Skip to content

fix: fix lite csa.py code doesn't match megatron-lm dev rescontruct code #29

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

Open
chenjiaoAngel wants to merge 2 commits into
base: lite
Choose a base branch
from
Open

fix: fix lite csa.py code doesn't match megatron-lm dev rescontruct code #29

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4f87e08
fix: fix lite csa.py code doesn't match megatron-lm dev rescontruct code
chenjiaoAngel Aug 13, 2026
17309bf
Merge branch 'lite' of https://github.com/chenjiaoAngel/Megatron-LM i…
chenjiaoAngel Aug 31, 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
10 changes: 5 additions & 5 deletions experimental/lite/megatron/lite/primitive/modules/attention/csa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,11 +9,11 @@
# Zero-copy imports of the DSv4 THD-CP helpers that live in Megatron Core. The
# lite CSA module reuses Core's differentiable kernels, CP row-mapping utilities,
# and CuTeDSL layout kernels rather than vendoring them; see the module docstring
# of ``csa_cp_utils`` / ``csa_cp_layout_kernels`` for the exact contracts.
# of ``cp_utils`` / ``cp_layout_kernels`` for the exact contracts.
from megatron.core.tensor_parallel.mappings import gather_from_sequence_parallel_region
from megatron.core.transformer.experimental_attention_variant import (
csa_cp_layout_kernels,
csa_cp_utils as cp_utils,
from megatron.core.transformer.experimental_attention_variant.csa_utils import (
cp_layout_kernels,
cp_utils,
)
from megatron.core.transformer.experimental_attention_variant.csa import (
_unfused_indexer_sparse_attn_from_topk,
Expand Down Expand Up @@ -977,7 +977,7 @@ def _forward_thd_cp(
else (max_seqlen_q // ratio if ratio > 1 else 0)
)
topk_idxs, topk_length, indexer_topk_rank_major = (
csa_cp_layout_kernels.build_attention_indices(
cp_layout_kernels.build_attention_indices(
cu_seqlens,
global_start,
l_local,
Expand Down