vLLM Significantly Improves Transformers Backend Performance

The ability to use Hugging Face's Transformers library as vLLM's modeling backend has been significantly improved. Now, model creators can use Transformers' implementations as-is to take advantage of vLLM's high-performance inference technology, without writing separate optimization code (porting).

Key updates and features are as follows:

• Performance improvement: In Qwen3 model benchmarks, the Transformers backend achieved throughput on par with or even higher than vLLM's existing native implementation.
• Automatic optimization: Supports maximizing hardware performance through GPU parallelization, compilation, and fused kernels.
• Ease of use: Can be applied instantly with just the --model-impl transformers flag, without changing existing vLLM settings.
• Versatility: Supports not only LLMs but also VLMs (Vision Language Models), and works in combination with continuous batching and custom attention kernels.