Skip to content

dsa(indexer_backward): restore range_constexpr in 8 kernel_gemm epilogue loops #549

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Anerudhan merged 1 commit into from
Aug 11, 2026
Merged

dsa(indexer_backward): restore range_constexpr in 8 kernel_gemm epilogue loops #549

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 8 additions & 8 deletions python/cudnn/deepseek_sparse_attention/indexer_backward/indexer_backward_sm100.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1074,7 +1074,7 @@ def _compute_warpgroup(
s_full_1_phase = Int32(0)

dw_accum = cute.make_rmem_tensor(tSrS_shape, Float32)
for ei in cutlass.range(cute.size(dw_accum), unroll_full=True):
for ei in cutlass.range_constexpr(cute.size(dw_accum)):
dw_accum[ei] = Float32(0.0)

tSrS = cute.make_rmem_tensor(tSrS_shape, Float32)
Expand Down Expand Up @@ -1129,16 +1129,16 @@ def _compute_warpgroup(

# Phase 2: Convert dS f32→bf16, write to sdS via coordinate mapping.
tSrS_f16 = cute.make_rmem_tensor(tSrS.shape, self.q_dtype)
for ei in cutlass.range(cute.size(tSrS), unroll_full=True):
for ei in cutlass.range_constexpr(cute.size(tSrS)):
tSrS_f16[ei] = self.q_dtype(tSrS[ei])

if bi % 2 == 0:
for ei in cutlass.range(cute.size(tSrS_f16), unroll_full=True):
for ei in cutlass.range_constexpr(cute.size(tSrS_f16)):
h = cute.get(tCcS[ei], mode=[0, 0])
n = cute.get(tCcS[ei], mode=[0, 1])
sdS_gemm_view_0[h, n] = tSrS_f16[ei]
else:
for ei in cutlass.range(cute.size(tSrS_f16), unroll_full=True):
for ei in cutlass.range_constexpr(cute.size(tSrS_f16)):
h = cute.get(tCcS[ei], mode=[0, 0])
n = cute.get(tCcS[ei], mode=[0, 1])
sdS_gemm_view_1[h, n] = tSrS_f16[ei]
Expand All @@ -1159,7 +1159,7 @@ def _compute_warpgroup(
cute.copy(tiled_tmem_load_dq, tDqDq_t2r, tDQrDQ)

tDQrDQ_bf16 = cute.make_rmem_tensor(tDQrDQ.shape, self.q_dtype)
for ei in cutlass.range(cute.size(tDQrDQ), unroll_full=True):
for ei in cutlass.range_constexpr(cute.size(tDQrDQ)):
tDQrDQ_bf16[ei] = self.q_dtype(tDQrDQ[ei] * Float32(sm_scale))

cute.arch.fence_view_async_tmem_load()
Expand All @@ -1169,7 +1169,7 @@ def _compute_warpgroup(
sdQ_epi_slice,
cute.make_layout((self.heads_padded, self.head_dim_padded)),
)
for ei in cutlass.range(cute.size(tDQrDQ_bf16), unroll_full=True):
for ei in cutlass.range_constexpr(cute.size(tDQrDQ_bf16)):
h = cute.get(tCcDQ[ei], mode=[0, 0])
d = cute.get(tCcDQ[ei], mode=[0, 1])
sdQ_gemm_view[h, d] = tDQrDQ_bf16[ei]
Expand All @@ -1188,7 +1188,7 @@ def _compute_warpgroup(
for h_local in cutlass.range_constexpr(HEADS_PER_WARP):
h = warp_base_h + h_local
my_partial = Float32(0.0)
for ei in cutlass.range(cute.size(dw_accum), unroll_full=True):
for ei in cutlass.range_constexpr(cute.size(dw_accum)):
if cute.get(tCcS[ei], mode=[0, 0]) == h:
my_partial = my_partial + dw_accum[ei]
total = cute.arch.warp_reduction_sum(my_partial)
Expand Down Expand Up @@ -1257,7 +1257,7 @@ def _reduce_warpgroup(
# local→global when topk_indices_global=False); gmem fallback
# mirrors that conversion via const_expr branch.
batch_offset_l2g = Int32(0) if const_expr(self.topk_indices_global) else batch_idx * (seqlen_k // batch_size)
for pair in cutlass.range(cute.size(tDKrDK) // 2, unroll_full=True):
for pair in cutlass.range_constexpr(cute.size(tDKrDK) // 2):
ei = pair * 2
n = cute.get(tCcDK[ei], mode=[0, 0])
d = cute.get(tCcDK[ei], mode=[0, 1])
Expand Down