flashinfer-ai/mlsys26-contest

---
license: apache-2.0
---

# MLSys 2026 FlashInfer-Bench Challenge Dataset

This repository contains the FlashInfer-Bench dataset for the MLSys 2026 Kenrel Generation Challenge.

This dataset targets to be used in the [FlashInfer-Bench](https://github.com/flashinfer-ai/flashinfer-bench) benchmark system.

It follows the [FlashInfer Trace Schema](https://bench.flashinfer.ai/docs/flashinfer-trace). To use the dataset in the competition, please refer to our [starter kit](https://github.com/flashinfer-ai/flashinfer-bench-starter-kit).

## Download

Use this command to download the dataset:
```bash
git lfs install
git clone https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest
```

Set the environment variable so that FlashInfer-Bench can find the dataset:
```bash
export FIB_DATASET_PATH=/path/to/mlsys26-contest
```

## Tasks

This dataset contains the definitions and workloads for these kernels:
* Fused Mixture of Experts (MoE)
* Gated Delta Network (GDN)
* DeepSeek Sparse Attention (DSA)

## Dataset Structure

It is organized as follows:
```
mlsys26-contest/
├── definitions/
└── workloads/
```

These components are provided in the dataset:
* **Definition**: describes the input, output, and computation logic of a kernel task.
* **Workload**: describes the inputs for a definition during real inference. This will be used to benchmark the **Solution** you provided.

During benchmarking, these components should be provided or generated:
* **Solution**: provided by participants, your implementation of the kernel task.
* **Trace**: generated by FlashInfer-Bench, the performance and correctness results of your solution on the workloads.