vllm - 💡(How to fix) Fix [Feature]: NVFP4 MoE LoRA support (Cutlass)

🚀 The feature, motivation and pitch

Feature Proposal & Motivation

Can we add support for NVFP4 MoE models with LoRA adapters in vLLM. Currently, there are issues tracking MoE + LoRA for models like Qwen3.5, but there is no Lora NVFP4 quantization support. Current NVFP4 lora support only uses the Marlin kernel which does not do native FP4 computation on supported hardware.

Motivation:

• NVFP4 improves throughput on latest hardware.
• Lora allow for efficient fine-tuned models.

Alternatives

No response

Additional context

• Related vLLM issues: #40005, #38520, #40038
• Relevant model families: Qwen3.6-MoE
• NVFP4 formats available via HuggingFace and quantization libraries
• Happy to help with test cases, model links, and validation if needed.

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.