
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": {
|
3 |
+
"content": "<s>",
|
4 |
+
"lstrip": false,
|
5 |
+
"normalized": true,
|
6 |
+
"rstrip": false,
|
7 |
+
"single_word": false
|
8 |
+
},
|
9 |
+
"eos_token": {
|
10 |
+
"content": "</s>",
|
11 |
+
"lstrip": false,
|
12 |
+
"normalized": true,
|
13 |
+
"rstrip": false,
|
14 |
+
"single_word": false
|
15 |
+
},
|
16 |
+
"unk_token": {
|
17 |
+
"content": "<unk>",
|
18 |
+
"lstrip": false,
|
19 |
+
"normalized": true,
|
20 |
+
"rstrip": false,
|
21 |
+
"single_word": false
|
22 |
+
}
|
23 |
+
}
|
tokenizer.model
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:36f79e0c70f73cdd2a8dd0fbe7bfe290da158eea746778d289e4ad76c8b383d9
|
3 |
+
size 2155861
|
tokenizer_config.json
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"add_bos_token": false,
|
3 |
+
"add_eos_token": false,
|
4 |
+
"bos_token": {
|
5 |
+
"__type": "AddedToken",
|
6 |
+
"content": "<s>",
|
7 |
+
"lstrip": false,
|
8 |
+
"normalized": true,
|
9 |
+
"rstrip": false,
|
10 |
+
"single_word": false
|
11 |
+
},
|
12 |
+
"clean_up_tokenization_spaces": false,
|
13 |
+
"eos_token": {
|
14 |
+
"__type": "AddedToken",
|
15 |
+
"content": "</s>",
|
16 |
+
"lstrip": false,
|
17 |
+
"normalized": true,
|
18 |
+
"rstrip": false,
|
19 |
+
"single_word": false
|
20 |
+
},
|
21 |
+
"model_max_length": 1000000000000000019884624838656,
|
22 |
+
"pad_token": null,
|
23 |
+
"sp_model_kwargs": {},
|
24 |
+
"tokenizer_class": "LlamaTokenizer",
|
25 |
+
"unk_token": {
|
26 |
+
"__type": "AddedToken",
|
27 |
+
"content": "<unk>",
|
28 |
+
"lstrip": false,
|
29 |
+
"normalized": true,
|
30 |
+
"rstrip": false,
|
31 |
+
"single_word": false
|
32 |
+
}
|
33 |
+
}
|
yuan_moe_hf_model.py
ADDED
@@ -0,0 +1,1454 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
5 |
+
# and OPT implementations in this library. It has been modified from its
|
6 |
+
# original forms to accommodate minor architectural differences compared
|
7 |
+
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
|
8 |
+
#
|
9 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
10 |
+
# you may not use this file except in compliance with the License.
|
11 |
+
# You may obtain a copy of the License at
|
12 |
+
#
|
13 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
14 |
+
#
|
15 |
+
# Unless required by applicable law or agreed to in writing, software
|
16 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
17 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
18 |
+
# See the License for the specific language governing permissions and
|
19 |
+
# limitations under the License.
|
20 |
+
""" PyTorch Yuan model."""
|
21 |
+
import math
|
22 |
+
from typing import List, Optional, Tuple, Union
|
23 |
+
import torch.nn.functional as F
|
24 |
+
import torch
|
25 |
+
import torch.utils.checkpoint
|
26 |
+
from torch import nn
|
27 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
28 |
+
from transformers.activations import ACT2FN
|
29 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
|
30 |
+
from transformers.modeling_utils import PreTrainedModel
|
31 |
+
from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
|
32 |
+
from configuration_yuan import YuanConfig
|
33 |
+
from einops import rearrange
|
34 |
+
#from flash_attn import flash_attn_varlen_func as flash_attn_unpadded_func
|
35 |
+
#from flash_attn import flash_attn_func
|
36 |
+
|
37 |
+
import copy
|
38 |
+
|
39 |
+
try:
|
40 |
+
from flash_attn import flash_attn_varlen_func as flash_attn_unpadded_func
|
41 |
+
from flash_attn import flash_attn_func
|
42 |
+
except ImportError:
|
43 |
+
flash_attn_unpadded_func = None
|
44 |
+
|
45 |
+
|
46 |
+
logger = logging.get_logger(__name__)
|
47 |
+
|
48 |
+
_CONFIG_FOR_DOC = "YuanConfig"
|
49 |
+
|
50 |
+
|
51 |
+
class LocalizedFiltering(torch.nn.Module):
|
52 |
+
"""
|
53 |
+
Mega's Exponential Moving Average layer, largely left unmodified from the original repo with the exception of
|
54 |
+
variable names and moving away from the stateful representation of incremental decoding state. See
|
55 |
+
"https://arxiv.org/abs/2209.10655" for more details.
|
56 |
+
"""
|
57 |
+
|
58 |
+
def __init__(self, hidden_size):
|
59 |
+
super().__init__()
|
60 |
+
|
61 |
+
self.embed_dim = hidden_size
|
62 |
+
self.lf_conv2d_group = 1
|
63 |
+
self.lf_conv2d_num_pad = 1
|
64 |
+
|
65 |
+
self.conv1 = torch.nn.Conv2d(self.embed_dim, self.embed_dim // 2, (2, 1), stride=(1, 1), padding=(self.lf_conv2d_num_pad, 0), groups=self.lf_conv2d_group)
|
66 |
+
self.conv2 = torch.nn.Conv2d(self.embed_dim // 2, self.embed_dim, (2, 1), stride=(1, 1), padding=(self.lf_conv2d_num_pad, 0), groups=self.lf_conv2d_group)
|
67 |
+
self.output_layernorm = YuanRMSNorm(self.embed_dim)
|
68 |
+
|
69 |
+
def _train_forward(self, inputs):
|
70 |
+
inputs = inputs.transpose(0,1)
|
71 |
+
seq_len, bsz, embed_dim = inputs.size()
|
72 |
+
if embed_dim != self.embed_dim:
|
73 |
+
raise ValueError(
|
74 |
+
f"Unexpected embedding dimension received: input is {embed_dim}, model expects {self.embed_dim}"
|
75 |
+
)
|
76 |
+
residual = inputs
|
77 |
+
|
78 |
+
inputs = inputs.view(seq_len, 1, bsz, embed_dim).permute(2, 3, 0, 1)
|
79 |
+
output1 = self.conv1(inputs)
|
80 |
+
output1 = output1[:, :, :seq_len, :]
|
81 |
+
|
82 |
+
output2 = self.conv2(output1)
|
83 |
+
output2 = output2[:, :, :seq_len, :].permute(2, 3, 0, 1).contiguous()
|
84 |
+
output2 = output2.view(seq_len, bsz, embed_dim)
|
85 |
+
assert output2.shape == residual.shape
|
86 |
+
|
87 |
+
lf_output = self.output_layernorm(output2 + residual)
|
88 |
+
lf_output = lf_output.transpose(0,1)
|
89 |
+
return lf_output
|
90 |
+
|
91 |
+
def _inference_forward(self, inputs, before_hidden_states):
|
92 |
+
|
93 |
+
if before_hidden_states is None:
|
94 |
+
inputs = inputs.transpose(0,1)
|
95 |
+
seq_len, bsz, embed_dim = inputs.size()
|
96 |
+
if embed_dim != self.embed_dim:
|
97 |
+
raise ValueError(
|
98 |
+
f"Unexpected embedding dimension received: input is {embed_dim}, model expects {self.embed_dim}"
|
99 |
+
)
|
100 |
+
residual = inputs
|
101 |
+
|
102 |
+
inputs = inputs.view(seq_len, 1, bsz, embed_dim).permute(2, 3, 0, 1)
|
103 |
+
output1 = self.conv1(inputs)
|
104 |
+
output1 = output1[:, :, :seq_len, :]
|
105 |
+
|
106 |
+
output2 = self.conv2(output1)
|
107 |
+
output2 = output2[:, :, :seq_len, :].permute(2, 3, 0, 1).contiguous()
|
108 |
+
output2 = output2.view(seq_len, bsz, embed_dim)
|
109 |
+
assert output2.shape == residual.shape
|
110 |
+
|
111 |
+
lf_output = self.output_layernorm(output2 + residual)
|
112 |
+
lf_output = lf_output.transpose(0,1)
|
113 |
+
return lf_output
|
114 |
+
else:
|
115 |
+
inputs = inputs.transpose(0,1)
|
116 |
+
before_hidden_states = before_hidden_states.transpose(0,1)
|
117 |
+
residual = inputs
|
118 |
+
|
119 |
+
seq_len, bsz, embed_dim = inputs.size()
|
120 |
+
seq_len_before, _, _ = before_hidden_states.size()
|
121 |
+
|
122 |
+
assert seq_len == 1 and seq_len_before == 2
|
123 |
+
|
124 |
+
inputs = torch.cat((before_hidden_states, inputs), dim=0)
|
125 |
+
inputs = inputs.view(3, 1, bsz, embed_dim).permute(2, 3, 0, 1)
|
126 |
+
|
127 |
+
output1 = self.conv1(inputs)
|
128 |
+
output2 = self.conv2(output1[:,:,1:-1,:])
|
129 |
+
output2 = output2[:,:,1:-1,:]
|
130 |
+
output2 = output2.view(1, bsz, embed_dim)
|
131 |
+
assert output2.shape == residual.shape
|
132 |
+
|
133 |
+
lf_output = self.output_layernorm(output2 + residual)
|
134 |
+
lf_output = lf_output.transpose(0,1)
|
135 |
+
|
136 |
+
return lf_output
|
137 |
+
|
138 |
+
|
139 |
+
|
140 |
+
def forward(
|
141 |
+
self,
|
142 |
+
inputs,
|
143 |
+
before_hidden_states
|
144 |
+
) -> torch.Tensor:
|
145 |
+
assert self.lf_conv2d_num_pad == 1
|
146 |
+
if self.training:
|
147 |
+
lf_output = self._train_forward(inputs)
|
148 |
+
else:
|
149 |
+
lf_output = self._inference_forward(inputs, before_hidden_states)
|
150 |
+
|
151 |
+
return lf_output
|
152 |
+
|
153 |
+
|
154 |
+
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
155 |
+
def _make_causal_mask(
|
156 |
+
input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
|
157 |
+
):
|
158 |
+
"""
|
159 |
+
Make causal mask used for bi-directional self-attention.
|
160 |
+
"""
|
161 |
+
bsz, tgt_len = input_ids_shape
|
162 |
+
mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
|
163 |
+
mask_cond = torch.arange(mask.size(-1), device=device)
|
164 |
+
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
|
165 |
+
mask = mask.to(dtype)
|
166 |
+
|
167 |
+
if past_key_values_length > 0:
|
168 |
+
mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
|
169 |
+
return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
|
170 |
+
|
171 |
+
|
172 |
+
# Copied from transformers.models.bart.modeling_bart._expand_mask
|
173 |
+
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
|
174 |
+
"""
|
175 |
+
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
|
176 |
+
"""
|
177 |
+
bsz, src_len = mask.size()
|
178 |
+
tgt_len = tgt_len if tgt_len is not None else src_len
|
179 |
+
|
180 |
+
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
|
181 |
+
|
182 |
+
inverted_mask = 1.0 - expanded_mask
|
183 |
+
|
184 |
+
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
|
185 |
+
|
186 |
+
|
187 |
+
def rotate_half(x):
|
188 |
+
"""Rotates half the hidden dims of the input."""
|
189 |
+
x1 = x[..., : x.shape[-1] // 2]
|
190 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
191 |
+
return torch.cat((-x2, x1), dim=-1)
|
192 |
+
|
193 |
+
def apply_rotary_pos_emb_0(q, k, cos, sin, position_ids):
|
194 |
+
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
|
195 |
+
rot_dim = sin.shape[-1]
|
196 |
+
|
197 |
+
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
|
198 |
+
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
|
199 |
+
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
|
200 |
+
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
|
201 |
+
|
202 |
+
q, q_pass = q[..., :rot_dim], q[..., rot_dim:]
|
203 |
+
k, k_pass = k[..., :rot_dim], k[..., rot_dim:]
|
204 |
+
|
205 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
206 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
207 |
+
|
208 |
+
return torch.cat((q_embed, q_pass), dim=-1), torch.cat((k_embed, k_pass), dim=-1)
|
209 |
+
|
210 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
211 |
+
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
|
212 |
+
#import pdb;pdb.set_trace()
|
213 |
+
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
|
214 |
+
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
|
215 |
+
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
|
216 |
+
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
|
217 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
218 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
219 |
+
return q_embed, k_embed
|
220 |
+
|
221 |
+
class YuanRMSNorm(nn.Module):
|
222 |
+
def __init__(self, hidden_size, eps=1e-6):
|
223 |
+
"""
|
224 |
+
YuanRMSNorm is equivalent to LlamaRMSNorm
|
225 |
+
"""
|
226 |
+
super().__init__()
|
227 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
228 |
+
self.variance_epsilon = eps
|
229 |
+
|
230 |
+
def forward(self, hidden_states):
|
231 |
+
input_dtype = hidden_states.dtype
|
232 |
+
hidden_states = hidden_states.to(torch.float32)
|
233 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
234 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
235 |
+
return self.weight * hidden_states.to(input_dtype)
|
236 |
+
|
237 |
+
class YuanRotaryEmbedding(torch.nn.Module):
|
238 |
+
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
239 |
+
|
240 |
+
"""
|
241 |
+
YuanRotaryEmbedding is equivalent to LlamaRotaryEmbedding in transformers v4.36
|
242 |
+
"""
|
243 |
+
|
244 |
+
super().__init__()
|
245 |
+
|
246 |
+
self.dim = dim
|
247 |
+
self.max_position_embeddings = max_position_embeddings
|
248 |
+
self.base = base
|
249 |
+
|
250 |
+
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
|
251 |
+
inv_freq = inv_freq.to(torch.bfloat16)
|
252 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
253 |
+
|
254 |
+
# Build here to make `torch.jit.trace` work.
|
255 |
+
self._set_cos_sin_cache(
|
256 |
+
seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
|
257 |
+
)
|
258 |
+
|
259 |
+
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
260 |
+
self.max_seq_len_cached = seq_len
|
261 |
+
t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
|
262 |
+
|
263 |
+
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
264 |
+
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
265 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
266 |
+
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
|
267 |
+
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
|
268 |
+
|
269 |
+
def forward(self, x, seq_len=None):
|
270 |
+
# x: [bs, num_attention_heads, seq_len, head_size]
|
271 |
+
if seq_len > self.max_seq_len_cached:
|
272 |
+
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
|
273 |
+
|
274 |
+
return (
|
275 |
+
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
|
276 |
+
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
|
277 |
+
)
|
278 |
+
|
279 |
+
# flash attn
|
280 |
+
class FlashSelfAttention(torch.nn.Module):
|
281 |
+
"""Implement the scaled dot product attention with softmax.
|
282 |
+
Arguments
|
283 |
+
---------
|
284 |
+
softmax_scale: The temperature to use for the softmax attention.
|
285 |
+
(default: 1/sqrt(d_keys) where d_keys is computed at
|
286 |
+
runtime)
|
287 |
+
attention_dropout: The dropout rate to apply to the attention
|
288 |
+
(default: 0.0)
|
289 |
+
"""
|
290 |
+
def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0,
|
291 |
+
device=None, dtype=None):
|
292 |
+
super().__init__()
|
293 |
+
assert flash_attn_unpadded_func is not None, ('Please install FlashAttention first, '
|
294 |
+
'e.g., with pip install flash-attn')
|
295 |
+
assert rearrange is not None, 'Please install einops first, e.g., with pip install einops'
|
296 |
+
self.causal = causal
|
297 |
+
self.softmax_scale = softmax_scale
|
298 |
+
self.dropout_p = attention_dropout
|
299 |
+
|
300 |
+
def forward(self, q, k, v):
|
301 |
+
"""Implements the multihead softmax attention.
|
302 |
+
Arguments
|
303 |
+
---------
|
304 |
+
q, k, v: The tensor containing the query, key, and value. (B, S, H, D)
|
305 |
+
"""
|
306 |
+
|
307 |
+
assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q,k,v)))
|
308 |
+
assert all((i.is_cuda for i in (q,k,v)))
|
309 |
+
|
310 |
+
batch_size, seqlen_q = q.shape[0], q.shape[1]
|
311 |
+
seqlen_k = k.shape[1]
|
312 |
+
|
313 |
+
q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [q, k, v]]
|
314 |
+
cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32,
|
315 |
+
device=q.device)
|
316 |
+
|
317 |
+
if self.training:
|
318 |
+
# during training q,k,v always have same seqlen
|
319 |
+
assert seqlen_k == seqlen_q
|
320 |
+
|
321 |
+
is_causal = self.causal
|
322 |
+
cu_seqlens_k = cu_seqlens_q
|
323 |
+
dropout_p = self.dropout_p
|
324 |
+
else:
|
325 |
+
# turn off FA causal mask after first inference autoregressive iteration
|
326 |
+
# only on first autoregressive step q,k,v have same seqlen
|
327 |
+
is_causal = seqlen_q == seqlen_k
|
328 |
+
cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32,
|
329 |
+
device=q.device)
|
330 |
+
dropout_p = 0
|
331 |
+
|
332 |
+
output = flash_attn_unpadded_func(
|
333 |
+
q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k,
|
334 |
+
dropout_p,
|
335 |
+
softmax_scale=self.softmax_scale, causal=is_causal
|
336 |
+
)
|
337 |
+
|
338 |
+
output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
|
339 |
+
return output
|
340 |
+
|
341 |
+
|
342 |
+
class ParallelAttention_router(nn.Module):
|
343 |
+
def __init__(self, config):
|
344 |
+
super(ParallelAttention_router, self).__init__()
|
345 |
+
layer_number=0
|
346 |
+
self.layer_number = max(1, layer_number)
|
347 |
+
|
348 |
+
|
349 |
+
self.flash_attn_drop = 0.01
|
350 |
+
self.hidden_size = config.hidden_size
|
351 |
+
self.projection_size = config.moe_config['moe_num_experts']
|
352 |
+
|
353 |
+
self.query = nn.Linear(self.hidden_size, self.projection_size, bias=False)
|
354 |
+
self.key = nn.Linear(self.hidden_size, self.projection_size, bias=False)
|
355 |
+
self.value = nn.Linear(self.hidden_size, self.projection_size, bias=False)
|
356 |
+
|
357 |
+
|
358 |
+
def forward(self, hidden_states, attention_mask=None, enc_position_ids=None,
|
359 |
+
encoder_output=None, inference_params=None,
|
360 |
+
rotary_pos_emb=None):
|
361 |
+
is_first_step = False
|
362 |
+
before_hidden_states = None
|
363 |
+
|
364 |
+
query_layer = self.query(hidden_states)
|
365 |
+
key_layer = self.key(hidden_states)
|
366 |
+
value_layer = self.value(hidden_states)
|
367 |
+
|
368 |
+
b = query_layer.size(0)
|
369 |
+
s = query_layer.size(1) # seq*batch = token_num
|
370 |
+
z = query_layer.size(2) # expert_num
|
371 |
+
|
372 |
+
# use fp32 router
|
373 |
+
query_layer = query_layer.float().view(b,s,z,1)
|
374 |
+
key_layer = key_layer.float().view(b,s,z,1)
|
375 |
+
value_layer = value_layer.float().view(b,s,z,1)
|
376 |
+
|
377 |
+
|
378 |
+
attn_weights = torch.matmul(query_layer, key_layer.transpose(2, 3))
|
379 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
|
380 |
+
|
381 |
+
attn_output = torch.matmul(attn_weights, value_layer)
|
382 |
+
|
383 |
+
router_output = attn_output.view(b*s, z)
|
384 |
+
|
385 |
+
return router_output
|
386 |
+
|
387 |
+
class YuanExpertMLP(nn.Module):
|
388 |
+
def __init__(self, config):
|
389 |
+
super(YuanExpertMLP, self).__init__()
|
390 |
+
|
391 |
+
self.gated_linear_unit = config.moe_config['gated_linear_unit']
|
392 |
+
self.ffn_hidden_size = config.moe_config['ffn_hidden_size']
|
393 |
+
|
394 |
+
|
395 |
+
if self.gated_linear_unit:
|
396 |
+
self.w1 = nn.Linear(config.hidden_size, self.ffn_hidden_size*2, bias=False)
|
397 |
+
|
398 |
+
|
399 |
+
else:
|
400 |
+
self.w1 = nn.Linear(config.hidden_size, self.ffn_hidden_size, bias=False)
|
401 |
+
|
402 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
403 |
+
self.w2 = nn.Linear(self.ffn_hidden_size, config.hidden_size, bias=False)
|
404 |
+
|
405 |
+
|
406 |
+
def forward(self, x):
|
407 |
+
x = self.w1(x)
|
408 |
+
if self.gated_linear_unit:
|
409 |
+
x = torch.chunk(x, 2, dim=-1)
|
410 |
+
x = self.act_fn(x[0]) * x[1]
|
411 |
+
else:
|
412 |
+
x = self.act_fn(x)
|
413 |
+
x = self.w2(x)
|
414 |
+
return x
|
415 |
+
|
416 |
+
|
417 |
+
|
418 |
+
class YuanMLP(nn.Module):
|
419 |
+
def __init__(
|
420 |
+
self,
|
421 |
+
hidden_size: int,
|
422 |
+
intermediate_size: int,
|
423 |
+
hidden_act: str
|
424 |
+
):
|
425 |
+
super().__init__()
|
426 |
+
self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
427 |
+
self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
428 |
+
self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
|
429 |
+
self.act_fn = ACT2FN[hidden_act]
|
430 |
+
|
431 |
+
def forward(self, x):
|
432 |
+
return self.down_proj(self.gate_proj(x) * self.act_fn(self.up_proj(x)))
|
433 |
+
|
434 |
+
|
435 |
+
class YuanAttention(nn.Module):
|
436 |
+
"""Localized Filtering-based Attention 'YUAN 2.0: A Large Language Model with Localized Filtering-based Attention' paper"""
|
437 |
+
|
438 |
+
def __init__(self, config: YuanConfig):
|
439 |
+
super().__init__()
|
440 |
+
self.config = config
|
441 |
+
self.hidden_size = config.hidden_size
|
442 |
+
self.num_heads = config.num_attention_heads
|
443 |
+
|
444 |
+
try:
|
445 |
+
self.attention_projection_size = config.attention_projection_size
|
446 |
+
except:
|
447 |
+
self.attention_projection_size = None
|
448 |
+
|
449 |
+
if self.attention_projection_size is None:
|
450 |
+
self.head_dim = self.hidden_size // self.num_heads
|
451 |
+
else:
|
452 |
+
self.head_dim = self.attention_projection_size // self.num_heads
|
453 |
+
|
454 |
+
self.max_position_embeddings = config.max_position_embeddings
|
455 |
+
self.causal_mask = config.causal_mask
|
456 |
+
self.softmax_scale = 1.0 / math.sqrt(self.head_dim)
|
457 |
+
self.use_flash_attention = config.use_flash_attention
|
458 |
+
try:
|
459 |
+
self.use_shareqk = config.use_shareqk
|
460 |
+
except Exception as e:
|
461 |
+
self.use_shareqk=False
|
462 |
+
self.dropout = 0.0
|
463 |
+
|
464 |
+
self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
|
465 |
+
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
|
466 |
+
|
467 |
+
if self.head_dim == self.hidden_size // self.num_heads:
|
468 |
+
self.rotary_emb = YuanRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
|
469 |
+
|
470 |
+
else:
|
471 |
+
self.rotary_emb = YuanRotaryEmbedding(self.hidden_size // self.num_heads, max_position_embeddings=self.max_position_embeddings)
|
472 |
+
|
473 |
+
if self.use_shareqk:
|
474 |
+
self.qk_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
|
475 |
+
self.qk_weight = nn.Parameter(torch.Tensor(2, self.hidden_size))
|
476 |
+
self.qk_bias = nn.Parameter(torch.Tensor(2, self.hidden_size))
|
477 |
+
else:
|
478 |
+
self.lf_gate = LocalizedFiltering(self.hidden_size)
|
479 |
+
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
|
480 |
+
self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
|
481 |
+
|
482 |
+
|
483 |
+
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
|
484 |
+
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
485 |
+
|
486 |
+
def forward(
|
487 |
+
self,
|
488 |
+
hidden_states: torch.Tensor,
|
489 |
+
attention_mask: Optional[torch.Tensor] = None,
|
490 |
+
position_ids: Optional[torch.LongTensor] = None,
|
491 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
492 |
+
output_attentions: bool = False,
|
493 |
+
use_cache: bool = False,
|
494 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
495 |
+
|
496 |
+
bsz, q_len, _ = hidden_states.size()
|
497 |
+
before_hidden_states = None
|
498 |
+
is_first_step = False
|
499 |
+
if use_cache:
|
500 |
+
if past_key_value is None:
|
501 |
+
inference_hidden_states_memory = torch.empty(bsz, 2, hidden_states.shape[2], dtype=hidden_states.dtype)
|
502 |
+
is_first_step = True
|
503 |
+
else:
|
504 |
+
before_hidden_states = past_key_value[2]
|
505 |
+
|
506 |
+
if use_cache:
|
507 |
+
if is_first_step:
|
508 |
+
if q_len >= 2:
|
509 |
+
inference_hidden_states_memory = hidden_states[ :, -2:, :]
|
510 |
+
else:
|
511 |
+
inference_hidden_states_memory[:, :, :] = 0
|
512 |
+
inference_hidden_states_memory[:, -1:, :] = hidden_states[:, -1:, :]
|
513 |
+
else:
|
514 |
+
hidden_states_tmp = before_hidden_states[:, -1:, :]
|
515 |
+
inference_hidden_states_memory = copy.deepcopy(torch.cat((hidden_states_tmp, hidden_states), dim=1))
|
516 |
+
|
517 |
+
value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
518 |
+
if self.use_shareqk:
|
519 |
+
qk_states = self.qk_proj(hidden_states).view(bsz, q_len, self.num_heads*self.head_dim)
|
520 |
+
query_key = qk_states.unsqueeze(2) * self.qk_weight + self.qk_bias
|
521 |
+
query_states, key_states = torch.unbind(query_key, dim=2)
|
522 |
+
|
523 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
524 |
+
key_states = key_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
525 |
+
else:
|
526 |
+
hidden_states = self.lf_gate(hidden_states,before_hidden_states)
|
527 |
+
query_states = self.q_proj(hidden_states)
|
528 |
+
key_states = self.k_proj(hidden_states)
|
529 |
+
qk_states = torch.cat([query_states, key_states], dim=-1)
|
530 |
+
qk_states = qk_states.view(bsz,q_len,self.num_heads,int(qk_states.shape[-1]//self.num_heads))
|
531 |
+
(query_states,key_states) = torch.chunk(qk_states, 2, dim=-1)
|
532 |
+
query_states = query_states.transpose(1, 2)
|
533 |
+
key_states = key_states.transpose(1, 2)
|
534 |
+
|
535 |
+
kv_seq_len = key_states.shape[-2]
|
536 |
+
if past_key_value is not None:
|
537 |
+
kv_seq_len += past_key_value[0].shape[-2]
|
538 |
+
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
539 |
+
|
540 |
+
query_states, key_states = apply_rotary_pos_emb_0(query_states, key_states, cos, sin, position_ids)
|
541 |
+
|
542 |
+
if past_key_value is not None:
|
543 |
+
# reuse k, v, self_attention
|
544 |
+
key_states = torch.cat([past_key_value[0], key_states], dim=2)
|
545 |
+
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
546 |
+
|
547 |
+
past_key_value = (key_states, value_states,inference_hidden_states_memory) if use_cache else None
|
548 |
+
if self.use_flash_attention:
|
549 |
+
attn_weights = None
|
550 |
+
query_states = query_states.transpose(1, 2)
|
551 |
+
key_states = key_states.transpose(1, 2)
|
552 |
+
value_states = value_states.transpose(1, 2)
|
553 |
+
|
554 |
+
batch_size, seqlen_q = query_states.shape[0], query_states.shape[1]
|
555 |
+
seqlen_k = key_states.shape[1]
|
556 |
+
|
557 |
+
q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [query_states, key_states, value_states]]
|
558 |
+
|
559 |
+
cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int,
|
560 |
+
device=q.device)
|
561 |
+
|
562 |
+
if self.training:
|
563 |
+
assert seqlen_k == seqlen_q
|
564 |
+
cu_seqlens_k = cu_seqlens_q
|
565 |
+
is_causal = self.causal_mask
|
566 |
+
else:
|
567 |
+
is_causal = seqlen_q == seqlen_k
|
568 |
+
cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int,
|
569 |
+
device=q.device)
|
570 |
+
self.dropout=0
|
571 |
+
|
572 |
+
output = flash_attn_unpadded_func(
|
573 |
+
q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k, self.dropout, causal=is_causal
|
574 |
+
)
|
575 |
+
|
576 |
+
attn_output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
|
577 |
+
else:
|
578 |
+
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
579 |
+
|
580 |
+
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
581 |
+
raise ValueError(
|
582 |
+
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
|
583 |
+
f" {attn_weights.size()}"
|
584 |
+
)
|
585 |
+
if attention_mask is not None:
|
586 |
+
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
|
587 |
+
raise ValueError(
|
588 |
+
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
589 |
+
)
|
590 |
+
attn_weights = attn_weights + attention_mask
|
591 |
+
attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
|
592 |
+
|
593 |
+
# upcast attention to fp32
|
594 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
595 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
596 |
+
|
597 |
+
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
598 |
+
raise ValueError(
|
599 |
+
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
|
600 |
+
f" {attn_output.size()}"
|
601 |
+
)
|
602 |
+
|
603 |
+
attn_output = attn_output.transpose(1, 2)
|
604 |
+
|
605 |
+
if self.attention_projection_size is None:
|
606 |
+
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
607 |
+
else:
|
608 |
+
attn_output = attn_output.reshape(bsz, q_len, self.attention_projection_size)
|
609 |
+
|
610 |
+
attn_output = self.o_proj(attn_output)
|
611 |
+
|
612 |
+
if not output_attentions:
|
613 |
+
attn_weights = None
|
614 |
+
return attn_output, attn_weights, past_key_value
|
615 |
+
|
616 |
+
|
617 |
+
|
618 |
+
class YuanMoeLayer(nn.Module):
|
619 |
+
def __init__(self, config):
|
620 |
+
super().__init__()
|
621 |
+
self.num_experts = config.moe_config['moe_num_experts']
|
622 |
+
self.top_k = config.moe_config['moe_top_k']
|
623 |
+
self.norm_topk_prob = config.moe_config['norm_topk_prob']
|
624 |
+
self.hidden_size = config.hidden_size
|
625 |
+
|
626 |
+
|
627 |
+
self.gate = ParallelAttention_router(config)
|
628 |
+
self.experts = nn.ModuleList(
|
629 |
+
[YuanExpertMLP(config) for _ in range(self.num_experts)]
|
630 |
+
)
|
631 |
+
|
632 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
633 |
+
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
634 |
+
|
635 |
+
# router_logits: (batch * sequence_length, n_experts)
|
636 |
+
router_logits = self.gate(hidden_states)
|
637 |
+
hidden_states = hidden_states.view(-1, hidden_dim)
|
638 |
+
|
639 |
+
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
|
640 |
+
routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
|
641 |
+
if self.norm_topk_prob:
|
642 |
+
routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
|
643 |
+
# we cast back to the input dtype
|
644 |
+
routing_weights = routing_weights.to(hidden_states.dtype)
|
645 |
+
|
646 |
+
final_hidden_states = torch.zeros(
|
647 |
+
(batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
|
648 |
+
)
|
649 |
+
|
650 |
+
# One hot encode the selected experts to create an expert mask
|
651 |
+
# this will be used to easily index which expert is going to be sollicitated
|
652 |
+
expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
|
653 |
+
|
654 |
+
# Loop over all available experts in the model and perform the computation on each expert
|
655 |
+
for expert_idx in range(self.num_experts):
|
656 |
+
expert_layer = self.experts[expert_idx]
|
657 |
+
idx, top_x = torch.where(expert_mask[expert_idx])
|
658 |
+
|
659 |
+
if top_x.shape[0] == 0:
|
660 |
+
continue
|
661 |
+
|
662 |
+
# in torch it is faster to index using lists than torch tensors
|
663 |
+
top_x_list = top_x.tolist()
|
664 |
+
idx_list = idx.tolist()
|
665 |
+
|
666 |
+
# Index the correct hidden states and compute the expert hidden state for
|
667 |
+
# the current expert. We need to make sure to multiply the output hidden
|
668 |
+
# states by `routing_weights` on the corresponding tokens (top-1 and top-2)
|
669 |
+
current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
|
670 |
+
current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]
|
671 |
+
|
672 |
+
# However `index_add_` only support torch tensors for indexing so we'll use
|
673 |
+
# the `top_x` tensor here.
|
674 |
+
final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
|
675 |
+
|
676 |
+
final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
|
677 |
+
return final_hidden_states, router_logits
|
678 |
+
|
679 |
+
|
680 |
+
class YuanDecoderLayer(nn.Module):
|
681 |
+
def __init__(self, config: YuanConfig):
|
682 |
+
super().__init__()
|
683 |
+
self.hidden_size = config.hidden_size
|
684 |
+
self.self_attn = YuanAttention(config=config)
|
685 |
+
|
686 |
+
if config.moe_config['moe_num_experts'] > 0:
|
687 |
+
self.mlp = YuanMoeLayer(config)
|
688 |
+
else:
|
689 |
+
self.mlp = YuanMLP(
|
690 |
+
hidden_size=self.hidden_size,
|
691 |
+
intermediate_size=config.intermediate_size,
|
692 |
+
hidden_act=config.hidden_act,
|
693 |
+
)
|
694 |
+
|
695 |
+
|
696 |
+
self.input_layernorm = YuanRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
697 |
+
self.post_attention_layernorm = YuanRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
698 |
+
|
699 |
+
def forward(
|
700 |
+
self,
|
701 |
+
hidden_states: torch.Tensor,
|
702 |
+
attention_mask: Optional[torch.Tensor] = None,
|
703 |
+
position_ids: Optional[torch.LongTensor] = None,
|
704 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
705 |
+
output_attentions: Optional[bool] = False,
|
706 |
+
use_cache: Optional[bool] = False,
|
707 |
+
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
708 |
+
"""
|
709 |
+
Args:
|
710 |
+
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
711 |
+
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
712 |
+
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
|
713 |
+
output_attentions (`bool`, *optional*):
|
714 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
715 |
+
returned tensors for more detail.
|
716 |
+
use_cache (`bool`, *optional*):
|
717 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
718 |
+
(see `past_key_values`).
|
719 |
+
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
|
720 |
+
"""
|
721 |
+
residual = hidden_states
|
722 |
+
hidden_states = self.input_layernorm(hidden_states)
|
723 |
+
|
724 |
+
# Self Attention
|
725 |
+
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
726 |
+
hidden_states=hidden_states,
|
727 |
+
attention_mask=attention_mask,
|
728 |
+
position_ids=position_ids,
|
729 |
+
past_key_value=past_key_value,
|
730 |
+
output_attentions=output_attentions,
|
731 |
+
use_cache=use_cache,
|
732 |
+
)
|
733 |
+
|
734 |
+
hidden_states = residual + hidden_states
|
735 |
+
|
736 |
+
# Fully Connected
|
737 |
+
residual = hidden_states
|
738 |
+
|
739 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
740 |
+
|
741 |
+
hidden_states, router_logits = self.mlp(hidden_states)
|
742 |
+
|
743 |
+
hidden_states = residual + hidden_states
|
744 |
+
|
745 |
+
outputs = (hidden_states,)
|
746 |
+
|
747 |
+
if output_attentions:
|
748 |
+
outputs += (self_attn_weights,)
|
749 |
+
|
750 |
+
if use_cache:
|
751 |
+
outputs += (present_key_value,)
|
752 |
+
|
753 |
+
return outputs
|
754 |
+
|
755 |
+
|
756 |
+
YUAN_START_DOCSTRING = r"""
|
757 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
758 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
759 |
+
etc.)
|
760 |
+
|
761 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
762 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
763 |
+
and behavior.
|
764 |
+
|
765 |
+
Parameters:
|
766 |
+
config ([`YuanConfig`]):
|
767 |
+
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
768 |
+
load the weights associated with the model, only the configuration. Check out the
|
769 |
+
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
770 |
+
"""
|
771 |
+
|
772 |
+
|
773 |
+
@add_start_docstrings(
|
774 |
+
"The bare Yuan Model outputting raw hidden-states without any specific head on top.",
|
775 |
+
YUAN_START_DOCSTRING,
|
776 |
+
)
|
777 |
+
class YuanPreTrainedModel(PreTrainedModel):
|
778 |
+
config_class = YuanConfig
|
779 |
+
base_model_prefix = "model"
|
780 |
+
supports_gradient_checkpointing = True
|
781 |
+
_no_split_modules = ["YuanDecoderLayer"]
|
782 |
+
_skip_keys_device_placement = "past_key_values"
|
783 |
+
_keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
|
784 |
+
|
785 |
+
def _init_weights(self, module):
|
786 |
+
std = self.config.initializer_range
|
787 |
+
if isinstance(module, nn.Linear):
|
788 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
789 |
+
if module.bias is not None:
|
790 |
+
module.bias.data.zero_()
|
791 |
+
elif isinstance(module, nn.Embedding):
|
792 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
793 |
+
if module.padding_idx is not None:
|
794 |
+
module.weight.data[module.padding_idx].zero_()
|
795 |
+
|
796 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
797 |
+
if isinstance(module, YuanModel):
|
798 |
+
module.gradient_checkpointing = value
|
799 |
+
|
800 |
+
|
801 |
+
YUAN_INPUTS_DOCSTRING = r"""
|
802 |
+
Args:
|
803 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
804 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
805 |
+
it.
|
806 |
+
|
807 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
808 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
809 |
+
|
810 |
+
[What are input IDs?](../glossary#input-ids)
|
811 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
812 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
813 |
+
|
814 |
+
- 1 for tokens that are **not masked**,
|
815 |
+
- 0 for tokens that are **masked**.
|
816 |
+
|
817 |
+
[What are attention masks?](../glossary#attention-mask)
|
818 |
+
|
819 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
820 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
821 |
+
|
822 |
+
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
823 |
+
`past_key_values`).
|
824 |
+
|
825 |
+
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
826 |
+
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
827 |
+
information on the default strategy.
|
828 |
+
|
829 |
+
- 1 indicates the head is **not masked**,
|
830 |
+
- 0 indicates the head is **masked**.
|
831 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
832 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
833 |
+
config.n_positions - 1]`.
|
834 |
+
|
835 |
+
[What are position IDs?](../glossary#position-ids)
|
836 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
837 |
+
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
838 |
+
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
|
839 |
+
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
|
840 |
+
|
841 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
842 |
+
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
843 |
+
|
844 |
+
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
845 |
+
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
846 |
+
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
847 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
848 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
849 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
850 |
+
model's internal embedding lookup matrix.
|
851 |
+
use_cache (`bool`, *optional*):
|
852 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
853 |
+
`past_key_values`).
|
854 |
+
output_attentions (`bool`, *optional*):
|
855 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
856 |
+
tensors for more detail.
|
857 |
+
output_hidden_states (`bool`, *optional*):
|
858 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
859 |
+
more detail.
|
860 |
+
return_dict (`bool`, *optional*):
|
861 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
862 |
+
"""
|
863 |
+
|
864 |
+
|
865 |
+
@add_start_docstrings(
|
866 |
+
"The bare Yuan Model outputting raw hidden-states without any specific head on top.",
|
867 |
+
YUAN_START_DOCSTRING,
|
868 |
+
)
|
869 |
+
class YuanModel(YuanPreTrainedModel):
|
870 |
+
"""
|
871 |
+
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`YuanDecoderLayer`]
|
872 |
+
|
873 |
+
Args:
|
874 |
+
config: YuanConfig
|
875 |
+
"""
|
876 |
+
|
877 |
+
def __init__(self, config: YuanConfig):
|
878 |
+
super().__init__(config)
|
879 |
+
self.padding_idx = config.pad_token_id
|
880 |
+
self.vocab_size = config.vocab_size
|
881 |
+
|
882 |
+
#TODO: control it by config
|
883 |
+
self.eod_token = config.eod_token
|
884 |
+
self.reset_attention_mask = config.reset_attention_mask
|
885 |
+
self.reset_position_ids = config.reset_position_ids
|
886 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
887 |
+
self.layers = nn.ModuleList([YuanDecoderLayer(config) for _ in range(config.num_hidden_layers)])
|
888 |
+
self.norm = YuanRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
889 |
+
self.gradient_checkpointing = False
|
890 |
+
# Initialize weights and apply final processing
|
891 |
+
self.post_init()
|
892 |
+
|
893 |
+
def get_input_embeddings(self):
|
894 |
+
return self.embed_tokens
|
895 |
+
|
896 |
+
def set_input_embeddings(self, value):
|
897 |
+
self.embed_tokens = value
|
898 |
+
|
899 |
+
# Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
|
900 |
+
def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
|
901 |
+
# create causal mask
|
902 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
903 |
+
combined_attention_mask = None
|
904 |
+
if input_shape[-1] > 1:
|
905 |
+
combined_attention_mask = _make_causal_mask(
|
906 |
+
input_shape,
|
907 |
+
inputs_embeds.dtype,
|
908 |
+
device=inputs_embeds.device,
|
909 |
+
past_key_values_length=past_key_values_length,
|
910 |
+
)
|
911 |
+
|
912 |
+
if attention_mask is not None:
|
913 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
914 |
+
expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
|
915 |
+
inputs_embeds.device
|
916 |
+
)
|
917 |
+
combined_attention_mask = (
|
918 |
+
expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
|
919 |
+
)
|
920 |
+
|
921 |
+
return combined_attention_mask
|
922 |
+
|
923 |
+
def _prepare_decoder_attention_mask_training(self, input_id, inputs_embeds, eod_token, reset_mask_flag ,reset_attention_mask=True, reset_position_ids=True):
|
924 |
+
|
925 |
+
micro_batch_size, seq_length = input_id.size()
|
926 |
+
|
927 |
+
attention_mask = torch.tril(torch.ones(
|
928 |
+
(micro_batch_size, seq_length, seq_length), device=inputs_embeds.device)).view(
|
929 |
+
micro_batch_size, 1, seq_length, seq_length)
|
930 |
+
|
931 |
+
position_ids = torch.arange(seq_length, dtype=torch.long,
|
932 |
+
device=inputs_embeds.device)
|
933 |
+
position_ids = position_ids.unsqueeze(0).expand_as(input_id)
|
934 |
+
|
935 |
+
if reset_position_ids:
|
936 |
+
position_ids = position_ids.clone()
|
937 |
+
|
938 |
+
if reset_position_ids or reset_attention_mask:
|
939 |
+
# Loop through the batches:
|
940 |
+
for b in range(micro_batch_size):
|
941 |
+
|
942 |
+
# Find indecies where EOD token is.
|
943 |
+
eod_index = position_ids[b, input_id[b] == eod_token]
|
944 |
+
|
945 |
+
# Detach indecies from positions if going to modify positions.
|
946 |
+
if reset_position_ids:
|
947 |
+
eod_index = eod_index.clone()
|
948 |
+
# Loop through EOD indecies:
|
949 |
+
prev_index = 0
|
950 |
+
for j in range(eod_index.size()[0]):
|
951 |
+
i = eod_index[j]
|
952 |
+
# Mask attention loss.
|
953 |
+
if reset_attention_mask:
|
954 |
+
attention_mask[b, 0, (i + 1):, :(i + 1)] = 0
|
955 |
+
# Reset positions.
|
956 |
+
if reset_position_ids:
|
957 |
+
position_ids[b, (i + 1):] -= (i + 1 - prev_index)
|
958 |
+
prev_index = i + 1
|
959 |
+
|
960 |
+
inverted_mask = 1 - attention_mask
|
961 |
+
output_attn_mask = inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(inputs_embeds.dtype).min)
|
962 |
+
if reset_mask_flag:
|
963 |
+
output_attn_mask = output_attn_mask[:,:,-1:,:]
|
964 |
+
return output_attn_mask, position_ids
|
965 |
+
|
966 |
+
@add_start_docstrings_to_model_forward(YUAN_INPUTS_DOCSTRING)
|
967 |
+
def forward(
|
968 |
+
self,
|
969 |
+
input_ids: torch.LongTensor = None,
|
970 |
+
attention_mask: Optional[torch.Tensor] = None,
|
971 |
+
position_ids: Optional[torch.LongTensor] = None,
|
972 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
973 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
974 |
+
use_cache: Optional[bool] = None,
|
975 |
+
output_attentions: Optional[bool] = None,
|
976 |
+
output_hidden_states: Optional[bool] = None,
|
977 |
+
output_router_logits: Optional[bool] = None,
|
978 |
+
return_dict: Optional[bool] = None,
|
979 |
+
) -> Union[Tuple, BaseModelOutputWithPast]:
|
980 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
981 |
+
output_router_logits = (
|
982 |
+
output_router_logits if output_router_logits is not None else self.config.output_router_logits
|
983 |
+
)
|
984 |
+
output_hidden_states = (
|
985 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
986 |
+
)
|
987 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
988 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
989 |
+
input_ids1 = copy.deepcopy(input_ids)
|
990 |
+
reset_mask_flag = False
|
991 |
+
if past_key_values:
|
992 |
+
input_ids = input_ids[:, -1:]
|
993 |
+
if use_cache:
|
994 |
+
reset_mask_flag = True
|
995 |
+
# retrieve input_ids and inputs_embeds
|
996 |
+
if input_ids is not None and inputs_embeds is not None:
|
997 |
+
raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
|
998 |
+
elif input_ids is not None:
|
999 |
+
|
1000 |
+
batch_size, seq_length = input_ids.shape
|
1001 |
+
elif inputs_embeds is not None:
|
1002 |
+
batch_size, seq_length, _ = inputs_embeds.shape
|
1003 |
+
else:
|
1004 |
+
raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
|
1005 |
+
|
1006 |
+
seq_length_with_past = seq_length
|
1007 |
+
past_key_values_length = 0
|
1008 |
+
|
1009 |
+
if past_key_values is not None:
|
1010 |
+
past_key_values_length = past_key_values[0][0].shape[2]
|
1011 |
+
seq_length_with_past = seq_length_with_past + past_key_values_length
|
1012 |
+
|
1013 |
+
if position_ids is None:
|
1014 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
1015 |
+
position_ids = torch.arange(
|
1016 |
+
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
|
1017 |
+
)
|
1018 |
+
position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
|
1019 |
+
else:
|
1020 |
+
position_ids = position_ids.view(-1, seq_length).long()
|
1021 |
+
|
1022 |
+
if inputs_embeds is None:
|
1023 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
1024 |
+
if self.training or self.reset_position_ids:
|
1025 |
+
attention_mask, _ = self._prepare_decoder_attention_mask_training(input_ids1, inputs_embeds, self.eod_token, reset_mask_flag, self.reset_attention_mask, self.reset_position_ids)
|
1026 |
+
|
1027 |
+
else:
|
1028 |
+
if attention_mask is None:
|
1029 |
+
attention_mask = torch.ones(
|
1030 |
+
(batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
|
1031 |
+
)
|
1032 |
+
attention_mask = self._prepare_decoder_attention_mask(
|
1033 |
+
attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
|
1034 |
+
)
|
1035 |
+
|
1036 |
+
hidden_states = inputs_embeds
|
1037 |
+
|
1038 |
+
if self.gradient_checkpointing and self.training:
|
1039 |
+
if use_cache:
|
1040 |
+
logger.warning_once(
|
1041 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
1042 |
+
)
|
1043 |
+
use_cache = False
|
1044 |
+
|
1045 |
+
# decoder layers
|
1046 |
+
all_hidden_states = () if output_hidden_states else None
|
1047 |
+
all_self_attns = () if output_attentions else None
|
1048 |
+
next_decoder_cache = () if use_cache else None
|
1049 |
+
|
1050 |
+
for idx, decoder_layer in enumerate(self.layers):
|
1051 |
+
if output_hidden_states:
|
1052 |
+
all_hidden_states += (hidden_states,)
|
1053 |
+
|
1054 |
+
past_key_value = past_key_values[idx] if past_key_values is not None else None
|
1055 |
+
|
1056 |
+
if self.gradient_checkpointing and self.training:
|
1057 |
+
|
1058 |
+
def create_custom_forward(module):
|
1059 |
+
def custom_forward(*inputs):
|
1060 |
+
# None for past_key_value
|
1061 |
+
return module(*inputs, output_attentions, None)
|
1062 |
+
|
1063 |
+
return custom_forward
|
1064 |
+
|
1065 |
+
layer_outputs = torch.utils.checkpoint.checkpoint(
|
1066 |
+
create_custom_forward(decoder_layer),
|
1067 |
+
hidden_states,
|
1068 |
+
attention_mask,
|
1069 |
+
position_ids,
|
1070 |
+
None,
|
1071 |
+
)
|
1072 |
+
else:
|
1073 |
+
layer_outputs = decoder_layer(
|
1074 |
+
hidden_states,
|
1075 |
+
attention_mask=attention_mask,
|
1076 |
+
position_ids=position_ids,
|
1077 |
+
past_key_value=past_key_value,
|
1078 |
+
output_attentions=output_attentions,
|
1079 |
+
use_cache=use_cache,
|
1080 |
+
)
|
1081 |
+
|
1082 |
+
hidden_states = layer_outputs[0]
|
1083 |
+
|
1084 |
+
if use_cache:
|
1085 |
+
next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
|
1086 |
+
|
1087 |
+
if output_attentions:
|
1088 |
+
all_self_attns += (layer_outputs[1],)
|
1089 |
+
hidden_states = self.norm(hidden_states)
|
1090 |
+
|
1091 |
+
# add hidden states from the last decoder layer
|
1092 |
+
if output_hidden_states:
|
1093 |
+
all_hidden_states += (hidden_states,)
|
1094 |
+
next_cache = next_decoder_cache if use_cache else None
|
1095 |
+
if not return_dict:
|
1096 |
+
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
1097 |
+
return BaseModelOutputWithPast(
|
1098 |
+
last_hidden_state=hidden_states,
|
1099 |
+
past_key_values=next_cache,
|
1100 |
+
hidden_states=all_hidden_states,
|
1101 |
+
attentions=all_self_attns,
|
1102 |
+
)
|
1103 |
+
|
1104 |
+
|
1105 |
+
class YuanForCausalLM(YuanPreTrainedModel):
|
1106 |
+
def __init__(self, config):
|
1107 |
+
super().__init__(config)
|
1108 |
+
self.eod_token = config.eod_token
|
1109 |
+
self.sep_token = config.sep_token
|
1110 |
+
self.use_loss_mask = config.use_loss_mask
|
1111 |
+
self.model = YuanModel(config)
|
1112 |
+
|
1113 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
1114 |
+
|
1115 |
+
# Initialize weights and apply final processing
|
1116 |
+
self.post_init()
|
1117 |
+
|
1118 |
+
def get_input_embeddings(self):
|
1119 |
+
return self.model.embed_tokens
|
1120 |
+
|
1121 |
+
def set_input_embeddings(self, value):
|
1122 |
+
self.model.embed_tokens = value
|
1123 |
+
|
1124 |
+
def get_output_embeddings(self):
|
1125 |
+
return self.lm_head
|
1126 |
+
|
1127 |
+
def set_output_embeddings(self, new_embeddings):
|
1128 |
+
self.lm_head = new_embeddings
|
1129 |
+
|
1130 |
+
def set_decoder(self, decoder):
|
1131 |
+
self.model = decoder
|
1132 |
+
|
1133 |
+
def get_decoder(self):
|
1134 |
+
return self.model
|
1135 |
+
|
1136 |
+
def get_loss_mask(self, input_ids, labels, eod_token, sep_token):
|
1137 |
+
micro_batch_size, seq_length = input_ids.size()
|
1138 |
+
loss_mask = torch.ones(input_ids.size(), dtype=torch.float, device=input_ids.device)
|
1139 |
+
|
1140 |
+
position_ids = torch.arange(seq_length, dtype=torch.long,
|
1141 |
+
device=input_ids.device)
|
1142 |
+
position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
|
1143 |
+
|
1144 |
+
|
1145 |
+
"""modify loss_mask to only calculate the loss of the answer (separated with [SEP])"""
|
1146 |
+
|
1147 |
+
for b in range(micro_batch_size):
|
1148 |
+
eod_indexs = position_ids[b, input_ids[b] == eod_token]
|
1149 |
+
sep_indexs = position_ids[b, input_ids[b] == sep_token]
|
1150 |
+
|
1151 |
+
if len(eod_indexs) == 0 or len(sep_indexs) == 0:
|
1152 |
+
loss_mask[b] = 1.0
|
1153 |
+
else:
|
1154 |
+
if eod_indexs[0] > sep_indexs[0]:
|
1155 |
+
loss_mask[b, 0:sep_indexs[0]] = 0
|
1156 |
+
|
1157 |
+
if len(eod_indexs) == len(sep_indexs):
|
1158 |
+
for ii, eod_index in enumerate(eod_indexs):
|
1159 |
+
start_index = eod_index
|
1160 |
+
if ii == (len(sep_indexs) - 1):
|
1161 |
+
stop_index = seq_length
|
1162 |
+
else:
|
1163 |
+
stop_index = sep_indexs[ii + 1]
|
1164 |
+
loss_mask[b, start_index:stop_index] = 0.0
|
1165 |
+
else:
|
1166 |
+
if len(eod_indexs) > len(sep_indexs):
|
1167 |
+
loss_mask[b,:] = 1.0
|
1168 |
+
else:
|
1169 |
+
for ii, eod_index in enumerate(eod_indexs):
|
1170 |
+
start_index = eod_index
|
1171 |
+
stop_index = sep_indexs[ii + 1]
|
1172 |
+
|
1173 |
+
loss_mask[b, start_index:stop_index] = 0.0
|
1174 |
+
|
1175 |
+
elif eod_indexs[0] < sep_indexs[0]:
|
1176 |
+
|
1177 |
+
if len(eod_indexs) == len(sep_indexs):
|
1178 |
+
for ii, eod_index in enumerate(eod_indexs):
|
1179 |
+
start_index = eod_index
|
1180 |
+
stop_index = sep_indexs[ii]
|
1181 |
+
loss_mask[b, start_index:stop_index] = 0.0
|
1182 |
+
|
1183 |
+
else:
|
1184 |
+
if len(eod_indexs) < len(sep_indexs):
|
1185 |
+
loss_mask[b,:] = 1.0
|
1186 |
+
else:
|
1187 |
+
for ii, eod_index in enumerate(eod_indexs):
|
1188 |
+
start_index = eod_index
|
1189 |
+
if ii >= len(sep_indexs):
|
1190 |
+
stop_index = seq_length
|
1191 |
+
else:
|
1192 |
+
stop_index = sep_indexs[ii]
|
1193 |
+
loss_mask[b, start_index:stop_index] = 0.0
|
1194 |
+
|
1195 |
+
loss_mask[input_ids == eod_token] = 1.0
|
1196 |
+
return loss_mask
|
1197 |
+
@add_start_docstrings_to_model_forward(YUAN_INPUTS_DOCSTRING)
|
1198 |
+
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
1199 |
+
def forward(
|
1200 |
+
self,
|
1201 |
+
input_ids: torch.LongTensor = None,
|
1202 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1203 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1204 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1205 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1206 |
+
labels: Optional[torch.LongTensor] = None,
|
1207 |
+
use_cache: Optional[bool] = None,
|
1208 |
+
output_attentions: Optional[bool] = None,
|
1209 |
+
output_hidden_states: Optional[bool] = None,
|
1210 |
+
output_router_logits: Optional[bool] = None,
|
1211 |
+
return_dict: Optional[bool] = None,
|
1212 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
1213 |
+
r"""
|
1214 |
+
Args:
|
1215 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1216 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
1217 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
1218 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
1219 |
+
|
1220 |
+
Returns:
|
1221 |
+
|
1222 |
+
Example:
|
1223 |
+
|
1224 |
+
```python
|
1225 |
+
>>> from transformers import AutoTokenizer, YuanForCausalLM
|
1226 |
+
|
1227 |
+
>>> model = YuanForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
|
1228 |
+
>>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
|
1229 |
+
|
1230 |
+
>>> prompt = "Hey, are you consciours? Can you talk to me?"
|
1231 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
1232 |
+
|
1233 |
+
>>> # Generate
|
1234 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
1235 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
1236 |
+
"Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
|
1237 |
+
```"""
|
1238 |
+
|
1239 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1240 |
+
|
1241 |
+
output_hidden_states = (
|
1242 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1243 |
+
)
|
1244 |
+
|
1245 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1246 |
+
|
1247 |
+
outputs = self.model(
|
1248 |
+
input_ids=input_ids,
|
1249 |
+
attention_mask=attention_mask,
|
1250 |
+
position_ids=position_ids,
|
1251 |
+
past_key_values=past_key_values,
|
1252 |
+
inputs_embeds=inputs_embeds,
|
1253 |
+
use_cache=use_cache,
|
1254 |
+
output_attentions=output_attentions,
|
1255 |
+
output_hidden_states=output_hidden_states,
|
1256 |
+
return_dict=return_dict,
|
1257 |
+
)
|
1258 |
+
|
1259 |
+
hidden_states = outputs[0]
|
1260 |
+
|
1261 |
+
logits = self.lm_head(hidden_states)
|
1262 |
+
loss = None
|
1263 |
+
if labels is not None:
|
1264 |
+
if self.use_loss_mask:
|
1265 |
+
loss_mask = self.get_loss_mask(input_ids, labels, self.eod_token, self.sep_token)
|
1266 |
+
# Shift so that tokens < n predict n
|
1267 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
1268 |
+
shift_labels = labels[..., 1:].contiguous()
|
1269 |
+
# Flatten the tokens
|
1270 |
+
if self.use_loss_mask:
|
1271 |
+
loss_fct = CrossEntropyLoss(reduction='none')
|
1272 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
1273 |
+
shift_labels = shift_labels.view(-1)
|
1274 |
+
# Enable model parallelism
|
1275 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
1276 |
+
loss = loss_fct(shift_logits, shift_labels)
|
1277 |
+
loss = torch.sum(loss * loss_mask) / loss_mask.sum()
|
1278 |
+
else:
|
1279 |
+
loss_fct = CrossEntropyLoss()
|
1280 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
1281 |
+
shift_labels = shift_labels.view(-1)
|
1282 |
+
# Enable model parallelism
|
1283 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
1284 |
+
loss = loss_fct(shift_logits, shift_labels)
|
1285 |
+
if not return_dict:
|
1286 |
+
output = (logits,) + outputs[1:]
|
1287 |
+
return (loss,) + output if loss is not None else output
|
1288 |
+
|
1289 |
+
return CausalLMOutputWithPast(
|
1290 |
+
loss=loss,
|
1291 |
+
logits=logits,
|
1292 |
+
past_key_values=outputs.past_key_values,
|
1293 |
+
hidden_states=hidden_states,
|
1294 |
+
attentions=outputs.attentions,
|
1295 |
+
)
|
1296 |
+
|
1297 |
+
def prepare_inputs_for_generation(
|
1298 |
+
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
1299 |
+
):
|
1300 |
+
|
1301 |
+
position_ids = kwargs.get("position_ids", None)
|
1302 |
+
if attention_mask is not None and position_ids is None:
|
1303 |
+
# create position_ids on the fly for batch generation
|
1304 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
1305 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
1306 |
+
if past_key_values:
|
1307 |
+
position_ids = position_ids[:, -1].unsqueeze(-1)
|
1308 |
+
|
1309 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
1310 |
+
if inputs_embeds is not None and past_key_values is None:
|
1311 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
1312 |
+
else:
|
1313 |
+
model_inputs = {"input_ids": input_ids}
|
1314 |
+
|
1315 |
+
model_inputs.update(
|
1316 |
+
{
|
1317 |
+
"position_ids": position_ids,
|
1318 |
+
"past_key_values": past_key_values,
|
1319 |
+
"use_cache": kwargs.get("use_cache"),
|
1320 |
+
"attention_mask": attention_mask,
|
1321 |
+
}
|
1322 |
+
)
|
1323 |
+
return model_inputs
|
1324 |
+
|
1325 |
+
@staticmethod
|
1326 |
+
def _reorder_cache(past_key_values, beam_idx):
|
1327 |
+
reordered_past = ()
|
1328 |
+
for layer_past in past_key_values:
|
1329 |
+
reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
|
1330 |
+
return reordered_past
|
1331 |
+
|
1332 |
+
|
1333 |
+
@add_start_docstrings(
|
1334 |
+
"""
|
1335 |
+
The Yuan Model transformer with a sequence classification head on top (linear layer).
|
1336 |
+
|
1337 |
+
[`YuanForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
1338 |
+
(e.g. GPT-2) do.
|
1339 |
+
|
1340 |
+
Since it does classification on the last token, it requires to know the position of the last token. If a
|
1341 |
+
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
1342 |
+
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
1343 |
+
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
1344 |
+
each row of the batch).
|
1345 |
+
""",
|
1346 |
+
YUAN_START_DOCSTRING,
|
1347 |
+
)
|
1348 |
+
class YuanForSequenceClassification(YuanPreTrainedModel):
|
1349 |
+
#_keys_to_ignore_on_load_missing = [r"lm_head.weight"]
|
1350 |
+
|
1351 |
+
def __init__(self, config):
|
1352 |
+
super().__init__(config)
|
1353 |
+
self.num_labels = config.num_labels
|
1354 |
+
self.model = YuanModel(config)
|
1355 |
+
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
|
1356 |
+
|
1357 |
+
# Initialize weights and apply final processing
|
1358 |
+
self.post_init()
|
1359 |
+
|
1360 |
+
def get_input_embeddings(self):
|
1361 |
+
return self.model.embed_tokens
|
1362 |
+
|
1363 |
+
def set_input_embeddings(self, value):
|
1364 |
+
self.model.embed_tokens = value
|
1365 |
+
|
1366 |
+
@add_start_docstrings_to_model_forward(YUAN_INPUTS_DOCSTRING)
|
1367 |
+
def forward(
|
1368 |
+
self,
|
1369 |
+
input_ids: torch.LongTensor = None,
|
1370 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1371 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1372 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1373 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1374 |
+
labels: Optional[torch.LongTensor] = None,
|
1375 |
+
use_cache: Optional[bool] = None,
|
1376 |
+
output_attentions: Optional[bool] = None,
|
1377 |
+
output_hidden_states: Optional[bool] = None,
|
1378 |
+
return_dict: Optional[bool] = None,
|
1379 |
+
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
1380 |
+
r"""
|
1381 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1382 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
1383 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
1384 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
1385 |
+
"""
|
1386 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1387 |
+
transformer_outputs = self.model(
|
1388 |
+
input_ids,
|
1389 |
+
attention_mask=attention_mask,
|
1390 |
+
position_ids=position_ids,
|
1391 |
+
past_key_values=past_key_values,
|
1392 |
+
inputs_embeds=inputs_embeds,
|
1393 |
+
use_cache=use_cache,
|
1394 |
+
output_attentions=output_attentions,
|
1395 |
+
output_hidden_states=output_hidden_states,
|
1396 |
+
return_dict=return_dict,
|
1397 |
+
)
|
1398 |
+
hidden_states = transformer_outputs[0]
|
1399 |
+
logits = self.score(hidden_states)
|
1400 |
+
|
1401 |
+
if input_ids is not None:
|
1402 |
+
batch_size = input_ids.shape[0]
|
1403 |
+
else:
|
1404 |
+
batch_size = inputs_embeds.shape[0]
|
1405 |
+
|
1406 |
+
if self.config.pad_token_id is None and batch_size != 1:
|
1407 |
+
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
|
1408 |
+
if self.config.pad_token_id is None:
|
1409 |
+
sequence_lengths = -1
|
1410 |
+
else:
|
1411 |
+
if input_ids is not None:
|
1412 |
+
sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
|
1413 |
+
else:
|
1414 |
+
sequence_lengths = -1
|
1415 |
+
|
1416 |
+
pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
|
1417 |
+
|
1418 |
+
loss = None
|
1419 |
+
if labels is not None:
|
1420 |
+
labels = labels.to(logits.device)
|
1421 |
+
if self.config.problem_type is None:
|
1422 |
+
if self.num_labels == 1:
|
1423 |
+
self.config.problem_type = "regression"
|
1424 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
1425 |
+
self.config.problem_type = "single_label_classification"
|
1426 |
+
else:
|
1427 |
+
self.config.problem_type = "multi_label_classification"
|
1428 |
+
|
1429 |
+
if self.config.problem_type == "regression":
|
1430 |
+
loss_fct = MSELoss()
|
1431 |
+
if self.num_labels == 1:
|
1432 |
+
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
|
1433 |
+
else:
|
1434 |
+
loss = loss_fct(pooled_logits, labels)
|
1435 |
+
elif self.config.problem_type == "single_label_classification":
|
1436 |
+
loss_fct = CrossEntropyLoss()
|
1437 |
+
loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
|
1438 |
+
elif self.config.problem_type == "multi_label_classification":
|
1439 |
+
loss_fct = BCEWithLogitsLoss()
|
1440 |
+
loss = loss_fct(pooled_logits, labels)
|
1441 |
+
if not return_dict:
|
1442 |
+
output = (pooled_logits,) + transformer_outputs[1:]
|
1443 |
+
return ((loss,) + output) if loss is not None else output
|
1444 |
+
|
1445 |
+
return SequenceClassifierOutputWithPast(
|
1446 |
+
loss=loss,
|
1447 |
+
logits=pooled_logits,
|
1448 |
+
past_key_values=transformer_outputs.past_key_values,
|
1449 |
+
hidden_states=transformer_outputs.hidden_states,
|
1450 |
+
attentions=transformer_outputs.attentions,
|
1451 |
+
)
|
1452 |
+
|
1453 |
+
|
1454 |
+
|