FlashInfer Trace We provide an official dataset called FlashInfer Trace with kernels and workloads in real world AI system deployment environments. FlashInfer Bench can use this dataset to measure and compare the performance of kernels. It follows the FlashInfer Trace Schema. It is organized as follows: Example solutions and traces directories, featuring reference implementations and benchmark logs, are available on Hugging Face: https://huggingface.co/datasets/flashinfer ai/flashinfer trace The layout is: solutions/ / / / .json traces/ / / .jsonl Each Definition describes a computation task and reference logic. Each Solution specifies a kernel or agent implementation for a definition. Each Workload contains the inputs for a definition during real inference. Each Trace records a benchmark result: input config, performance, correctness, environment, etc. Components Definition This component provides a formal definition for a specific computational workload encountered in a model's forward pass. It specifies the expected input and output formats. We also include a mathematical specification of the workload in the form of PyTorch code. This serves as both a precise description of the…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy