PyTorch Taps Helion to Unify High-Performance AI Kernel Development

PARIS, FRANCE – April 07, 2026 – The PyTorch Foundation, the open-source AI hub operating under the Linux Foundation, today announced it has welcomed Helion as its newest foundation-hosted project. Contributed by Meta, Helion is a new language designed to standardize and simplify the complex process of writing AI kernels, the fundamental code blocks that dictate performance on specialized hardware. The move signals a major push to make high-performance AI more portable and accessible to developers across a rapidly diversifying hardware landscape.

Helion joins a prestigious portfolio of projects within the PyTorch Foundation, including DeepSpeed, Ray, and vLLM, reinforcing the foundation's role in building a comprehensive, open-source AI stack. The announcement comes as the AI industry grapples with a significant shift from model training to an “inference boom,” where the efficient deployment of models at scale has become a critical bottleneck.

“Helion joining the PyTorch Foundation as its newest project reflects where the open AI ecosystem needs to go next: higher-level performance portability for kernel authors,” said Matt White, Global CTO of AI at the Linux Foundation and CTO of the PyTorch Foundation. “Helion gives engineers a much more productive path to writing high-performance kernels.”

The Kernel Conundrum

At the heart of every AI model's performance are kernels—small, highly optimized programs that run directly on hardware like GPUs. Writing these kernels has traditionally been a dark art, requiring deep expertise in low-level programming languages like CUDA and an intimate understanding of specific chip architectures. This complexity creates a high barrier to entry, slows down development, and often leads to code that is locked into a single hardware vendor.

As AI models become more sophisticated and the hardware used to run them more varied—spanning GPUs from NVIDIA, AMD, and Intel, as well as custom accelerators—this challenge has intensified. Engineering teams face significant hurdles in ensuring their models run efficiently and consistently across different platforms. The manual effort required to port and re-optimize kernels for each new piece of hardware results in immense technical debt and stifles innovation.

Helion aims to solve this problem by raising the level of abstraction. It provides a vital software layer that sits between the AI model and the hardware, automating many of the most difficult parts of kernel authoring. By doing so, it promises to democratize performance engineering, allowing a broader range of developers to create highly efficient models without getting bogged down in hardware-specific details.

Inside Helion: Automation and Portability

Helion is a Python-embedded domain-specific language (DSL) that feels native to developers already working within the PyTorch ecosystem. Described as a “higher-level Triton,” it builds upon existing technologies but introduces powerful new capabilities, chief among them being its sophisticated autotuning engine.

While other frameworks like OpenAI's Triton simplified kernel writing compared to raw CUDA, they still often require developers to manually define the optimization search space. Helion's key innovation is its automated, ahead-of-time (AOT) autotuning. From a single, high-level Helion kernel, the system can automatically generate and test hundreds or even thousands of potential low-level implementations, benchmarking them to find the most performant configuration for a specific target hardware. This automated exploration of the optimization space allows Helion to often match or exceed the performance of kernels that were painstakingly tuned by human experts.

“Helion brings kernel authoring into PyTorch – making it simpler, portable, and accessible to every developer,” noted Jana van Greunen, Director of PyTorch Engineering at Meta. “Joining the PyTorch Foundation opens Helion to the broader hardware ecosystem, so developers write one kernel and it runs fast everywhere.”

This capability directly enables Helion's second core promise: hardware performance portability. The language is designed to compile down to multiple backends, including Triton and TileIR, ensuring that a kernel written once can be efficiently deployed across a wide array of hardware. This frees developers from the cycle of rewriting code for every new chip, future-proofing their work and dramatically accelerating the deployment process.

Meta's Strategic Gift to the Open Source Ecosystem

Meta's decision to contribute Helion to the PyTorch Foundation is a strategic move that aligns with its long-standing commitment to open-source AI. By placing this critical technology under a neutral, community-driven governance model, Meta is not just giving away a tool but is actively shaping the future of the entire AI software stack. The move reinforces PyTorch's position as a leading end-to-end platform for AI development.

This strategy is further evidenced by the simultaneous announcement that ExecuTorch, another Meta-led project focused on enabling PyTorch models on edge and mobile devices, is being integrated into PyTorch Core. Together, the Helion and ExecuTorch initiatives show a concerted effort to ensure PyTorch can power AI applications everywhere, from massive data centers to the smartphone in your pocket, all while maintaining performance and developer productivity.

Hosting Helion within the PyTorch Foundation provides a vendor-neutral ground for collaboration. It invites hardware manufacturers, cloud providers, and academic researchers to contribute to and benefit from a shared, open standard for performance engineering.

“By bringing Helion into the PyTorch Foundation community, we are meeting the technical frontier of AI head on,” said Mark Collier, Executive Director of the PyTorch Foundation. “The project provides a vital layer of abstraction that makes it easier for developers to target different architectures and accelerate AI adoption.”

This collaborative approach is essential for tackling the systemic challenges of hardware fragmentation. Instead of individual companies building siloed solutions, the foundation provides a forum for creating common infrastructure that benefits everyone. As part of this community, Helion is expected to evolve with contributions from across the industry, ensuring it remains a robust and relevant solution as the AI landscape continues to change at a breakneck pace. Developers interested in contributing are encouraged to participate in upcoming events like the PyTorch Conferences in Shanghai and San Jose later this year.