[None][feat] Add TransformerEngine FP8 attention backend for VisualGen - #17849

Open

wu6u3tw wants to merge 11 commits into

NVIDIA:mainfrom

wu6u3tw:feat/vg-te-attention-backend

Commits on Aug 28, 2026

[None][test] Add LSE correctness test for TEAttention.forward_with_lse
wu6u3tw
committed
f692c29
[None][fix] Fix test_attention_te import path and remove invalid .cuda() call
wu6u3tw
committed
fd415e8
[None][fix] Compute LSE manually in TEAttention.forward_with_lse
wu6u3tw
committed
ad1e608
[None][fix] Address review findings in TEAttention
wu6u3tw
committed
327ede6
[None][test] Add Attention2D x72 single-stream tests for the TE backend
wu6u3tw
committed
9bd2032
[None][fix] Cache TE attention ops per trait and stop swallowing configs
wu6u3tw
committed
ceed0c3
[None][fix] Drop TEAttention.forward_with_lse and consolidate its tests
wu6u3tw
committed
9e9b438
[None][fix] Declare transformer_engine as a dependency
wu6u3tw
committed
35ddf5c
[None][test] Do not register the TE tests in B200 CI
wu6u3tw
committed
2b9dafd
[None][fix] Use current scaling, not delayed scaling, for TE FP8 attention
wu6u3tw
committed
f04142d
[None][fix] Put TE's DotProductAttention in eval mode
wu6u3tw
committed
4063693