April 2025 monthly summary for NVIDIA/TensorRT-LLM: delivered FP8 layout optimizations for Blackwell to improve performance and flexibility on SM100, and strengthened FusedMoE routing reliability with router input weight support and correct expert weight loading. Added tests to validate smaller M dimensions and ensure robustness. Result: higher FP8 throughput, more reliable MoE routing, and improved test coverage.

Monthly summary for 2025-01: NVIDIA/physicsnemo delivered Graph Partitioning: Matrix Decomposition, introducing a 1D row-wise decomposition of the adjacency matrix to reduce overhead and improve efficiency in distributed graph processing for square adjacency matrices. Implemented in commit b7b6265b9b2e8f82ee239cffdee4464065c843bb ([Feature] Add row-decomposition of adj. matrix to reduce graph partitioning overhead (#720)). Impact includes improved scalability, reduced inter-partition communication, and faster distributed runs. No major bugs fixed this month. Technologies demonstrated include graph algorithms, matrix decomposition, distributed processing, and performance optimization.