VGen: Automated Verilog RTL Code Generation

For optimized memory and compute usage with lower latency, use the customized version of fauxpilot. This method allows you to query the fine-tuned LLMs via an OpenAI-compatible API.

1. Clone the customized repository: git clone https://github.com/shailja-thakur/fauxpilot_changes.git.
2. Follow the specific setup instructions in the fauxpilot_changes repository.
3. Once the server is running, use the openai Python library to send completion requests to the local endpoint (default http://127.0.0.1:5000/v1).

import openai

openai.api_key = 'dummy'
openai.api_base = 'http://127.0.0.1:5000/v1'

result = openai.Completion.create(
    engine='codegen',
    prompt='//module half adder',
    max_tokens=100,
    temperature=0.1,
    n=3,
    top_p=1.0,
    stop=["endmodule"]
)

print(result)

To use BigQuery in your Python environment, you must provide your Google Cloud credentials via the GOOGLE_APPLICATION_CREDENTIALS environment variable. This variable should point to the path of your service account JSON key file.

Note: If you encounter an ImportError: No module named 'google', ensure you have installed the necessary libraries and loaded the BigQuery extension in your notebook environment using %load_ext google.cloud.bigquery.

import os
from google.cloud import bigquery

# Set the path to your service account JSON key
os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = r"path/to/your-service-account-key.json"

# Initialize the BigQuery client
bqclient = bigquery.Client()

Load the fine-tuned CodeGen_2B_Verilog model and its corresponding tokenizer using the transformers library. It is recommended to move the model to a CUDA device if available.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

if torch.cuda.is_available():
    device = torch.device("cuda")
    print("Using GPU")
else:
    device = torch.device("cpu")
    print("Using CPU")

model_name = "shailja/CodeGen_2B_Verilog"
# Note: The demo uses 'shailja/fine-tuned-codegen-2B-Verilog' for both tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("shailja/fine-tuned-codegen-2B-Verilog")
model = AutoModelForCausalLM.from_pretrained("shailja/fine-tuned-codegen-2B-Verilog").to(device)

Use the transformers library to load the pre-trained Verilog model and its corresponding tokenizer. The model shailja/fine-tuned-codegen-6B-Verilog is specifically fine-tuned for Verilog generation tasks.

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "shailja/CodeGen_6B_Verilog"

tokenizer = AutoTokenizer.from_pretrained("shailja/fine-tuned-codegen-6B-Verilog")
model = AutoModelForCausalLM.from_pretrained("shailja/fine-tuned-codegen-6B-Verilog").to(device)

Before loading the model, check if a CUDA-enabled GPU is available to ensure efficient inference. If torch.cuda.is_available() returns true, set the device to cuda; otherwise, fallback to cpu.

import torch

if torch.cuda.is_available():
    device = torch.device("cuda")
else:
    device = torch.device("cpu")

To use the VGen models, you need to install the transformers library directly from the HuggingFace GitHub repository to ensure compatibility with the fine-tuned models.

Ensure you have a GPU available for inference, as the models are designed for CUDA execution.

!pip install git+https://github.com/huggingface/transformers.git

Generate Verilog code by providing a text prompt (e.g., a comment describing a module). Use model.generate() with parameters like max_length, temperature, and top_p to control the output. You can use truncate_before_pattern to clean up the output based on specific Verilog keywords like endmodule.

prompt = "//a half adder module "
n_steps = 1
n = 5
end_pattern='endmodule'

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"]) + end_pattern)

To collect Verilog-related discussions, you can join the posts_questions and posts_answers tables from the bigquery-public-data.stackoverflow dataset, filtering by tags or body content containing 'Verilog'.

SELECT 
    q.tags question_tag, q.id question_id, q.title question_title, q.body question_body, 
    a.id answer_id, a.body 
FROM `bigquery-public-data.stackoverflow.posts_questions` as q
LEFT JOIN `bigquery-public-data.stackoverflow.posts_answers` as a
ON a.parent_id = q.id
WHERE regexp_contains(q.tags, 'Verilog');