biondizzle 4fed910c9c WIP: remove prepack cache, remap SFB per-call inside CUTLASS ...

- Deleted _prepack_weight_sf() and all cache/LRU logic
- L1/L2 pass raw scales, sfb_prepacked param removed
- cutlass_grouped_nvfp4_gemm always uses remap path
- README: big warning table explaining why prepack cache must not return
- Updated all doc references

NOT PUSHED YET — pending Mike review

2026-05-15 10:46:25 +00:00

..

nvfp4_megamoe_kernel

WIP: remove prepack cache, remap SFB per-call inside CUTLASS

2026-05-15 10:46:25 +00:00

nvfp4_megamoe_kernel.egg-info

Implement TileLang NVFP4 mega_moe L1/L2 kernels

2026-05-13 22:36:58 +00:00