# **NanoTranslator-XL**

[English](README.md) | 简体中文

## Introduction

这是 NanoTranslator 的 **XX-Large** 型号，目前仅支持**英译中**。仓库中同时提供了 ONNX 版本的模型。

所有模型均收录于 [NanoTranslator Collection](https://huggingface.co/collections/Mxode/nanotranslator-66e1de2ba352e926ae865bd2) 中。

|  | P. | Arch. | Act. |  V.  |  H.  |  I.  |  L.  | A.H. | K.H. | Tie |
| :--: | :-----: | :--: | :--: | :--: | :-----: | :---: | :------: | :--: | :--: | :--: |
|  [XXL](https://huggingface.co/Mxode/NanoTranslator-XXL)  |  100  |  LLaMA  |  SwiGLU  | 16000 | 768  | 4096 |  8   |  24  |  8   | True |
|  [XL](https://huggingface.co/Mxode/NanoTranslator-XL)  |  78  | LLaMA | GeGLU  | 16000 | 768  | 4096 |  6   |  24  |  8   | True |
| [L](https://huggingface.co/Mxode/NanoTranslator-L) |  49  | LLaMA | GeGLU  | 16000 | 512  | 2816 |  8   |  16  |  8   | True |
| [M2](https://huggingface.co/Mxode/NanoTranslator-M2) | 22 | Qwen2 | GeGLU | 4000  | 432  | 2304 |  6   |  24  |  8   | True |
|  [M](https://huggingface.co/Mxode/NanoTranslator-M)   |  22  |  LLaMA  |  SwiGLU  | 8000  | 256  | 1408 |  16  |  16  |  4   | True |
|  [S](https://huggingface.co/Mxode/NanoTranslator-S)   | 9 | LLaMA | SwiGLU | 4000  | 168  | 896  |  16  |  12  |  4   | True |
| [XS](https://huggingface.co/Mxode/NanoTranslator-XS) | 2 | LLaMA | SwiGLU | 2000 | 96 | 512 | 12 | 12 | 4 | True |

- **P.** - Parameters (in million)
- **V.** - vocab size
- **H.** - hidden size
- **I.** - intermediate size
- **L.** - num layers
- **A.H.** - num attention heads
- **K.H.** - num kv heads
- **Tie** - tie word embeddings



## How to use

Prompt 格式如下：

```
<|im_start|> {English Text} <|endoftext|>
```

### Directly using transformers

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_path = 'Mxode/NanoTranslator-XL'

tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path)

def translate(text: str, model, **kwargs):
    generation_args = dict(
        max_new_tokens = kwargs.pop("max_new_tokens", 512),
        do_sample = kwargs.pop("do_sample", True),
        temperature = kwargs.pop("temperature", 0.55),
        top_p = kwargs.pop("top_p", 0.8),
        top_k = kwargs.pop("top_k", 40),
        **kwargs
    )

    prompt = "<|im_start|>" + text + "<|endoftext|>"
    model_inputs = tokenizer([prompt], return_tensors="pt").to(model.device)

    generated_ids = model.generate(model_inputs.input_ids, **generation_args)
    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]
    return response

text = "I love to watch my favorite TV series."

response = translate(text, model, max_new_tokens=64, do_sample=False)
print(response)
```


### ONNX

根据实际测试，使用 ONNX 模型推理会比直接使用 transformers 推理要**快 2～10 倍**。

如果希望使用 ONNX 模型，那么你需要手动切换到 [onnx 分支](https://huggingface.co/Mxode/NanoTranslator-XL/tree/onnx)并从本地加载。

参考文档：

- [Export to ONNX](https://huggingface.co/docs/transformers/serialization)
- [Inference pipelines with the ONNX Runtime accelerator](https://huggingface.co/docs/optimum/main/en/onnxruntime/usage_guides/pipelines)

**Using ORTModelForCausalLM**

```python
from optimum.onnxruntime import ORTModelForCausalLM
from transformers import AutoTokenizer

model_path = "your/folder/to/onnx_model"

ort_model = ORTModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)

text = "I love to watch my favorite TV series."

response = translate(text, ort_model, max_new_tokens=64, do_sample=False)
print(response)
```

**Using pipeline**

```python
from optimum.pipelines import pipeline

model_path = "your/folder/to/onnx_model"
pipe = pipeline("text-generation", model=model_path, accelerator="ort")

text = "I love to watch my favorite TV series."

response = pipe(text, max_new_tokens=64, do_sample=False)
response
```