enable Qwen3-VL-MoE model quantization, save and vllm loading

[WeiweiZhang1]

For the Qwen3-VL-MoE models (e.g., Qwen/Qwen3-VL-30B-A3B-Instruct), the fused MoE architecture (similar to LLaMA 4 and GPT-OSS) requires additional support for quantization.

Since MoE parameters constitute the majority of the model, this significantly affects the achievable compression ratio.

To ensure compatibility, expert quantization should follow the VLLM-supported tensor shape for proper model loading.

Reference quantized model and example:

Model: QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ

Example : vllm-project/llm-compressor — qwen3-vl-30b-a3b-Instruct-example.py

[wenhuach21]

https://huggingface.co/Intel/Qwen3-VL-235B-A22B-Instruct-int4-AutoRound

[WeiweiZhang1]

https://huggingface.co/Intel/Qwen3-VL-235B-A22B-Instruct-int4-AutoRound

So, we won’t modify the modeling logic inside the AutoRound API. Instead, we’ll provide a solution that users can follow to manually adjust the architecture for FusedMoE quantization, right? If this is the final decision, this issue can be closed now.

[wenhuach21]

no, you could also support the converting in AutoRound.