Native-speed vLLM transformers modeling backend

Model developers can reuse compatible Transformers model code in vLLM, a system for serving model responses, without writing a separate optimized implementation. The updated backend examines the model's computation structure and replaces recognized operation patterns with optimized vLLM operations. Its authors report throughput matching or exceeding vLLM's handwritten implementations on three tested Qwen3 configurations, up to a 235-billion-parameter mixture-of-experts model on eight H100 GPUs. Because Transformers code can also be used for training, the same model code can serve training and deployment. Compatibility is limited: the July 2026 article says linear-attention models are not currently supported and that custom model code stored in a Hub repository is unlikely to work. Implementation and test account