File size: 3,211 Bytes
a509f01 4b85e0b 7b1f6e0 4b85e0b 96364e0 17629dc 96364e0 97b297e 96364e0 17629dc 293270e 17629dc 96364e0 f2403c0 96364e0 9382a3f 96364e0 9382a3f 96364e0 ca5ec62 64e8d68 ca5ec62 96364e0 |
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 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 |
---
license: gpl-3.0
base_model: deepseek-ai/deepseek-coder-7b-instruct-v1.5
library_name: peft
---
# OriGen: Enhancing RTL Code Generation with Code-to-Code Augmentation and Self-Reflection
### Introduction
OriGen is a fine-tuned lora model designed for Verilog code generation. It is trained on top of DeepSeek Coder 7B using datasets generated from code-to-code augmentation and self-reflection.
The model has been uploaded to Hugging Face, and the repository contains the inference scripts. The dataset and data generation flow will be released soon.
- **Huggingface**: https://huggingface.co/henryen/OriGen
- **Repository**: https://github.com/pku-liang/OriGen
### Evaluation Results
<img src="figures/evaluation.png" alt="evaluation" width="1000"/>
### Quick Start
Before running the following code, please install the required packages:
```bash
conda create -n origen python=3.11
conda activate origen
pip install -r requirements.txt
```
Here is an example of how to use the model. Please note that the base model, DeepSeek Coder 7B, is loaded in float16 precision, even though its default precision is bfloat16. This choice was made because our experiments showed that Lora trained in float16 outperforms those trained in bfloat16.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
import torch
from peft import PeftModel
model_name = "deepseek-ai/deepseek-coder-7b-instruct-v1.5"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
low_cpu_mem_usage=True,
torch_dtype=torch.float16,
attn_implementation="flash_attention_2",
device_map="auto",
).to("cuda")
model = PeftModel.from_pretrained(model, model_id="henryen/OriGen")
model.eval()
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
prompt = "### Instruction: Please act as a professional Verilog designer and provide Verilog code based on the given description. Create a 8 bit full adder with only one assign statement.\n### Response: "
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(
**inputs,
max_new_tokens=1000,
do_sample=False,
temperature=0,
eos_token_id=tokenizer.eos_token_id,
pad_token_id=tokenizer.pad_token_id,
streamer=streamer
)
```
The output will be:
```verilog
module full_adder(
input [7:0] a,
input [7:0] b,
input cin,
output [7:0] sum,
output cout
);
assign {cout, sum} = a + b + cin;
endmodule
```
### Verilog-Eval Benchmark
We have released the scripts for the Verilog-Eval benchmark. Please refer to the [README](https://github.com/pku-liang/OriGen/blob/main/evaluation/README.md) for details.
### Paper
**Arxiv:** https://arxiv.org/abs/2407.16237
Please cite our paper if you use this model.
```
@article{2024origen,
title={OriGen: Enhancing RTL Code Generation with Code-to-Code Augmentation and Self-Reflection},
author={Cui, Fan and Yin, Chenyang and Zhou, Kexing and Xiao, Youwei and Sun, Guangyu and Xu, Qiang and Guo, Qipeng and Song, Demin and Lin, Dahua and Zhang, Xingcheng and others},
journal={arXiv preprint arXiv:2407.16237},
year={2024}
}
```
|