GPUMODE/pytorch_scrape_inductor_data

This dataset is composed of scraping code off of github containing pytorch code, and then running torch compile on it in order to have pairs of pytorch and triton code.

To spot check the data you can run

from datasets import load_dataset

# Load the dataset
ds = load_dataset("GPUMODE/pytorch_scrape_inductor_data")

# Get the first row
row = ds["train"][0]

print("\n=== UUID ===")
print(row['uuid'])

print("\n=== PYTHON CODE ===")
print(row['python_code'])

print("\n=== TRITON CODE ===")
print(row['triton_code'])

Furthermore, you can run the python code directly by running python run_code.py dataset.parquet <index> where index is number between 0 and the length of the dataset.

Todo

• Add license information to data
• Expand dataset to entirety of github