Hugging Face Kernels: swap kernels via mapping changes, no model rewrite needed
RisingSayak · x · 2026-09-18
- Hugging Face's Kernels library lets you swap in compatible kernels for existing layers as a mapping change, not a model rewrite — try one, benchmark your workload, keep what helps.
- You can specify separate kernels for inference vs. training, and different kernels per device type, registered globally or locally in a mapping.
- Mechanism: annotate a layer with usekernelforwardfromhub() (e.g. SiluAndMul); kernelize() looks up Hub kernels by name to replace the forward method.
Related event: Hugging Face Launches Kernels Library for Drop-in Faster Kernels(2 posts)→
More from Infra
- Nvidia struggles to hire RL talent from Anthropic, highlighting chipmakers' reinforcement learning gap — teortaxesTex · 2026-11-04
- Sam Altman? No—Bezos Calls for Nuclear-Powered Offshore Data Center Platforms — beffjezos · 2026-11-04
- SpecPilot agent boosts speculative decoding from 4.46× to 10.87×, no retraining — pmttyji · 2026-11-04
- NEURA launches Neuron AI, but Europe's Physical AI test is whether the stack matures together — ingliguori · 2026-11-04
- S3-based databases cache in memory and NVMe, making disk vs S3 latency gap smaller than expected — mgill25 · 2026-11-04
- gpu-unlock project claims up to 2x FP16 tensor throughput for RTX 20/30/40 cards — paranoidray · 2026-11-04