biondizzle

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

2026-05-22 17:26:56 +00:00

8740ab5b27 Stage C: manual kv_coord + correct K GMEM slice + O rescale fence

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

2026-05-22 17:24:39 +00:00

f39c3a8e38 Add example4: manual kv_coord Int32 for GMEM tile indexing

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

2026-05-22 17:09:54 +00:00

711480a4fb README: add test harness instructions

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

2026-05-22 17:08:15 +00:00

83f8a13add run_test.sh: SIGKILL all children of screen session on cleanup

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

2026-05-22 17:07:25 +00:00

f4cfeae262 Add check_log.sh convenience script

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

2026-05-22 17:06:02 +00:00

61122e10c6 Fix quoting in run_test.sh

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

2026-05-22 17:05:45 +00:00

af475affab Add run_test.sh harness (screen + log)

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

2026-05-22 16:57:35 +00:00

d0626e0434 FIX: only slice GMEM tensors (SMEM already 2D from tma_partition)

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

2026-05-22 16:56:41 +00:00

15a2fdadbc FIX: consistent GMEM/SMEM slicing for K and V TMA partitions

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

2026-05-22 16:51:59 +00:00

5c1423b2c5 FIX: keep GMEM iteration dimension FREE in TMA K/V partition slices

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

2026-05-22 16:47:10 +00:00

018c09644b Add diagnostic test for multi-tile TMA pipeline (identity softmax)

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

2026-05-22 16:42:50 +00:00

ec895e831e FIX: acc_scale was double-multiplying by scale_log2

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

2026-05-22 16:39:45 +00:00

2147cce95d Stage C: integrate example3 multi-tile fixes into unit test

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

2026-05-22 16:32:35 +00:00

dad35818f3 README + MEMORY: update Stage C status to single-tile only, document multi-tile blocker

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

2026-05-22 16:23:39 +00:00

e4c82873bb FMHA Stage-C multi-tile: combined K+V barrier, final_o_bar, acc_pipe producer

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

2026-05-22 15:54:05 +00:00

39fa5b96b0 restore tBgK to kh.count indexing (single-tile working), add TODO for multi-tile

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

2026-05-22 15:52:58 +00:00

f1854bab26 FIX: use unsliced tBgK with (None, kt, None, 0) for proper GMEM tile indexing

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

2026-05-22 15:52:09 +00:00

3d2cb0e52b CRITICAL FIX: keep GMEM iteration dim free in tBgK/tVgV slice

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

2026-05-22 15:49:49 +00:00

a04b219f0f add explicit acc_pipe.consumer_wait before final normalize

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

2026-05-22 15:41:21 +00:00

ff27a261b1 FMHA Stage-C multi-tile: Fix 1 (s_k=n), Fix 2 (TMA kt indexing), Fix 3 (O rescale)