Summary for 2026-04: Delivered a critical bug fix to the Mxfp4 scale shuffle path in ROCm/aiter to restore correct GEMM reads when K is not divisible by 256. Implemented layout changes in _dynamic_mxfp4_quant_kernel_asm_layout to align the row-group stride with the kernel’s width (scaleN_pad) and padded out-of-bounds scale values with 0, eliminating misreads that previously propagated across most output rows. The patch, tracked in commit d6cefa97836633edbfe92aa8a42b6f453586304d (#2900), was co-authored by Wei Zhu and Claude Sonnet. This work enhances reliability of quantized inference paths and reduces risk in production deployments, while maintaining clean code with explicit commit messages for traceability.