FP4 weight stages in shared memory

Packed FP4 halves each weight stage, so one SM's shared memory holds 11 pipeline stages instead of 7.

Schematic, not to scale. Weight stages of the routed-expert GEMM in one SM's shared memory.

Both bars have the same length; stage counts as reported, stage widths schematic.

Weight stage, 1 byte per value (SM100) Weight stage, packed FP4 (SM107) Other buffers