Hugging Face just made GPT-OSS's best tricks part of every model

Remember how GPT-OSS got those impressive efficiency numbers? The secret wasn’t just the architecture — it was a stack of custom kernels that kept memory use low and throughput high. Hugging Face just baked those same tricks into transformers in a way that works across the library, not just for OpenAI’s model. And they’re downloadable pre-compiled from the Hub, so you skip the dependency hell entirely.

This is a genuinely big deal for anyone running open models locally. The update introduces a use_kernels=True flag that pulls pre-built binaries for things like Liger RMSNorm and MegaBlocks MoE operations. When you enable it, you’ll see log messages confirming the library is swapping in community-contributed kernels optimized for your hardware — CUDA or ROCm, training or inference. No compiling CUDA code yourself, no chasing compatible versions of five different kernel libraries.

OpenAI’s GPT-OSS relies heavily on this approach. The blog post walks through how decorators like @use_kernel_forward_from_hub("RMSNorm") and @use_kernel_forward_from_hub("MegaBlocksMoeMLP") point directly to repos from liger_kernels and megablocks. The system checks your device and workload, then downloads the right binary on first use. It’s clean enough that the RMSNorm kernel is already being reused across multiple models, and the MoE kernel could apply to any future mixture-of-experts architecture.

There’s a catch, though. These kernels aren’t compatible with the new MXFP4 quantization — inference falls back to bfloat16 if you use them. The team is refreshingly direct about this: benchmark your own system to find the sweet spot between memory and throughput for your specific setup. In their testing, the kernel advantage really showed at larger batch sizes. For smaller workloads, you might be better off with the quantization. The broader point is that transformers is becoming less of a model library and more of an optimization platform. Features like Flash Attention 3 with attention sink support, tensor parallelism, and continuous batching are all landing as composable pieces, not one-off integrations for a single model.