biondizzle

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:57:51 +00:00

fe0588d906 fix: simplify UMMA dump script

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:55:44 +00:00

948a3f8a7a add UMMA descriptor dump script

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:53:37 +00:00

e5ba0ca119 debug: clean QK verify with scalar sanity + MMA result

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:49:38 +00:00

a04d794979 debug: skip TMEM alloc — test SMEM loads only

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:48:10 +00:00

72c97f2546 debug: minimal UMMA descriptor (just start_addr + version)

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:47:00 +00:00

9a51bfa578 fix: align SMEM layout properly (128B aligned tmem + Q)

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:44:56 +00:00

2a765be715 fix: correct SMEM size for row-major (not swizzled)

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:43:41 +00:00

c64bd7b875 debug: read Q/K directly from SMEM

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:41:32 +00:00

58b610c96c fix: proper early return for SMEM load test

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:40:19 +00:00

82bc2c4a49 debug: verify SMEM loads + scalar QK sanity check

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:39:13 +00:00

53139d24bf debug: verify TMEM r/w works before MMA

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:38:10 +00:00

a9d71ff6ab debug: print TMEM values after MMA

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:37:03 +00:00

bfb1e177ce debug: try all-lane MMA + print tmem_base

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:35:32 +00:00

d3510980e4 feat: SWIZZLE_NONE UMMA descriptors with row-major SMEM

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:23:39 +00:00

8c67c31497 add CuTe descriptor printing script

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:20:57 +00:00

d29d6b575f add UMMA descriptor diagnostic script

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:18:48 +00:00

ab84ad0f86 feat: implement canonical UMMA SMEM layout with SWIZZLE_128B

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:07:55 +00:00

ecbc75255c fix: correct UMMA descriptor format from CUTLASS source

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:03:55 +00:00

fe7d561143 debug: print UMMA descriptor values for diagnosis

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-28 08:02:56 +00:00

c5f7a9a15c fix: align SMEM buffers to 16 bytes for UMMA descriptors