VeriGen
Table of Contents
Model Summary
The VeriGen model is 2B parameter model is a fine-tuned version of CodeGen-multi-2B, trained on Verilog Dataset, with a context length of 2048.
- Repository: shailja-thakur/VGen
- Baseline LLM SalesForce/CodeGen
- Paper: Benchmarking Large Language Models for Automated Verilog RTL Code Generation
- Point of Contact: contact@shailja
- Languages: Verilog (Hardware Description Language)
Use
Intended use
The model was trained on Verilog from GitHub and textbooks. As such it is not an instruction model and commands like "Write a module that implements a 2-to-1 Mux." do not work well. However, by additing a partial line of module header like "module mux" in addition with the text in the prompt turns it into a capable Verilog teaching assistant.
Feel free to share your generations in the Community tab!
Generation
# pip install -q transformers
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
# Prompt
prompt = "//module half adder "
device='cuda'
# Load model and tokenizer
model_name = "shailja/CodeGen_2B_Verilog"
tokenizer = AutoTokenizer.from_pretrained("shailja/fine-tuned-codegen-2B-Verilog")
model = AutoModelForCausalLM.from_pretrained("shailja/fine-tuned-codegen-2B-Verilog").to(device)
# Sample
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
sample = model.generate(input_ids, max_length=128, temperature=0.5, top_p=0.9)
print(tokenizer.decode(sample[0], truncate_before_pattern=[r"endmodule"]) + "endmodule")
Attribution & Other Requirements
The pretraining dataset of the model was not filtered for permissive licenses only. Nevertheless, the model can generate source code verbatim from the dataset. The code's license might require attribution and/or other specific requirements that must be respected.
Limitations
The model has been trained on Verilog source code from open sources. The predominant natural language in source code is English, although other languages are also present. As such the model is capable of generating Verilog snippets provided some context but the generated code is not guaranteed to work as intended. It can be inefficient, contain bugs or exploits. See the paper for an in-depth discussion of the model limitations.
Training
Model
- Architecture: GPT-2 model with multi-query attention
- Pretraining steps: 150k
- Pretraining tokens: ~72B
- Precision: fp16
Hardware
- GPUs: 3 Tesla A100
- Training time: 8 days
License
The model is licensed under the BigCode OpenRAIL-M v1 license agreement. You can find the full agreement here.
Citation
@misc{https://doi.org/10.48550/arxiv.2212.11140,
doi = {10.48550/ARXIV.2212.11140},
url = {https://arxiv.org/abs/2212.11140},
author = {Thakur, Shailja and Ahmad, Baleegh and Fan, Zhenxing and Pearce, Hammond and Tan, Benjamin and Karri, Ramesh and Dolan-Gavitt, Brendan and Garg, Siddharth},
title = {Benchmarking Large Language Models for Automated Verilog RTL Code Generation},
publisher = {arXiv},
year = {2022},
copyright = {arXiv.org perpetual, non-exclusive license}
}