a2d063a48b
D1: N-tile support for HEAD_DIM>256
...
- pv_n_tile = min(head_dim, 256) — MMA instruction N limit
- n_pv_tiles = head_dim // pv_n_tile — outer loop count
- V FMHA layout uses pv_n_tile (not head_dim) for N-tile slicing
- Test loops over N-tiles at Python level, kernel processes (128, pv_n_tile)
- For hd=512: 2 kernel launches with V[:,0:256] and V[:,256:512]
2026-05-23 03:22:23 +00:00
7bc097163d
d1: add hd=512 test
2026-05-23 03:20:46 +00:00
32995c2ba3
d1: add quick regression test (hd=64 only)
2026-05-23 03:20:12 +00:00
eed981bee5
D1: Parameterize HEAD_DIM in FmhaKernel (64→512)
...
- Promote HEAD_DIM from module constant to constructor parameter
- FmhaKernel(head_dim=64, s_k=128, ...) — default 64 for regression
- All references to HEAD_DIM replaced with self.head_dim
- PV MMA tiler, V layout, softmax corr_tiles all parameterized
- TMEM budget warning when num_tmem_alloc_cols > 512
- New test: test_fmha_v3_stage_d1.py tests hd=64 (regression) and hd=512
- Stage C test preserved as-is for reference
2026-05-23 03:19:52 +00:00
1a6c5e3822
docs: revised Stage D/E plan — indexer removes paged TMA, one kernel for CSA/HCA/SWA, sink merge
2026-05-23 03:10:41 +00:00
a846193c4a
cleanup: remove archive/ (240 stale files), stale example9/10, fix test table, add Stage D plan
2026-05-23 03:05:08 +00:00
f3d0d67ae9
docs: update README with Stage C TMEM layout mismatch findings and status
2026-05-23 03:01:04 +00:00
9c331de7ba
fix: revert to composition layout for hand-constructed atoms (matching CUTLASS)
2026-05-23 02:54:54 +00:00
3a2d3c66da
fix: use logical_divide (not composition) for O rescale/normalize atoms to match get_tmem_load_op layout
2026-05-23 02:53:59 +00:00
3aba5cc6da
fix: add NO-OP TMEM round-trip to re-map O from MMA to epilog layout
2026-05-23 02:50:53 +00:00
45cf89a556
fix: use TMEM round-trip normalize + epilogue_tma_store (known ~3% error)
2026-05-23 02:49:46 +00:00
350c7c36ac
fix: correct bSG_gC indexing (6 modes)
2026-05-23 02:45:30 +00:00
6318b4da29
diag: print bSG shapes for TMA store indexing
2026-05-23 02:44:47 +00:00
28060dd944
fix: typo from_dlcap -> from_dlpack
2026-05-23 02:44:00 +00:00
048a546e76
fix: correction_epilog with paired atoms + pre-partitioned TMA store outside if block
2026-05-23 02:41:07 +00:00
0700745852
test: NO-OP round-trip + normalize at n=128 and n=256
2026-05-23 02:37:50 +00:00
2ebfcb2278
fix: correction_epilog with paired atoms + pre-partitioned TMA store
2026-05-23 02:34:33 +00:00
49bf6e8294
diag: NO-OP round-trip before normalize on 2D pattern
2026-05-23 02:32:40 +00:00
6cf1f17904
fix: O rescale uses 2D register tensor pattern, remove fence_view_async_tmem_load
2026-05-23 02:31:28 +00:00
7842d86294
fix: use paired atoms for correction_epilog + cute.copy TMA store
2026-05-23 02:26:57 +00:00
1f4e40decc
diag: add CUDA_LAUNCH_BLOCKING for crash debug
2026-05-23 02:25:46 +00:00
728a24db6a
fix: inline epilogue_tma_store with inv_row_sum multiply using paired atoms
2026-05-23 02:24:36 +00:00
0ecde542f1
fix: use cute.copy instead of cpasync.copy for TMA store
2026-05-23 02:23:16 +00:00
702bf8aa29
fix: correction_epilog with get_tmem_load_op paired atoms + direct TMA store
2026-05-23 02:19:41 +00:00
ea66b6ee8d
diag: NO-OP TMEM round-trip test — load+store back unchanged
2026-05-23 02:15:28 +00:00
6ee28d8423
fix: inline epilogue with paired atoms + inv_row_sum normalize, no TMEM round-trip
2026-05-23 02:13:52 +00:00
043b66406a
fix: all epilogue warps do TMA store, no dynamic if inside method
2026-05-23 01:41:36 +00:00
db3572bafb
fix: correction_epilog with get_tmem_load_op paired atoms, no TMEM round-trip
2026-05-23 01:40:13 +00:00
d99a90ade5
fix: use attn_raw (not softmax'd) for unnorm computation
2026-05-23 01:36:27 +00:00
7becdaf739
diag: skip kernel normalize, do Python-side normalize to isolate TMEM round-trip issue
2026-05-23 01:35:18 +00:00
039c8b90ce
diag: print expected unnorm P@V for comparison with raw kernel output
2026-05-23 01:28:32 +00:00
ec5b892e32
diag: skip final normalize, test raw PV output via epilogue_tma_store
2026-05-23 01:27:03 +00:00
e50644afde
fix: O rescale uses 2D register tensor pattern (matching CUTLASS correction_rescale)
2026-05-23 01:25:53 +00:00
b77c8d83f5
fix: pre-compute tmem_load_epi_atom in __call__, pass to kernel
2026-05-23 01:24:33 +00:00
c9271ffbf4
fix: index into TMA partitioned tensors for copy
2026-05-23 01:23:04 +00:00
e01ff282b7
fix: use flat_divide+group_modes(0,2) for TMA store, matching CUTLASS
2026-05-23 01:22:22 +00:00
5efa9c9297
fix: use gC not tCgC for TMA partition, group modes 0-3
2026-05-23 01:20:52 +00:00
7a894c4bf6
fix: use tma_partition for TMA store in correction_epilog
2026-05-23 01:20:09 +00:00
3c134f7e90
fix: replace TMEM round-trip normalize with CUTLASS correction_epilog pattern
2026-05-23 01:18:56 +00:00
690fd77e6c
diag: inv_row_sum=1.0 to test raw PV, n=128 only
2026-05-23 01:17:14 +00:00
2b93b10199
diag: test original code n=128+256 to confirm baseline
2026-05-23 01:13:29 +00:00
9bcddb68e1
diag: disable O rescale properly, test n=128+256 baseline
2026-05-23 01:12:50 +00:00
0ef41266de
diag: test n=128 and n=256 both with rescale disabled
2026-05-23 01:12:00 +00:00
dc44fa187a
fix: indentation error in diag disable
2026-05-23 01:11:25 +00:00
a9cace316d
diag: disable O rescale to isolate the issue (n=256 only)
2026-05-23 01:11:00 +00:00
08d4af90ca
debug: add wide-search diagnostics for n=256 O rescale
2026-05-23 01:02:33 +00:00
f026c1824c
🚀 MULTI-TILE SOFTMAX + O RESCALE WORKING: n=128 cos 0.999998, n=256 cos 0.80
...
Fixed ALL loops to use self.n_kv_tiles (Python int) instead of
cute.size(gK, mode=[3]) which returned 1 for all n values.
Results:
n=128: cos 0.999998 ✅ PASS (single tile, full softmax + normalize)
n=256: cos 0.801156 (2 tiles, O rescale partially working)
n=512: CUDA launch failure (pipeline can't cycle past kv_stage=2)
The n=256 improvement (0.71 → 0.80) confirms:
1. TMA fix (None,0,None,0) loads both KV tiles correctly
2. Softmax processes both tiles with online row_max/row_sum tracking
3. O rescale (O *= acc_scale for kt > 0) is partially working
4. Final normalize (O *= 1/row_sum) works correctly
Remaining:
- n=256 cos 0.80 → 0.9999: O rescale precision issue
- n≥384: pipeline cycling (kv_stage=2 can only hold 2 tiles)
- Need to increase kv_stage or fix pipeline state cycling
2026-05-23 00:35:42 +00:00
d511ebe387
Debug: add row_sum/inv_row_sum printf at final normalize
2026-05-23 00:34:38 +00:00
c2ff8e072e
Fix ALL loops: use self.n_kv_tiles everywhere
...
The MMA loop (cutlass.range) and MMA consumer loop (range) also used
cute.size(gK, mode=[3]) which returns 1 for all n. Fixed all 3 loops:
1. TMA load loop (cutlass.range, line 215)
2. MMA consumer loop (range, line 231)
3. Softmax loop (range, line 324)
This was causing the deadlock — MMA only produced S[0] while softmax
waited for S[1].
2026-05-23 00:33:38 +00:00
a63f452c86
Fix softmax loop: use self.n_kv_tiles not cute.size(gK, mode=[3])
...
cute.size(gK, mode=[3]) returns 1 for ALL n values — mode 3 is batch,
not KV tiles. self.n_kv_tiles = s_k // 128 is the correct Python int.
This is why softmax only processed kt=0 for all n.
2026-05-23 00:30:49 +00:00
