license: apache-2.0
MLSys 2026 FlashInfer-Bench Challenge Dataset (Unofficial Patch)
This is an unofficial patch of the original mlsys26-contest dataset. It is not affiliated with or endorsed by the FlashInfer team.
What was changed
The original GDN prefill workloads cause numerical explosion in the reference
implementation. Random keys generated by the harness (torch.randn) have
||k||^2 ~ 128, making the delta rule's state transition matrix spectrally
unstable whenever the gate g exceeds 1 / (beta * ||k||^2) ~ 0.005.
Fix: The dt_bias tensor in all 100 GDN prefill safetensors files was
adjusted per-head so that g < 0.005 for every timestep. No other tensors,
workload entries, definitions, or non-GDN data were modified.
Original README
This repository contains the FlashInfer-Bench dataset for the MLSys 2026 Kenrel Generation Challenge.
This dataset targets to be used in the FlashInfer-Bench benchmark system.
It follows the FlashInfer Trace Schema. To use the dataset in the competition, please refer to our starter kit.
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.