biondizzle 0ac73a82f9 fix: L1 output uses unpacked E2M1 (1 byte/element) like FP8 ...

- float_e2m1_unpacksmem_t: sizeof=1, SMEM is 1 byte/element (not packed)
- TMA load unpacks 2 E2M1/global-byte → 2 SMEM bytes
- UMMA reads unpacked SMEM, packs internally for mxf4nvf4
- L1→L2 handoff: unpacked format (same byte count as FP8)
- Epilogue: 4 E2M1 bytes per uint32 STSM atom, same as FP8
- Dispatch TMA: kHidden bytes (unpacked), not kHidden/2
- Added static_assert on sizeof(a_dtype_t) and sizeof(b_dtype_t)
- Note: no bandwidth savings at L1→L2 boundary for v1

2026-05-11 21:27:35 +00:00

..

attention.hpp

Add various optimizations and Mega MoE benchmarks (#316)

2026-04-24 18:41:37 +08:00

einsum.hpp

[Public release 26/04] Introducing Mega MoE, FP4 Indexer and other features/fixes (#304)

2026-04-17 09:45:14 +08:00

gemm.hpp

[Public release 26/04] Introducing Mega MoE, FP4 Indexer and other features/fixes (#304)

2026-04-17 09:45:14 +08:00

hyperconnection.hpp

[Public release 26/04] Introducing Mega MoE, FP4 Indexer and other features/fixes (#304)

2026-04-17 09:45:14 +08:00

layout.hpp

fix: gran_k=16 in transform_sf + sm_100a arch for NVFP4 mega_moe

2026-05-11 16:11:11 +00:00

mega_nvfp4.hpp

fix: L1 output uses unpacked E2M1 (1 byte/element) like FP8

2026-05-11 21:27:35 +00:00

mega.hpp

Add various optimizations and Mega MoE benchmarks (#316)

2026-04-24 18:41:37 +08:00

runtime.hpp

[Public release 26/04] Introducing Mega MoE, FP4 Indexer and other features/fixes (#304)

2026-04-17 09:45:14 +08:00