SGLang adds Hugging Face transformers backend for high-throughput LLM inference

AI Impact Summary

SGLang now integrates the Hugging Face transformers backend, enabling high-throughput, low-latency inference for HF models while preserving native SGLang paths. If a model isn’t natively supported, SGLang can automatically fall back to transformers or you can explicitly set impl='transformers', reducing migration friction. The docs demonstrate usage with meta-llama/Llama-3.2-1B-Instruct and kyutai/helium-1-preview-2b, and note that the integration relies on features like RadixAttention and trust_remote_code for custom HF models, with ongoing performance optimizations and plans for LoRA and VLM support.