[PyT] [Common] Enable sm120 support for fused attn if cuDNN is 9.18.1+#2693

Draft

KshitijLakhani wants to merge 12 commits intoNVIDIA:mainfrom

KshitijLakhani:klakhani/maint/sm120-thd-flash-support

Commits on Mar 11, 2026

Add support for sm120 correct batch, seq dims
KshitijLakhani
committed
Add support for sm120 BHS1 style max logit even QKV are THD to avoid incorrect max logit calculation (includes padded tokens in max calculation)
KshitijLakhani
committed
Disable fused and flash attn for sm120 filter:kv cache
KshitijLakhani
committed
For CP P2P attn, set softmax_lse_in_packed_format to False if sm120+
KshitijLakhani
committed
Assert in TE if T3HD/TH3D layout is used on sm120 before cuDNN F16 sdpa arbitrary kernel call
KshitijLakhani
committed
Modify is_ragged_q && cudnn_runtime_version >= 90600 check to also include a check for sm120
KshitijLakhani
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
nit: Code clean up
KshitijLakhani
committed