Hugging Face Transformers Backend for vLLM Achieves Native Performance

The integration of Hugging Face Transformers as a modeling backend in vLLM now matches or exceeds native vLLM throughput across various Qwen3 models, including dense and mixture-of-experts architectures, without requiring custom porting. This is achieved through dynamic layer fusions at runtime using torch.fx and AST manipulation, allowing any compatible model to run at native speed with a single flag.

Hugging Face

NVIDIA uses Vera processor to accelerate design of future CPUs and GPUs

NVIDIA is collaborating with Cadence and Synopsys to optimize EDA applications for the new Vera processor. Early tests show up to 1.5x performance improvements on selected workloads. Vera is already being used in NVIDIA's own design processes to create future chips.

NVIDIA Cadence Design Systems Synopsys, Inc.