Support for "Qwen3" architecture and vLLM versions >= 0.11.0

[HardenGale]

Problem Description

I am trying to fine-tune a model based on the Qwen3 architecture (e.g., Qwen3-0.6B) using Verl, but I encountered ValueError related to unsupported architecture and vLLM version compatibility.

1. Unsupported Architecture: The model config uses "model_type": "qwen3", which is not present in Verl's model registry whitelist.
2. Unsupported vLLM Version: Verl throws an error because vLLM 0.11.0 is not in the supported version list (max 0.6.3).

Error Logs

Error 1: Model Architecture

ValueError: Model architecture qwen3 is not supported for now. RMPad supported architectures: dict_keys(['llama', 'mistral', 'gemma', 'qwen2']).
Please set `use_remove_padding=False` in the model config.

Error 2: vLLM Version

ValueError: vllm version 0.11.0 not supported. Currently supported versions are 0.3.1, 0.4.2, 0.5.4 and 0.6.3.

Environment

• Python: 3.9
• vLLM: 0.11.0
• Transformers: 4.57.3
• Model: Qwen3-0.6B