vllm - 💡(How to fix) Fix [Feature]: Adding VLA models support into vLLM

🚀 The feature, motivation and pitch

Background

vLLM already provides a high-throughput inference and serving stack for LLMs and multimodal language models. However, Vision-Language-Action (VLA) models used in robotics and autonomous driving are not currently supported as a first-class model family in vLLM.

VLA models extend VLMs by taking multimodal observations such as images, language instructions, and robot/vehicle state, and returning actions or trajectories rather than only text tokens. Supporting this class of models in vLLM would enable:

• high-throughput batched rollouts for robotics simulation, evaluation, and RL;
• a unified serving backend for VLA models instead of one-off policy servers;
• reuse of vLLM’s scheduler, request queueing, tensor/pipeline parallelism, metrics, and OpenAI-compatible serving infrastructure;
• a clean extension point for future action-output models.

There is already implementation work in SGLang that can be used as a reference for scoping and parity testing:

• Alpamayo-R1 VLA support: https://github.com/sgl-project/sglang/pull/21059
• π0 VLA support: https://github.com/sgl-project/sglang/pull/23504
• GR00T-N1.7 VLA support: https://github.com/sgl-project/sglang/pull/23589
• pi05 VLA support: https://github.com/sgl-project/sglang/pull/24455

Related SGLang tracking issue:

• https://github.com/sgl-project/sglang/issues/18266

There is also a previous vLLM issue requesting VLA/OpenVLA support, but it was closed as not planned:

• https://github.com/vllm-project/vllm/issues/14739

This issue proposes a scoped path to add first-class VLA inference support to vLLM, initially targeting four models that already have SGLang implementations.

Alternatives

No response

Additional context

Relevant SGLang work:

• VLA tracking issue: https://github.com/sgl-project/sglang/issues/18266
• Alpamayo-R1: https://github.com/sgl-project/sglang/pull/21059
• π0: https://github.com/sgl-project/sglang/pull/23504
• GR00T-N1.7: https://github.com/sgl-project/sglang/pull/23589
• π0.5 / pi05: https://github.com/sgl-project/sglang/pull/24455

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.