---
library_name: transformers
tags:
- trl
- sft
license: apache-2.0
datasets:
- Mike0307/alpaca-en-zhtw
language:
- zh
pipeline_tag: text-generation
base_model:
- microsoft/Phi-3-mini-4k-instruct
---


## Download Model

The base-model [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct) currently relies on 
the latest dev-version transformers and torch.<br>
Also, it needs *trust_remote_code=True* as an argument of the from_pretrained function.
```
pip install git+https://github.com/huggingface/transformers accelerate
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu
```

Additionally, LoRA adapter requires the peft package.
```
pip install peft
```

Now, let's start to download the adapter. 

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

model_id = "Mike0307/Phi-3-mini-4k-instruct-chinese-lora"
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    device_map="mps", # mps is for MacOS users
    torch_dtype=torch.float32,  # try float16 if needed
    trust_remote_code=True,
    attn_implementation="eager", # without flash_attn
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
```

## Inference Example

```python
# M2 pro takes about 3 seconds in this example.
input_text = "<|user|>將這五種動物分成兩組。\n老虎、鯊魚、大象、鯨魚、袋鼠 <|end|>\n<|assistant|>"

inputs = tokenizer(
    input_text, 
    return_tensors="pt"
).to(torch.device("mps")) # mps is for MacOS users

outputs = model.generate(
    **inputs, 
    temperature = 0.0,
    max_length = 500,
    do_sample = False
)

generated_text = tokenizer.decode(
    outputs[0], 
    skip_special_tokens=True
)
print(generated_text)
```


## Streaming Example
```python
from transformers import TextStreamer
streamer = TextStreamer(tokenizer)

input_text = "<|user|>將這五種動物分成兩組。\n老虎、鯊魚、大象、鯨魚、袋鼠 <|end|>\n<|assistant|>"

inputs = tokenizer(
    input_text, 
    return_tensors="pt"
).to(torch.device("mps")) # Change mps if not MacOS

outputs = model.generate(
    **inputs, 
    temperature = 0.0,
    do_sample = False,
    streamer=streamer,
    max_length=500,
)

generated_text = tokenizer.decode(
    outputs[0], 
    skip_special_tokens=True
)
```

## Example of RAG with Langchain

[This reference](https://huggingface.co/Mike0307/text2vec-base-chinese-rag#example-of-langchain-rag) shows how to customize langchain llm with this phi-3 lora model.

![image/png](https://cdn-uploads.huggingface.co/production/uploads/6414866f1cbd604c9217c7d0/RrBoHJINfrSWtCNkePs7g.png)