[model] feat: add registration and config converter for Qwen 2.5-Omni

[martinzhang03]

What does this PR do?

This PR introduces the initial registration and configuration conversion logic for the Qwen 2.5-Omni model within the veRL framework.

Key Changes

• Model Registration: Added QWEN2_5_OMNI to the SupportedVLM and SupportedModel enums in verl/models/mcore/registry.py.

• Config Converter: Implemented hf_to_mcore_config_qwen2_5_omni in verl/models/mcore/config_converter.py.

  • This converter specifically addresses the nested configuration structure of the Omni model: Qwen2_5OmniConfig → thinker_config → text_config.
  • It correctly extracts mrope_section from rope_parameters to support multimodal rotary positional embeddings (mRoPE).

• Forward Registry Skeleton: Added commented-out entries for MODEL_FORWARD registries as placeholders for the upcoming implementation.

Implementation Details

The implementation was developed by:

• Analyzing the Qwen2_5OmniForConditionalGeneration structure in the Hugging Face transformers library.
• Referencing established patterns in ms-swift for Omni model training support to ensure compatibility with existing community standards.

Future Work (Next Steps)

This PR serves as the foundation for full Qwen 2.5-Omni support. Immediate follow-up PRs will include:

• Weight Converter: Implementation to map HF checkpoints to Megatron-Core tensors.
• Model Initializer: Implementation of the initialization logic.
• Forward Activation: Enabling the Forward Registries (standard, no-pad, and fused).

Technical Summary (中文)

本次 PR 实现了 Qwen 2.5-Omni 模型在 veRL 中的基础注册与配置转换逻辑：

• 嵌套配置解析：成功处理了 Omni 模型特有的 thinker_config 嵌套结构，准确提取核心文本参数。
• mRoPE 支持：从配置中提取并映射了多模态旋转位置编码参数。
• 架构对齐：参考了 ms-swift 的实现经验，确保与现有社区多模态训练标准保持一致。

[CLAassistant]

All committers have signed the CLA.

[wuxibin89]

We're now use mbridge and NVIDIA-NeMo/Megatron-Bridge to convert hf model to mcore GPTModel, please submit a PR to them.

[martinzhang03]

We're now use mbridge and NVIDIA-NeMo/Megatron-Bridge to convert hf model to mcore GPTModel, please submit a PR to them.

Thanks for the heads-up, @wuxibin89! I'll port the configuration logic over to that repository and submit a PR there instead. Should I keep this PR open for a moment in case the registration part (the Enum changes) is still needed here, or should I close this entirely?

[wuxibin89]

We're now use mbridge and NVIDIA-NeMo/Megatron-Bridge to convert hf model to mcore GPTModel, please submit a PR to them.

Thanks for the heads-up, @wuxibin89! I'll port the configuration logic over to that repository and submit a PR there instead. Should I keep this PR open for a moment in case the registration part (the Enum changes) is still needed here, or should I close this entirely?

Close this PR since we're going to cleanup mcore model registration in verl #4496 #4530