Nvfp4 four over six grouped experts - #9

Draft

wolfcomos wants to merge 6 commits into

nvfp4-four-over-six-converterfrom

nvfp4-four-over-six-grouped-experts

Commits on Aug 24, 2026

Add NVFP4 four-over-six grouped-experts converter and RL recipes

wolfcomos
and
claude
committed
ca0326d
Slice grouped activations to offs[-1] for four-over-six expert GEMMs
wolfcomos
committed
da8674a
Disable CUDA graphs in the four-over-six DSv3 recipes
wolfcomos
committed
0532b02
Reject compile with row-scaled four-over-six grouped experts
wolfcomos
committed
c259be8
Count four-over-six converters in has_quantization
wolfcomos
committed
8906831
Unit-test the compile rejection and has_quantization coverage
wolfcomos
committed
8717093