a17dca508d
D1.3: Revert to zero-fill for sP - need to verify sP→PV pipeline first
2026-05-23 23:26:07 +00:00
5be5d42e94
D1.3: Compute (m,k) directly from thread mapping instead of identity tensor
2026-05-23 23:24:54 +00:00
23964d28c0
D1.3: Add debug prints for SMEM-P coordinate mapping
2026-05-23 23:24:02 +00:00
1e5635b93f
D1.3: Add SMEM-P coordinate diagnostic test
2026-05-23 23:23:05 +00:00
e0a11e32f8
D1.3: Fix coord extraction - identity tensor stores (m,k) pairs as values
2026-05-23 23:21:15 +00:00
a7171fa5e1
D1.3: Fix coordinate indexing - tTMEM_LOADcS first mode is (32,1) nested tuple
2026-05-23 23:20:12 +00:00
df7bc40d37
D1.3: Direct coordinate-indexed SMEM-P write using tTMEM_LOADcS coords
...
Each softmax thread writes its P values to sP using the (m,k) coordinates
from tTMEM_LOADcS. The k coordinate is decomposed into (k0,k1,k2) to
match sP's ((128,16),1,(4,2)) layout. CuTeDSL tensor indexing handles
the swizzle automatically. No make_tiled_copy needed.
2026-05-23 23:19:21 +00:00
2bbe55b08c
D1.3: Use make_cotiled_copy for SMEM-P — custom TV layout from TMEM-load coords to sP
...
Per CUTLASS guidance:
- make_tiled_copy_C/D encode wrong invariants for this transfer
- Build custom R→S copy where TV map comes from tTMEM_LOADcS (softmax thread
ownership) and destination addresses come from sP layout (PV A-operand swizzled SMEM)
- Use composition(sP_2d_layout, p_coord_layout) for atom_layout_tv
- Start with scalar BF16 (16-bit) stores — vectorize later
- Zero-fill source for compile test, will fill with actual P values next
2026-05-23 23:17:30 +00:00
2e86ed939e
Add SMEM-P guidance request document for CUTLASS LLM consultation
2026-05-23 23:03:35 +00:00
029c21a2af
D1.3: Use const_expr for lse None check
2026-05-23 22:30:55 +00:00
1720a0e86b
D1.3: Fix LSE with const_expr, always create valid mLSE tensor
2026-05-23 22:30:14 +00:00
bce31176aa
D1.3: Try make_tiled_copy_C(qk_mma) for SMEM-P copy - zero-fill source for compile test
2026-05-23 22:29:10 +00:00
c80bd021c9
D1.3: Define SMEM-P copy atoms unconditionally (CuTeDSL scoping)
2026-05-23 22:28:12 +00:00
43bb501acb
D1.3: Use full sP (4D) for make_tiled_copy_D partition
2026-05-23 22:27:11 +00:00
06fd2f63e9
D1.3: SMEM-P via get_smem_store_op + make_tiled_copy_D
...
Uses the CUTLASS blackwell_helpers pattern:
- get_smem_store_op creates a SMEM store atom paired with the TMEM load
- make_tiled_copy_D uses the same thread partition as the TMEM load
- Softmax warps write P to sP using the same thread mapping they use for reading S
- MMA warp reads P from sP via pv_mma.make_fragment_A(sP)
- Replaces the zero-fill stub with a proper register→SMEM copy
2026-05-23 22:26:09 +00:00
c507d0640c
D1.3: Enhanced diagnostic - test QK C-fragment as source for make_tiled_copy_C
2026-05-23 22:24:15 +00:00
bf896c0894
D1.3: Skip fragment creation in diagnostic, just print layouts
2026-05-23 22:21:31 +00:00
0c435b3e51
D1.3: Fix diagnostic - use dummy ptr 0 for shape analysis
2026-05-23 22:20:16 +00:00
55caf8be38
D1.3: Fix sP allocation - p_smem_s.outer is already a layout
2026-05-23 22:19:11 +00:00
d1c600f599
D1.3: Fix layout diagnostic - compute c_major outside kernel
2026-05-23 22:17:54 +00:00
d3d0020b4e
D1.3: Layout diagnostic v2 - run inside JIT-compiled kernel
2026-05-23 22:16:57 +00:00
ec8fd1474c
D1.3: Fix layout diagnostic - remove JIT-dependent code
2026-05-23 22:15:47 +00:00
c3a7c30f20
D1.3: Layout diagnostic - print all QK C-fragment and PV A-operand shapes
2026-05-23 22:14:35 +00:00
b59aca4655
Update all .md files with D5a/D5b progress, tOrP0 fix, LSE formula
...
- README.md: Updated Stage status table (D1 🟡 , D5 🟢 ), D5 section with
D5a/D5b results, tOrP0 bug fix docs, new CuTeDSL constraints #11-12
- STAGE_D1.3.md: Added progress update - TMEM-P works, SMEM-P still blocked,
recommended next steps
- STAGE_D.md was already updated
2026-05-23 22:07:53 +00:00
9264023e3b
Update STAGE_D.md with D5b results: merge cos 0.961, LSE err=0.0
2026-05-23 21:45:22 +00:00
2ced9d0da7
D5b: Fix reference computation - use logsumexp for stable LSE, fix o_unnorm definition
2026-05-23 21:43:04 +00:00
2883e042ca
D5b MILESTONE: SWA+sink merge works! cos 0.969
...
- Run FMHA twice (compressed KV + SWA KV) with normalized O + LSE
- Merge with sink weights in Python
- LSE err=0.0, merge cos=0.969 PASS
- Update STAGE_D.md: D5b done, D5c/D5d are optimizations
2026-05-23 21:36:26 +00:00
70763030c0
D5b: Use normalized O + LSE for merge (correct formula), always output LSE
2026-05-23 21:35:40 +00:00
28949da6e4
D5b: Clean up merge test - stable formula for both ref and kernel
2026-05-23 21:33:45 +00:00
9e1859827f
D5b: Use reference per-row LSE for proper O normalization
2026-05-23 21:31:52 +00:00
48d37d652e
D5b: Fix kernel_obj reference
2026-05-23 21:30:59 +00:00
caf89c65bf
D5b: Fix syntax error
2026-05-23 21:30:00 +00:00
3dd9cd6a94
D5b: Debug reference formula mismatch, add numerically stable merge
2026-05-23 21:19:25 +00:00
98390df27e
D5b: Python SWA+sink merge test
...
- Run FMHA twice (compressed KV + SWA KV, normalize=False)
- Merge with sink weights in Python
- Verify end-to-end correctness vs FP32 reference
2026-05-23 21:18:06 +00:00
60e03fe84a
Update STAGE_D.md: D5a done, CG-2/CG-3 status updated, tOrP0 offset rule added
2026-05-23 21:16:52 +00:00
edc283e6c1
D5a: Fix LSE formula - lse = ln(row_sum) + row_max * ln(2)
...
row_max is in scale_log2 domain, need to convert to natural log domain.
attn_max = row_max * ln(2), so lse = ln(row_sum) + row_max * ln(2).
2026-05-23 21:15:14 +00:00
6ca294ed6d
D5a: Use tensor indexing for LSE write
2026-05-23 21:13:52 +00:00
7e91d76669
D5a: Use cute.store for LSE write
2026-05-23 21:13:07 +00:00
751abd9b18
D5a: Fix LSE - compute row_max_safe from final row_max, remove mLSE None check
2026-05-23 21:12:29 +00:00
d6ea7f3ebd
D5a: Fix - add normalize param to __init__
2026-05-23 21:11:37 +00:00
c80f223d08
D5a: Add normalize flag + LSE output
...
- normalize=True (default): O = softmax(P) @ V (existing behavior)
- normalize=False: O = P @ V (un-normalized) + lse = log(row_sum) + row_max
- LSE tensor passed as optional parameter
- Test includes D5a normalize=False verification with LSE comparison
- Cleaned up SMEM-P debug prints and broken make_tiled_copy_C code
- hd=64 TMEM-P regression: cos 0.973 PASS
2026-05-23 21:10:40 +00:00
542bc7b1b0
D1.3: Use const_expr if for tOrP0 compile-time selection
2026-05-23 21:07:58 +00:00
37edd783ce
D1.3: Pre-compute tOrP0_offset in _setup, use const_expr for compile-time selection
2026-05-23 21:07:10 +00:00
972fbd48b9
D1.3: Use const_expr for tOrP0 offset (compile-time conditional)
2026-05-23 21:06:16 +00:00
a762820352
D1.3: Use MLIR-compatible expression for tOrP0 offset (same as Stage C)
2026-05-23 21:05:12 +00:00
4fa4239f95
D1.3: Initialize tOrP0 before conditional for CuTeDSL scoping
2026-05-23 21:03:53 +00:00
2bb3eb95ed
D1.3: Fix tOrP0 for SMEM-P - skip make_tensor when offset is 0
...
CuTeDSL doesn't support OpResult + int. When offset is 0 (SMEM-P),
just use tOrP directly.
2026-05-23 21:03:00 +00:00
eabea91b64
D1.3: Fix tOrP0 offset - scale FP32 columns to BF16 elements
...
tmem_p0_offset is in FP32 columns, but tOrP uses BF16 elements.
Offset = p0_offset * (32/16) = p0_offset * 2.
2026-05-23 21:02:04 +00:00
47eade4afc
D1.3: Fix CuTeDSL scoping - define tOrP0 unconditionally with p0 offset
2026-05-23 21:01:18 +00:00
0e81fc18aa
D1.3: Fix critical bug - add TMEM column offset for P0 in PV GEMM
...
The softmax warps store P at tmem_p0_offset=32. PV MMA must read from
the same offset. tOrP0 was missing the offset, causing PV to read from
TMEM column 0 (where S is) instead of column 32 (where P is).
This was the root cause of NaN/zeros in D1 tests.
2026-05-23 21:00:29 +00:00
