Exploring collaboration: Integrating FlagOS Kernels into HuggingFace Kernels

[Dongxu-H]

Hello Hugging Face Kernels team,

We are contributors from the FlagOS project and would like to explore a potential collaboration with huggingface/kernels GitHub repository.

At FlagOS, we are working on infrastructure to automatically generate high-performance kernels based on Triton (programming language).
Our project KernelGen focuses on automatically generating optimized operators.
Through our runtime and abstraction layer FlagTree, we support multiple (10+) accelerator platforms, such as Ascend, Moore Threads, NVIDIA and so on, with plans to support more hardware backends.

In our current experiments, KernelGen already demonstrates promising results:
Up to 75.5% execution correctness across generated operators
74.2% of operators achieve speedup > 0.8
68.5% of operators achieve speedup > 1.0
Median speedup: 1.04×, average speedup: 1.07×

We recently came across the HuggingFace Kernels initiative and found its goals very aligned with ours — especially the idea of packaging and publishing reusable kernels through the kernels ecosystem and supporting diverse hardware via the builder framework and kernels community.

We are wondering whether there might be a way for kernels generated by KernelGen to be published within the HuggingFace Kernels ecosystem. In addition, we are planning to evolve KernelGen into an MCP-style service interface with corresponding skills, which could make kernel generation and integration more accessible to external systems.

If this sounds interesting, we would love to discuss possible integration approaches or contribution guidelines.

Thanks for building this ecosystem!

Best regards,
Dongxu Han