Implemented native tensor parallelism for custom multi-axis Transformers composed with FSDP, packed attention/MLP projections with checkpoint conversion and adapter support, and grouped-GEMM layouts for unaligned adapter dimensions. See Training Frameworks.