Port DeepSeek Sparse Attention to `MambaModel` - #3553

Merged

Phlip79 merged 12 commits into

NVIDIA:mainfrom

janEbert:mamba-dsa

Apr 17, 2026

Commits on Apr 17, 2026

Fix Mamba MLA handling
janEbert
committed
b004e05
Fix return type
janEbert
committed
786c50f
Improve error type
janEbert
committed
bfab53e
Fix DSA dispatch
janEbert
committed
67b145c
Port DeepSeek Sparse Attention to `MambaModel`
janEbert
committed
5b5a328
Future-proof `get_hybrid_layer_counts` return dict
janEbert
committed
69d7742
Prefer dict to list/tuple for large return values
janEbert
committed
44d344e
Do not treat DSA layers as Attention layers
janEbert
committed
25dc43b
Handle hybrid standard Attention/MLA models
janEbert
committed
75bbc7f
Add DSA Mamba tests
janEbert
committed
4258b92
Add DSA GPT/Mamba logprob equivalence tests
janEbert
committed
2e5f90e
Add DSA+MoE GPT/Mamba equivalence tests
janEbert
committed
6fc1c7c