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.
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).
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.
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
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.
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.
Both GMEM and SMEM sides must be sliced to the same rank for cute.copy.
K (QK MMA B-partition): slice [(None,None,0,0)] keeps modes 0,1
- mode 1 = GMEM iteration, indexed by kvh.count
V (PV MMA B-partition): slice [(None,0,None,0)] keeps modes 0,2
- mode 2 = GMEM iteration, indexed by kvh.count
Q: only 1 tile, (None,0,None,0) hardcode is fine.
Root cause of multi-tile failure: (None,0,None,0) slice hardcodes the
GMEM tile dimension to 0, so TMA always loads from tile 0 regardless
of kvh.count. K from QK MMA has GMEM iter at mode 1, V from PV MMA
has it at mode 2 (different layouts: K,D,L vs D,K,L).
Fix follows CUTLASS FMHA reference:
- K: tBgK[(None,None,None,0)] + tBgK[(None, kvh.count, None)]
- V: tVgV[(None,0,None,0)] + tVgV[(None, kvh.count)]
row_max is already in log2(scaled) space (S * scale_log2), so
old_row_max - row_max_safe is the correct exponent for exp2.
The old code computed exp2(scale_log2 * (old_row_max - row_max_safe))
which is exp2(scale_log2^2 * (old_max_S - new_max_S)) — wrong.
- Combined K+V barrier (one acquire per kt, kvh.count == kt)
- O rescale for kt > 0 (online softmax O correction)
- final_o_bar sync (MMA signals before producer_tail)
- s_k as constructor param (compile-time for V layout)
- kv_tx_bytes covers both K and V transfers
- Test covers n=128, 256, 512, 1024
Key changes from Mike:
1. Combined K+V TMA barrier: one acquire per kt, both cute.copys share
kvh.barrier. kvh.count naturally == kt (no interleaving problem).
tx_count = K_bytes + V_bytes. Also fixes the sK[0]/sV[1] slot quirk.
2. final_o_bar NamedBarrier: MMA .arrive() after acc_pipe.producer_tail;
softmax .arrive_and_wait() before reading O for normalize. Prevents
softmax racing MMA's PV[N-1] on the final O read.
3. acc_pipe producer in MMA: producer_acquire before loop, commit+advance
after loop, producer_tail after. Consumer in epilogue as before.
4. O rescale re-enabled for kt>0 with acc_scale before softmax_done_bar.
CuTeDSL TMA copy API doesn't support dynamic GMEM tile indexing.
kh.count works for single tile. For multi-tile, need to either:
1. Map pipeline count to tile index (kh.count // 2 for interleaved K/V)
2. Separate K and V into non-interleaved TMA loops
3. Use gK/gV layouts that iterate naturally with pipeline count
This is the architectural blocker for multi-tile FMHA.
The pre-slice (None,0,None,0) hardcoded GMEM iteration to tile 0.
Instead, keep the original tBgK and index with (None, kt, None, 0)
inside the TMA loop, where kt selects the correct GMEM tile.
This preserves 2D rank matching with the SMEM tensor.
The slice (None,0,None,0) was hardcoding the GMEM iteration dim to 0,
meaning TMA always loaded K/V from tile 0 regardless of kt.
Changed to (None,None,None,0) to keep gmem_iter free,
then index with (None, kt, None) in the TMA copy loop.
This is the root cause of multi-tile failure: TMA was always reading
the first 128 tokens for ALL KV tiles.
Race condition: softmax reads O to normalize while MMA may still be
writing PV[N-1]. Single-tile wins by luck; multi-tile drifts.
Move acc_cons_st construction before the wait so epilogue reuses it.
Fix 1: s_k must equal actual n. With s_k < n, v_fmha layout only spans
first s_k V tokens and TMA reads OOB on later tiles.
Fix 2: TMA producer indexes K and V by kt (loop variable), NOT by the
pipeline's interleaved count. The kv pipeline interleaves K and V, so
pipeline count goes 0,1,2,3 but GMEM tiles should be K[0],V[0],K[1],V[1].
Fix 3: Online O rescale before softmax_done_bar. When row_max grows,
O must be multiplied by exp2(old_max - new_max) before MMA starts next PV.
row_max is in scaled domain (s_val * scale_log2). The O rescaling
should be exp2(old_max - new_max) without extra scale_log2 because
the max values already include the scaling factor.
row_max should be the max of the raw QK scores, not pre-scaled.
The scale_log2 is applied during exp2 and rescaling, not stored in row_max.
This fixes the double-scaling bug that broke multi-tile O rescaling.
- Move O TMEM load/store setup before softmax loop
- After P store: rescale O in TMEM by exp2((old_max - new_max) * scale)
- Only rescale for kt > 0 (first tile has no prior O to rescale)
- Use same TMEM load/modify/store pattern as final normalization
- Test both n=128 (1 tile) and n=256 (2 tiles)
MMA must wait for softmax to produce P in TMEM before starting PV.
Without this, MMA reads stale P data from TMEM, causing deadlock.
softmax_done_bar: softmax warps arrive after P store, MMA waits before PV.
- Add acc_bar to SS struct
- Create acc_pipe (full pipeline) before if blocks
- Pass acc_pipe to epilogue_tma_store (needs full pipeline, not participant)
