This pull request updates the way the text model is referenced and used in the generate_text_embeddings method, improving clarity and robustness. The most important change is using the unwrapped model's text_model attribute directly, which helps avoid ambiguity and potential errors when handling different model wrappers.

Improvements to model handling:

• Changed the model reference in generate_text_embeddings from a hardcoded string to dynamically use unwrap_model(self.model).text_model, ensuring the correct model is used and improving maintainability.<!--
  Thank you 🙏 for your contribution to Ultralytics 🚀! Your effort in enhancing our repositories is greatly appreciated. To streamline the process and assist us in integrating your Pull Request (PR) effectively, please follow these steps:

1. Check for Existing Contributions: Before submitting, kindly explore existing PRs to ensure your contribution is unique and complementary.

2. Link Related Issues: If your PR addresses an open issue, please link it in your submission. This helps us better understand the context and impact of your contribution.

3. Elaborate Your Changes: Clearly articulate the purpose of your PR. Whether it's a bug fix or a new feature, a detailed description aids in a smoother integration process.

4. Ultralytics Contributor License Agreement (CLA): To uphold the quality and integrity of our project, we require all contributors to sign the CLA. Please confirm your agreement by commenting below:

   I have read the CLA Document and I sign the CLA

For more detailed guidance and best practices on contributing, refer to our ✅ Contributing Guide. Your adherence to these guidelines ensures a faster and more effective review process.
-->

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Fix YOLOE trainer text-embedding cache key to use the actual configured text_model instead of a hardcoded string 🧠

📊 Key Changes

• Updated generate_text_embeddings() to derive model from unwrap_model(self.model).text_model rather than the hardcoded "mobileclip:blt"
• Cache filename now reflects the active text model, preventing cache collisions across different configurations 📦
• Removed a redundant assert self.model is not None before calling unwrap_model(self.model).get_text_pe(...)

🎯 Purpose & Impact

• Ensures text-embedding caching is correct when training YOLOE from scratch with different text model settings ✅
• Avoids silently reusing incompatible cached embeddings that could degrade training quality or cause confusing behavior 🔍
• Improves robustness/consistency by relying on the model’s actual attributes rather than constants 🛠️