fix: optimize attention decomposition - #4448
Open
fs-eire wants to merge 3 commits into
Open
Conversation
Restore the original -inf value and invalid-mask polarity while retaining scalar Select inputs required for compact TensorRT MHA masks.
github-actions
Bot
added
component: tests
narendasan
requested review from
micwill755 and
zewenli98
and removed request for
lanluo-nvidia
July 31, 2026 19:11
micwill755
reviewed
Aug 3, 2026
| assert attn_mask is None, "attn_mask must be None when is_causal=True" | ||
| attn_bias = torch.zeros((L, S), dtype=query.dtype, device=device) | ||
| temp_mask = torch.ones((L, S), dtype=torch.bool, device=device).tril(diagonal=0) | ||
| attn_bias = attn_bias.masked_fill(temp_mask.logical_not(), float("-inf")) |
Collaborator
There was a problem hiding this comment.
For the bool attn_mask path we switched to scalar where(0 / -inf), but is_causal still uses masked_fill. Should the causal path use the same scalar where pattern?
zewenli98
reviewed
Aug 4, 2026
| fill_layer.set_input(1, start_tensor) | ||
| fill_layer.set_input(2, step_tensor) | ||
| set_layer_name(fill_layer, target, f"{name}_arange_fill", source_ir) | ||
| return fill_layer.get_output(0) |
Collaborator
There was a problem hiding this comment.
Good finding! Is this arange op upstream of attention? If so, it should resolve the perf issue for both IAttention path and Decomposed path, right?
Besides, I have an addition to arange converter. Can you review #4456?
Contributor
Author
There was a problem hiding this comment.
Is this arange op upstream of attention? If so, it should resolve the perf issue for both IAttention path and Decomposed path, right?
yes
The change should be good. I think you can merge it first and then I can combine my change based on your change.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Align attention decomposition behavior with the ONNX graph.
This PR does the following:
With this change, Torch-TensorRT is able to generate optimized engine with a pattern that an attention using a causal mask AND a padding mask, when
decompose_attention=True.Fixes # (issue)
Type of change
Checklist: