README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
tags:
  - code
  - cuda
  - distributed-systems
  - gpu-kernels
  - benchmark
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: problem_id
      dtype: int64
    - name: stem
      dtype: string
    - name: reference_code
      dtype: string
    - name: reference_path
      dtype: string
    - name: input_tensor_spec_path
      dtype: string
    - name: world_size
      dtype: int64
    - name: default_m
      dtype: int64
    - name: default_n
      dtype: int64
    - name: default_dtype
      dtype: string
    - name: default_trials
      dtype: int64
  splits:
    - name: train
      num_bytes: 280010
      num_examples: 87
  download_size: 87290
  dataset_size: 280010

ParallelKernelBench (benchmark)

Reference problems for ParallelKernelBench: a benchmark for LLM-generated multi-GPU CUDA kernels.

This dataset contains 87 reference implementations in reference/ and the input tensor specification in utils/input_output_tensors.py. Inputs are deterministic — reproduce them with create_input_tensor(rank, world_size, problem_id, base_shape, dtype, trial) from that file; you do not need stored .pt files.

Files

Path	Description
data/problems.parquet	One row per problem (tabular access)
reference/*.py	Reference solution() implementations
utils/input_output_tensors.py	Input/output tensor generation for every problem

Columns (data/problems.parquet)

• problem_id, stem — problem identity
• reference_code — full Python source
• reference_path — path to the same file in this repo
• input_tensor_spec_path — path to utils/input_output_tensors.py (same on every row)
• world_size, default_m, default_n, default_dtype, default_trials — default eval settings (8× H100, 1024×1024, bfloat16, 5 trials)

Usage

from datasets import load_dataset
from huggingface_hub import hf_hub_download

ds = load_dataset("YOUR_ORG/ParallelKernelBench", split="train")
print(ds[0]["stem"], ds[0]["reference_code"][:200])

# Fetch the input tensor spec (same file on disk in this dataset repo)
spec_path = hf_hub_download("YOUR_ORG/ParallelKernelBench", "utils/input_output_tensors.py", repo_type="dataset")

Reproduce inputs locally (add the downloaded utils/ folder to PYTHONPATH, or clone this repo):

from utils.input_output_tensors import create_input_tensor
import torch

x = create_input_tensor(
    rank=0, world_size=8, problem_id=17,
    base_shape=(1024, 1024), dtype=torch.bfloat16,
)

Related

Net-new LLM-generated kernels live in a separate dataset: ParallelKernelBench-kernels (same org).

Eval

python run_local.py --mode eval --problem 17 --solution cuda \
  --solutions-root path/to/solutions_dir --dtype bfloat16 --trials 5