biondizzle

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

2026-05-28 15:17:43 +00:00

bafd26707b FMHA HD=64 with BLOCK_MN_B=16, 4 N-tiles per K-tile

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

2026-05-28 15:16:29 +00:00

6896d1aebb Update CURRENT_ISSUE: HD=16 done, HD=64 in progress

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

2026-05-28 15:15:24 +00:00

6b9b06647a Clean up HD=64 debug prints, keep register-math PV check

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

2026-05-28 15:13:49 +00:00

5c9d471162 Add register-math PV reference for HD=64 debug

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

2026-05-28 15:12:22 +00:00

43e9efbc2b Fix string literal

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

2026-05-28 15:11:17 +00:00

906be7ce50 Add filtered cosine (exclude near-zero)

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

2026-05-28 15:09:57 +00:00

40c83c769a Fix: remove ×2 QK scale correction (MMA scale is 1.0, not 0.5)

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

2026-05-28 15:07:56 +00:00

6ea7356fdd Debug: print P values for HD=64

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

2026-05-28 15:06:39 +00:00

4b052f22a5 Fix: opt into >48KB shared memory for HD=64

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

2026-05-28 15:03:28 +00:00

7becbfc07e Fix: printf after var declarations

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

2026-05-28 15:02:01 +00:00

2d44f8e356 Debug: check if HD=64 kernel starts

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

2026-05-28 14:58:13 +00:00

46e4d07c71 Test PV SS MMA with B=(64,16) BLOCK_MN=64

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

2026-05-28 14:56:08 +00:00

465e089a2b Add launch error check for HD=64

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

2026-05-28 14:54:13 +00:00

2fd64c464d FMHA HD=64 with BLOCK_MN_B=64 for V, proper output dimensions

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

2026-05-28 14:52:32 +00:00

15ecc1f616 Full FMHA HD=64 with PV SS MMA (SMEM-P)

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

2026-05-28 14:50:44 +00:00

5b2e690936 Milestone: Full FMHA HD=16 with PV SS MMA (SMEM-P) — cosine 0.9997

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

2026-05-28 14:49:21 +00:00

78026839b7 Fix V canonical layout: swap g_mn/g_k indices (d=MN, lr=K)

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

2026-05-28 14:47:12 +00:00

9a3b43c42b Fix reference to also use uniform P

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

2026-05-28 14:46:23 +00:00

75bdcbf728 Debug: override P with uniform 1/128

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

2026-05-28 14:45:16 +00:00

af93c283c7 Enable all 8 PV K-tiles