Skip to content

[https://nvbugs/6902355][fix] Keep TopK rows in the DSA sparse MLA test KV pool #20283

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
lfr-0531 merged 1 commit into from
Oct 16, 2026
+6 −0
Merged

[https://nvbugs/6902355][fix] Keep TopK rows in the DSA sparse MLA test KV pool #20283

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: 6 additions & 0 deletions tests/unittest/_torch/attention/sparse/dsa/test_dsa_sparse_mla.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -777,6 +777,12 @@ def yarn_get_mscale(scale=1, mscale=1):
* kv_cache_tokens_per_block
* max_num_contexts
)
# The static token-sparse context FMHA may load KV rows beyond the selected ones, and its
# TMA descriptor does not bound rows to the pool. A pool sized only for these short
# sequences (a single 64-token block for context_sequence_lengths=[10]) can therefore be
# read past its end, which faults whenever the following address range is unmapped. Keep
# at least TopK rows in the pool, as any real deployment does.
max_tokens = max(max_tokens, SPARSE_TOPK)

pretrained_config = SimpleNamespace(
rms_norm_eps=1e-6,
Expand Down