Skip to content

fix(dsa): align packed-CP indexer causal masks #16

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
JackRao123 merged 3 commits into from
Jul 14, 2026
Merged

fix(dsa): align packed-CP indexer causal masks #16

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
545ba95
fix(dsa): pass q_causal_offsets to the cuDNN indexer in packed-CP top…
JackRao123 Jul 9, 2026
ccce860
docs(dsa): clarify packed CP causal offset comments
JackRao123 Jul 13, 2026
a1fab1b
docs(dsa): remove redundant causal offset comments
JackRao123 Jul 13, 2026
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
17 changes: 16 additions & 1 deletion megatron/core/transformer/experimental_attention_variant/dsa_cudnn_kernels.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -612,8 +612,19 @@ def _indexer_topk_from_score_chunks(
None if score_seq_lens is None else score_seq_lens[row_start:row_end].contiguous()
)
if bottom_right_key_start is not None:
q_causal_offsets = torch.full(
(b,),
bottom_right_key_start + row_start,
dtype=torch.int32,
device=q_chunk.device,
)
scores_chunk = _cudnn_dsa.indexer_forward_wrapper(
q_chunk, score_k_bshd, w_chunk, ratio=indexer_ratio, sm_scale=_INDEXER_SOFTMAX_SCALE
q_chunk,
score_k_bshd,
w_chunk,
ratio=indexer_ratio,
sm_scale=_INDEXER_SOFTMAX_SCALE,
q_causal_offsets=q_causal_offsets,
)["scores"]
elif score_seq_lens is None and row_start == 0 and row_end == sq:
scores_chunk = _cudnn_dsa.indexer_forward_wrapper(
Expand Down Expand Up @@ -765,6 +776,9 @@ def _indexer_topk_multi_packed_cp_thd(
max_segment_q = packed_max_seqlen_q // segment_divisor
max_k_half = packed_max_seqlen_k // segment_divisor
max_segment_k = max((cp_rank + 1) * max_k_half, packed_max_seqlen_k - cp_rank * max_k_half)
segment_q_causal_offsets = (segment_k_lengths - segment_q_lengths).to(
dtype=torch.int32, device=device
)
scores = _cudnn_dsa.indexer_forward_wrapper(
q_bshd[0],
segmented_k,
Expand All @@ -775,6 +789,7 @@ def _indexer_topk_multi_packed_cp_thd(
cu_seqlens_k=segment_cu_k,
max_seqlen_q=max_segment_q,
max_seqlen_k=max_segment_k,
q_causal_offsets=segment_q_causal_offsets,
)["scores"]

segment_topk = min(topk_k, max_segment_k)
Expand Down