Uploaded March 2026 | Updated November 2026, 6 days ago
In this talk, Ben Burtenshaw from Hugging Face breaks down why optimized kernels are critical for real-world deep learning performance and how the Hugging Face Kernels ecosystem makes them easier to build and use.
He covers memory-bound bottlenecks, the kernel-builder workflow, reproducible multi-hardware builds with Nix, and practical PyTorch/Transformers integration patterns that reduce setup time from hours to seconds.
## Chapters
0:00 Intro and speaker background
1:35 Why Hugging Face Kernels matters
2:05 Compute vs memory bottlenecks in deep learning
3:30 Fused kernels and why they speed things up
5:05 Talk agenda and ecosystem overview
5:35 Kernel pain points: fragmentation and long installs
7:12 Supporting older, cheaper hardware for the community
8:18 Goal: from CMake errors to one-line kernel usage
8:54 Kernels + kernel-builder architecture
10:00 Reproducible builds with Nix and support matrix
11:45 Kernel project structure (`build.toml`, sources, torch extension)
12:23 Publishing kernels to the Hugging Face Hub
13:25 Real-world gain: faster FlashAttention setup
14:18 Docs, repos, and how to get started
16:00 Verifying compatibility and loading kernels in Python
17:20 Managing local cache with `hf cache ls`
17:55 Kernelizing PyTorch layers with hub mappings
19:32 Transformers integration (`use_kernels=True`)
20:48 Performance chart and closing resources
Learn more:
aifoundry.org
huggingface.co/docs/kernels
github.com/huggingface/kernels
huggingface.co/kernels-community
The video was created by @AIFoundry Org
In this talk, Ben Burtenshaw from Hugging Face breaks down why optimized kernels are critical for real-world deep learning performance and how the Hugging Face Kernels ecosystem makes them easier to build and use.
He covers memory-bound bottlenecks, the kernel-builder workflow, reproducible multi-hardware builds with Nix, and practical PyTorch/Transformers integration patterns that reduce setup time from hours to seconds.
## Chapters
0:00 Intro and speaker background
1:35 Why Hugging Face Kernels matters
2:05 Compute vs memory bottlenecks in deep learning
3:30 Fused kernels and why they speed things up
5:05 Talk agenda and ecosystem overview
5:35 Kernel pain points: fragmentation and long installs
7:12 Supporting older, cheaper hardware for the community
8:18 Goal: from CMake errors to one-line kernel usage
8:54 Kernels + kernel-builder architecture
10:00 Reproducible builds with Nix and support matrix
11:45 Kernel project structure (`build.toml`, sources, torch extension)
12:23 Publishing kernels to the Hugging Face Hub
13:25 Real-world gain: faster FlashAttention setup
14:18 Docs, repos, and how to get started
16:00 Verifying compatibility and loading kernels in Python
17:20 Managing local cache with `hf cache ls`
17:55 Kernelizing PyTorch layers with hub mappings
19:32 Transformers integration (`use_kernels=True`)
20:48 Performance chart and closing resources
Learn more:
aifoundry.org
huggingface.co/docs/kernels
github.com/huggingface/kernels
huggingface.co/kernels-community
The video was created by @AIFoundry Org










