Jul 12

1. 21:36

   fp4-train: add fp4_mm, the native Blackwell NVFP4 tensor-core matmul (10x over the dequant path)

   CharlesCNorton
   3be46d1
2. 12:02

   Correct fp4-train card: SR/EF beat round-to-nearest in NVFP4 weight-storage training

   CharlesCNorton
   3cd3528
3. 01:42

   Add fp4-train source, tests, and model card

   CharlesCNorton
   7a139da