Hugging Face Transformers Backend for vLLM Achieves Native Performance

The Hugging Face Transformers library, which supports over 450 architectures, has been integrated as a modeling backend in vLLM since last year, enabling model authors to run Transformers models inside vLLM without porting. The latest iteration of this integration now achieves native vLLM inference speed for compatible models. Benchmarks comparing the Transformers backend head-to-head with vLLM's native implementations across three Qwen3 models (4B dense, 32B dense, and 235B FP8 MoE) show that the Transformers backend meets or beats native throughput in all cases. The performance improvement comes from dynamically applying inference-specific layer fusions at runtime, using torch.fx for static analysis and AST for source code manipulation. These fusions include operations mapped to vLLM's optimized kernels for expert parallelization in MoE models, as well as parallel linear layers for tensor and pipeline parallelism. The manipulated models remain fully compilable with torch.compile and CUDA Graphs. Any Hugging Face model can be run with the single flag --model-impl transformers, and it composes with existing parallelism options. Models with linear attention are not yet supported, and custom models from Hub repos are unlikely to work.