---
license: apache-2.0
task_categories:
- text-generation
tags:
- code
- cuda
- distributed-systems
- gpu-kernels
- benchmark
size_categories:
- n<1K
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
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# ParallelKernelBench (benchmark)

Reference problems for [ParallelKernelBench](https://github.com/your-org/ParallelKernelBench): a benchmark for LLM-generated multi-GPU CUDA kernels.

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

## Splits

Parquet files are sharded by difficulty level (problem id prefix):

| Split | Problem IDs | Count |
|-------|-------------|-------|
| level_1 | 1–18 | 18 |
| level_2 | 19–27 | 9 |
| level_3 | 28–99 | 60 |
| level_4 | 100+ | 0 |

## Columns

- `problem_id`, `stem`, `level` — problem identity
- `reference_code` — full Python source
- `reference_path` — path to the same file in this repo
- `world_size`, `default_m`, `default_n`, `default_dtype`, `default_trials` — default eval settings (8× H100, 1024×1024, bfloat16, 5 trials)

## Usage

```python
from datasets import load_dataset

ds = load_dataset("YOUR_ORG/ParallelKernelBench", "level_1")
print(ds[0]["stem"], ds[0]["reference_code"][:200])
```

Reproduce inputs locally (requires this repo's harness):

```python
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

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