biondizzle

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

2026-05-28 23:01:14 +00:00

d54bce6a6d fix: correct SMEM size for MMA 4-warp test

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

2026-05-28 23:00:29 +00:00

be45e87891 test: MMA→4-warp TMEM read — do warps see different rows?

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

2026-05-28 22:59:33 +00:00

6b0d57074a test: TMEM cross-warp visibility with different sync strategies

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

2026-05-28 22:58:50 +00:00

77d190278e test: simpler TMEM 4-warp read — direct store+load

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

2026-05-28 22:58:01 +00:00

91b03bd6bd test: verify 4-warp TMEM read with 32x32b.x8 after MMA

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

2026-05-28 22:56:30 +00:00

28e04a5ea8 fix: use __cvta_generic_to_shared directly for 64-bit compat

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

2026-05-28 22:56:17 +00:00

1d6a95df32 fix: typo in tmem row offset test

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

2026-05-28 22:56:01 +00:00

cf6fe71368 test: verify TMEM 32x32b.x8 row offset addressing

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

2026-05-28 22:53:51 +00:00

4cfb707405 fix: correct SMEM size calculation in multirow test

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

2026-05-28 22:52:53 +00:00

863a030c3b fmha_multirow: rewrite with 32x32b.x8 only, no s_p_vals, row_page addressing

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

2026-05-28 21:08:17 +00:00

1ba304db3e stuff

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

2026-05-28 20:13:54 +00:00

deaa3ec725 CRITICAL FIX: Q/K SMEM canonical layout must use local d (0..15) not full_d — UMMA descriptor reads from sQ0/sK0 start, not offset

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

2026-05-28 20:10:18 +00:00

08694b8136 Fix multi-row softmax v3: 32x32b.x8 with per-lane per-row (no wmax/wsum), per-row sRowMax/sRowSum arrays

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

2026-05-28 20:08:32 +00:00

aaa76c1af1 Rewrite multi-row softmax using 16x256b.x1 TMEM reads for proper multi-row access

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

2026-05-28 20:06:17 +00:00

5e3c61184c Fix multi-row softmax: remove cross-lane wmax/wsum — each lane handles its own row independently

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

2026-05-28 20:05:02 +00:00

bf4dfd131b Fix nvcc goto-bypasses-init: move var decls before goto targets

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

2026-05-28 20:04:46 +00:00

2b09d4f2ef Fix nvcc goto-bypasses-init in multi-row test

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

2026-05-28 20:04:31 +00:00

d8b421ccee Multi-row FMHA kernel (Milestone 4): T>1 prefill support with 4-warp parallel softmax

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

2026-05-28 19:35:07 +00:00

adc88613fa Milestone 5 COMPLETE: multi-head FMHA grid launch verified on B200

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

2026-05-28 19:33:06 +00:00

3fd302e7a0 Fix nvcc goto-bypasses-init errors in multi-head test