PyTorch Day India 2026 Beyond the Brrr: Building a Unified Ecosystem for Optimized Kernels Aritra Ro

As deep learning models grow in complexity, the "efficiency tax" of generic operators has become a critical bottleneck. While custom kernels offer the promise of hardware-saturating performance, making the hardware go "brrr", the reality of implementing them is fraught with challenges. Developers are often caught in a cycle of managing device-specific dependencies, handling complex compilation chains, and struggling with the lack of modularity when swapping kernels across different hardware backends. In this talk, we bridge the gap between low-level optimization and high-level developer productivity. We will explore why optimized kernels are essential for modern workloads and dissect the friction points that make them difficult to maintain and distribute. We will introduce the new Kernels library from Hugging Face, discussing how it makes the experience of working with pre-built kernels a seamless and enjoyable experience. Participants will learn how to discover, use, and share optimized kernels as easily as model checkpoints. They will also see how Kernels is integrated inside the mighty Transformers library and how it helps in shipping latency improvements.