TMA store uses cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group
NOT mbarrier::complete_tx::bytes. Completion tracked via:
- cp.async.bulk.commit_group (after issuing stores)
- cp.async.bulk.wait_group.read 0 (wait for all groups)
Removed sMbarStore from SMEM allocations (no longer needed).
- fmha_multitile_capi.cu: C API wrapper for TMA multi-tile kernel
Creates TMA descriptors per (head, batch), launches kernel
- fmha_multitile_op.py: nvcc precompile + ctypes loader
- production.py: dispatch to multitile for N>128 or hd=512
- Reverted fmha_6warp_multihead.cuh to working single-tile version
- The TMA multi-tile kernel already passes 72 configs (D1.5)
HD=64/128/256/512 × T=1/4/32/128 × s_k=128/256/384/512
Single-tile path is IDENTICAL to the working pre-P5 kernel.
Multi-tile path uses FA2 online softmax with sOacc accumulator.
Runtime branch on is_multi_tile = (n_kv_tiles > 1).
- Kernel loops over KV tiles internally with running max/sum rescale
- SMEM accumulator sOacc[hd] replaces TMEM accumulation across tiles
- P is UN-NORMALIZED for multi-tile (exp(s-max), not /sum)
- Per KV tile: QK→softmax→PV→TMEM→read→add to sOacc
- Final: O = sOacc / running_sum
- Single tile (n_kv_tiles=1): same as before, no rescale
- Updated CAPI, Python loader, production.py fast path
- Added multi-tile test cases (N=256, 512)
P was already normalized in softmax step. PV = P_norm @ V gives the
correct attention output. Dividing by row_sum again in the epilogue
produces O = O_correct / row_sum (128x too small for uniform data).
- fmha_multihead_capi.cu: SMEM formula matches standalone test
Added cudaFuncSetAttribute for dynamic SMEM > 48KB
- fmha_multihead_op.py: pad K/V to N=128 when N<128
(kernel softmax loop is hardcoded to SK_TILE=128)
- Removed fmha_multihead_launch.cu (ATen approach, didn't work)
- Removed test_p3_ctypes_minimal.py (superseded by main test)
- fmha_multihead_capi.cu: pure C API wrapper, no ATen/pybind11 deps
- fmha_multihead_op.py: nvcc precompile + ctypes load (sm_100a)
- Removed fmha_multihead_launch.cu (ATen approach didn't work)
- Updated test to call kernel directly via ctypes API
- fmha_multihead_launch.cu: PyTorch launch wrapper for fmha_6warp_multihead_kernel
(c10::BFloat16 boundary, uint16_t bf16_t inside kernel, zero-cost casts)
- fmha_multihead_op.py: torch.utils.cpp_extension JIT loader + custom_op registration
(dsv4::fmha_multihead_decode for torch.compile)
- production.py: fast path dispatch for T=1, n_segments==1, hd in {64,128,256}
Falls through to CuTeDSL slow path for multi-segment/prefill
- test_p3_fast_decode.py: integration test (MHA/MQA/GQA, cosine >= 0.999998)
Architecture:
Grid: dim3(1, n_h, batch_size) — one CTA per (head, batch)
MQA: k_head_stride=0 so all Q heads share same K/V
Single kernel launch, zero cudaDeviceSynchronize on hot path
Normalized output for single-segment decode
V is the B operand with block_mn=16 in the PV MMA. Its canonical layout
uses CORES_MN=16/8=2, not 128/8=16. The previous code used CORES_MN=16
which produced wrong canonical indexing → garbage PV output.
Also:
- V SMEM size is (16,16) canonical = 256 BF16, not (128,16) = 2048
- P written as 16 elements at row 0 (T=1 decode)
- V loaded from TMA (16,128) and sub-sampled to (16,16) canonical
- V TMA coord: {col_start, d_base} for (HD,s_k) tensor
The MMA expects Q sub-tiles from a full (128,HD) canonical buffer,
but we were only loading (128,16) sub-tiles into a (128,16) buffer.
The MMA descriptor with block_mn=128 describes a (128,128) matrix,
reading 128 columns from SMEM but only 16 had real data.
Now: load all HD/16 TMA tiles of Q into a full (128,HD) canonical
buffer before the QK loop. The MMA reads the kt-th sub-tile via
descriptor offset kt * 128 * 32 bytes.
Also: share single sTmaBuf staging buffer for all TMA loads (Q, K, V).
Removed separate sQ_tma, sK_tma, sV_tma buffers.
write_smem_canonical used NTHREADS=192 as the stride, but in the TMA
kernel only the load warp (32 threads) calls it. With threadIdx.x in
[160,191] and stride 192, only 32 out of 2048 elements got written.
Fix: template STRIDE parameter, default 192, TMA kernel uses 32.
