Transformers v5 Notes

This section documents VeOmni’s integration with HuggingFace transformers==5.2.0 (the only supported transformers version).

Included Notes

• Flash Attention custom-name handling: explains why _lazy_imports fails for VeOmni custom attention names and how the local hub-kernel loader adapter resolves it.

• Patchgen workflow: explains the modeling code generation workflow used for every supported model and how to regenerate.

• MoE weight loading: explains how VeOmni expects MoE expert weights to be laid out and documents qwen3_moe handling.

• Testing a new model: SOP for adding test cases in test_models_patch.py and test_e2e_parallel.py when onboarding a new model.

• VeOmni Flash Attention Custom Name Adapter (Transformers 5.x)
• Modeling Code Generation
• Transformers v5 MoE Weight Loading
• Testing a New Model