183292d919
O normalize using tmem_ptr base (same as epilogue) + CUTLASS sub-tile pattern
2026-05-22 19:18:16 +00:00
365e8f53af
O normalize with full layout (no sub-tiling), Repetition(64)
2026-05-22 19:17:03 +00:00
7bc94e610a
Disable ALL O copies to verify baseline
2026-05-22 19:14:33 +00:00
590d6e9fba
Disable O rescale too for NO-OP test
2026-05-22 19:13:43 +00:00
51cec1405d
DEBUG: O load+store NO-OP to verify TMEM copy correctness
2026-05-22 19:12:43 +00:00
e3e2668192
Re-enable O rescale + normalize with corr_tile_size=32
2026-05-22 19:11:17 +00:00
d77c965646
Disable O rescale + normalize, verify softmax P only
2026-05-22 19:09:33 +00:00
dcc64dd14d
FIX: O sub-tile count should be HEAD_DIM/corr_tile_size, not 128/corr_tile_size
2026-05-22 19:06:52 +00:00
48b24ba005
Full pipeline: O rescale + final normalize with CUTLASS sub-tile approach
2026-05-22 19:05:59 +00:00
a85894df89
Test softmax P vs unnormalized reference (no O normalize)
2026-05-22 19:05:27 +00:00
c0b39fc2bf
O normalize using CUTLASS reference sub-tile approach
2026-05-22 19:04:37 +00:00
3dbda0eebb
Fix O normalize: use 2D register tensor indexing
2026-05-22 18:59:15 +00:00
6b61d5274c
Add O normalization with sub-tile TMEM read-modify-write
2026-05-22 18:58:35 +00:00
b936c6220d
Simplify: softmax P only, no O rescale/normalize yet
2026-05-22 18:57:00 +00:00
e2fad84205
Real softmax test built on working identity diag
2026-05-22 18:55:31 +00:00
d3b662d3a8
CRITICAL FIX: remove extra scale_log2 in softmax (minus_row_max and acc_scale)
2026-05-22 18:52:58 +00:00
32869c7378
FIX: K slice (None,None,0,0) like working diag
2026-05-22 18:51:22 +00:00
4b1fc7ee1f
Diag: identity softmax on example6 pipeline to isolate softmax bug
2026-05-22 18:49:33 +00:00
912f92c6b5
Quick test: working v3 with n=256 multi-tile
2026-05-22 18:47:54 +00:00
5b6392beaa
DEBUG: add version marker to confirm code changes are running
2026-05-22 18:47:05 +00:00
c7d55a5f49
CRITICAL FIX: TMA pre-slice (None,0,None,0) → (None,None,0,0) to keep GMEM tile dim free
2026-05-22 18:45:55 +00:00
f734610268
Diag: TMA shapes with hardcoded major modes
2026-05-22 18:43:55 +00:00
18a589347c
Diag: simplified TMA shape analysis
2026-05-22 18:43:30 +00:00
7ad4ddb6ba
Diag: print TMA partition shapes for multi-tile debugging
2026-05-22 18:43:05 +00:00
67c5a0928d
FIX: Use Python range() in TMA warp for concrete per-iteration GMEM coords
2026-05-22 18:26:31 +00:00
54de81985f
FIX: Force SSA GMEM coord via n_kv_tiles - n_kv_tiles instead of cutlass.range kt
2026-05-22 18:25:13 +00:00
f3503fc1ee
FIX: TMEM offset bug in O rescale/normalize — use tOtO0.iterator not tOtO.iterator
2026-05-22 18:23:46 +00:00
0b7ae7c969
Diag: test n=384 (3 tiles) to find crash boundary
2026-05-22 18:07:07 +00:00
640ec3e96e
Diag: test all sizes 128-1024
2026-05-22 18:06:28 +00:00
02d993ecac
DEBUG: disable O rescale to isolate NaN cause
2026-05-22 18:05:46 +00:00
1c3970fe58
Add NaN/inf checking to stage C test
2026-05-22 18:01:11 +00:00
d7a0fc2bc2
CRITICAL FIX: K GMEM slice (None,None,0,0) not (None,0,None,0)
...
K from QK MMA B-partition has GMEM iter at mode 1, NOT mode 2.
(None,0,None,0) hardcodes mode 1 to 0 → TMA always loads tile 0.
(None,None,0,0) keeps mode 1 free → correct multi-tile loading.
Proof: diag n=256 went from cos 0.711 → 0.999999 with this one change.
2026-05-22 17:59:57 +00:00
b6a2904e93
Diag: try K slice (None,None,0,0) keeping mode 1 (CUTLASS ref style)
2026-05-22 17:59:01 +00:00
01621e1520
Diag: try runtime Int32(0+0) for kv_coord with cutlass.range
2026-05-22 17:57:58 +00:00
beecc4df47
Diag: use Python range() unrolling like stage C test
2026-05-22 17:56:59 +00:00
200430bd3f
Fix diagnostic test: same Int32(kt) + n_kv_tiles fixes
2026-05-22 17:56:15 +00:00
c23ebd5b57
Try cutlass.range with Int32(kt) — now n_kv_tiles is Python int
2026-05-22 17:51:25 +00:00
4a41df51c4
FIX: n_kv_tiles as Python int (s_k//128) for range() unrolling
...
cute.size() returns a CuTeDSL symbol, not a Python int.
range() on a symbol can't iterate — the loop never unrolls.
Now n_kv_tiles is computed in __init__ as s_k // 128 (Python int).
2026-05-22 17:50:07 +00:00
70409636f7
Option 2: Python range() with Int32(kt) for TMA GMEM coord
...
cutlass.range traces once - kv_coord/kt are trace-time values,
not runtime loop-carried state. Python range() fully unrolls at
trace time, emitting distinct Int32(k) constants per iteration.
Int32(1) hardcoded already proved TMA CAN load from tile 1.
2026-05-22 17:47:43 +00:00
b55a38c4c3
Add example5: use cutlass.range induction variable as TMA GMEM coord
2026-05-22 17:47:10 +00:00
aacad257ea
README: add fire_b200_test docs, update multi-tile blocker with real findings
2026-05-22 17:41:23 +00:00
93c28b9c29
Clean up debug prints, set kv_coord as Int32(0)
...
Key findings to relay to CUTLASS LLM:
- kv_coord=Int32(1) hardcode CHANGES the output (TMA CAN load from different tiles)
- kv_coord=Int32(0) + kv_coord += 1 does NOT increment at runtime
(all multi-tile outputs identical to kv_coord=0)
- kv_coord=0 (plain Python int) also doesn't work
- Pipeline handle .count doesn't work either
- The TMA GMEM tile coordinate must be dynamic at kernel runtime,
but CuTeDSL appears to constant-fold or not propagate the increment
2026-05-22 17:39:27 +00:00
1bba851911
DEBUG: try plain Python int kv_coord (like CUTLASS ref)
2026-05-22 17:34:30 +00:00
15b2a28d29
DEBUG: hardcode kv_coord=1 to test if TMA uses it
2026-05-22 17:32:53 +00:00
ff9ef6dcde
DEBUG: try K slice (None,0,None,0) keeping mode 2 free
2026-05-22 17:30:06 +00:00
cec6f59d66
DEBUG: print tBgK/tVgV shapes before/after slice
2026-05-22 17:28:45 +00:00
8740ab5b27
Stage C: manual kv_coord + correct K GMEM slice + O rescale fence
...
Key fixes:
1. GMEM tile coord: manual Int32 kv_coord (not kvh.count)
2. K GMEM slice: (None,None,0,0) keeps mode 1 free (GMEM iter)
3. V GMEM slice: (None,0,None,0) keeps mode 2 free (GMEM iter)
4. Add fence_view_async_tmem_load before O rescale for visibility
2026-05-22 17:26:56 +00:00
f39c3a8e38
Add example4: manual kv_coord Int32 for GMEM tile indexing
...
Pipeline handle .count is NOT a GMEM coordinate — it's opaque pipeline
state. CUTLASS FMHA reference uses a manual Int32 counter (kv_coord)
for GMEM and handle.index for SMEM. .count is never used as a
coordinate anywhere in the reference.
2026-05-22 17:24:38 +00:00
711480a4fb
README: add test harness instructions
2026-05-22 17:09:53 +00:00
83f8a13add
run_test.sh: SIGKILL all children of screen session on cleanup
...
Deadlocked GPU processes ignore SIGHUP from screen -X quit.
Now kills the entire process group with SIGKILL, plus a catch-all
pkill for any python test_ processes.
2026-05-22 17:08:12 +00:00
