Generalization Testing for Large-Model Frameworks

Developed a generalization testing framework for large-model training systems by decomposing models into reusable structural modules and generating semantically meaningful variants across model structure, parallelism, and optimization. Investigated configuration constraints spanning model, parallelism, optimization, and backend layers, distinguishing semantic requirements from implementation restrictions while preserving mutation intent during deep distributed execution.

Applied cross-framework differential testing across PTA and MindSpore-based stacks. Produced 52 issue reports, including 44 high-confidence suspected framework defects.