biondizzle

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

2026-05-28 12:09:47 +00:00

c2e41a858e test: force 2 K-tiles for debug

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

2026-05-28 12:07:51 +00:00

8b2200a6d3 test: HD=64 full 4 K-tile accumulate + full-HD scalar reference

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

2026-05-28 12:04:56 +00:00

afb18caf2d test: clean HD=64, 1 K-tile only, verify SMEM writes + compare vs scalar

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

2026-05-28 12:01:29 +00:00

e587e26b06 test: log canonical indices we write Q to

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

2026-05-28 11:59:10 +00:00

facd509c3c test: remove sanity check (zeroing loop overwrites), fix verify offsets

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

2026-05-28 11:57:15 +00:00

20ae390d32 test: fix compile error

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

2026-05-28 11:56:09 +00:00

7b16eceb91 test: more detailed SMEM sanity check

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

2026-05-28 11:54:16 +00:00

eb0ca18e23 test: sanity check sQ[0] write+read

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

2026-05-28 11:52:53 +00:00

8936a2dec7 test: clean SMEM write loops for HD=64

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

2026-05-28 11:51:10 +00:00

2ffbfda47d test: print SMEM verify data

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

2026-05-28 11:49:46 +00:00

4fd41365de test: add SMEM verify for HD=64 K-tile offsets

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

2026-05-28 11:47:57 +00:00

4483539f01 test: HD=64 random data, 4 K-tiles, accumulate

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

2026-05-28 11:46:14 +00:00

73bd21ce01 test: force 1 K-tile for HD=64 debug

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

2026-05-28 11:44:33 +00:00

abe1870429 test: HD=64 all-ones, expected S[0,j]=64 (unscaled) or 8.0 scaled

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

2026-05-28 11:42:32 +00:00

73f9ff98c9 test: UMMA QK HD=64 (4 K-tiles, accumulate) — multi-K-tile test

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

2026-05-28 11:41:20 +00:00

df34cae9c6 UMMA QK GEMM WORKING! Update docs — 4x was scale factor, not bug

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

2026-05-28 11:39:17 +00:00

1874a70a6d test: fix var ref

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

2026-05-28 11:38:32 +00:00

8426d13285 test: fix comparison — row 0 is S[0,c], rows 1-127 should be zero

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

2026-05-28 11:36:47 +00:00

6f40fafa91 test: verify ALL 128 rows × 8 cols match scalar reference

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

2026-05-28 11:34:46 +00:00

3c7d9d9303 test: apply 1/sqrt(HD) scale to MMA output — 4x was the scale factor, not a bug!