File size: 2,956 Bytes
82bf218
 
2de3761
 
 
 
 
 
 
 
82bf218
2de3761
4f31db2
2de3761
 
 
 
cabbdca
2de3761
c287120
2de3761
 
0740d09
2de3761
 
 
 
 
dcf47f5
2de3761
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dcf47f5
2de3761
 
 
 
 
3c4bbf6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
---
license: other
license_name: tongyi-qianwen
license_link: >-
  https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B-Chat/blob/main/LICENSE
language:
- en
pipeline_tag: text-generation
tags:
- chat
---

# Qwen1.5-MoE-A2.7B-Chat


## Introduction

Qwen1.5-MoE is a transformer-based MoE decoder-only language model pretrained on a large amount of data. 

For more details, please refer to our [blog post](https://qwenlm.github.io/blog/qwen-moe/) and [GitHub repo](https://github.com/QwenLM/Qwen1.5).

## Model Details
Qwen1.5-MoE employs Mixture of Experts (MoE) architecture, where the models are upcycled from dense language models. For instance, `Qwen1.5-MoE-A2.7B` is upcycled from `Qwen-1.8B`. It has 14.3B parameters in total and 2.7B activated parameters during runtime, while achieching comparable performance to `Qwen1.5-7B`, it only requires 25% of the training resources. We also observed that the inference speed is 1.74 times that of `Qwen1.5-7B`.

## Training details
We pretrained the models with a large amount of data, and we post-trained the models with both supervised finetuning and direct preference optimization. However, DPO leads to improvements in human preference evaluation but degradation in benchmark evaluation. In the very near future, we will fix both problems.

## Requirements
The code of Qwen1.5-MoE has been in the latest Hugging face transformers and we advise you to build from source with command `pip install git+https://github.com/huggingface/transformers`, or you might encounter the following error:
```
KeyError: 'qwen2_moe'.
```

## Quickstart

Here provides a code snippet with `apply_chat_template` to show you how to load the tokenizer and model and how to generate contents.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen1.5-MoE-A2.7B-Chat",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-MoE-A2.7B-Chat")

prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
```

For quantized models, we advise you to use the GPTQ correspondents, namely `Qwen1.5-MoE-A2.7B-Chat-GPTQ-Int4`.


## Tips

* If you encounter code switching or other bad cases, we advise you to use our provided hyper-parameters in `generation_config.json`.
*