Worked on stabilizing the model merge workflow for the Qwen3-VL-2B series in the volcengine/verl repository by addressing critical errors in configuration handling and tokenizer creation. Applied Python programming and error handling skills to resolve issues that previously caused failures during PyTorch FSDP-based model merges. The solution reduced manual intervention and enabled faster, more reliable integration of Qwen3-VL models, directly supporting downstream teams that depend on consistent model deployment. Demonstrated expertise in model merging, configuration management, and the tokenizer pipeline, contributing to a smoother release process and improved onboarding for teams relying on the volcengine/verl codebase.