verl
verl copied to clipboard
[sglang, rollout] fix: use right padding for response_position_ids
Fixes #4159
- Changed response_position_ids padding from left to right
- Ensures alignment with response_ids for variable-length sequences
- Critical for 2D position_ids in multimodal models (e.g., Qwen2-VL)
- Added comprehensive unit tests covering 1D and 2D cases
The bug caused misalignment when batch contained sequences of different lengths, as response_ids used right padding but response_position_ids used left padding.