Skip to content

sdpa fp8 sm100: fix THD O-descriptor row stride (latent) #577

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
vedaanta merged 1 commit into from
Aug 13, 2026
Merged

sdpa fp8 sm100: fix THD O-descriptor row stride (latent) #577

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
6 changes: 5 additions & 1 deletion python/cudnn/sdpa/fwd/kernels/prefill_d128_fp8_sm100.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1925,14 +1925,18 @@ def _tma_swz(byte_w: int):
# THD: build the per-batch O descriptor array, then launch the exact
# flat batch-outermost grid (n_thd_units host-computed); grid_x = units*CGA_M.
# Works at cga1 (CGA_M=1) and cga2.
# Per-token element stride of packed O (QH * d_v) — NOT CFG.TILE_O,
# which is only coincidentally right at QH == 1 and otherwise lands
# every batch >= 1's descriptor base inside earlier batches' rows
# (the f16 kernel carries the same fix).
_build_o_descs_kernel(
o_tensor,
tma_o_desc,
o_desc_words,
seq_kv_lens_tensor,
cutlass.Int32(QH),
cutlass.Int32(B),
cutlass.Int32(CFG.TILE_O),
cutlass.Int32(o_tensor.stride[1]),
).launch(grid=(1, 1, 1), block=(32, 1, 1), stream=stream)
grid_shape = (n_thd_units * cutlass.Int32(CFG.CGA_M), cutlass.Int32(1), cutlass.Int32(1))
else:
Expand Down