Add native DeepSeek-V4 hybrid attention orchestration - #6402

Draft

FDecaYed wants to merge 9 commits into

NVIDIA:mainfrom

FDecaYed:codex/dsv4-v2-d2-20260810

Commits on Aug 10, 2026

Add selective FP32 parameter lifecycle support
FDecaYed
committed
a02802d
fix offload

hxbai
authored and
FDecaYed
committed
e2dd1ca
Restore optimizer chain import
FDecaYed
committed
3afc50e
Reuse native FP32 parameter IDs
FDecaYed
committed
effa326
Fix optimizer checkpoint mapping for frozen parameters
FDecaYed
committed
a1df227
Make MLA RoPE application safe and explicit
FDecaYed
committed
e051716
Handle asymmetric packed RoPE sequence lengths
FDecaYed
committed
94ac9e4
Add unfused SBHD compressed sparse attention
FDecaYed
committed
c463cbe
Add native DSv4 hybrid attention orchestration
FDecaYed
committed
0ecb394