Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[sharktank] Use reshape instead of view #681

Merged
merged 2 commits into from
Dec 13, 2024
Merged

[sharktank] Use reshape instead of view #681

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4db7338
[sharktank] Use reshape instead of view
marbre Dec 12, 2024
31e7cab
Merge branch 'main' into mmdit
marbre Dec 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion sharktank/sharktank/layers/mmdit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,7 +41,7 @@ def attention(q, k, v, pe):
q=q, k=k, v=v, a=None, is_causal=True, scale=None
)
x = ops.permute(x, (0, 2, 1, 3))
x = x.view(x.shape[0], x.shape[1], -1)
x = x.reshape(x.shape[0], x.shape[1], -1)

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I played a bit with this example. What is going on is that the scaled dot product attention is returning an non-contiguous tensor.
In torch 2.5.1

x.shape = torch.Size([3, 24, 1536, 128]), x.stride() = (4718592, 196608, 128, 1)

While in torch 2.3.0 it produces

x.shape = torch.Size([3, 24, 1536, 128]), x.stride() = (4718592, 128, 3072, 1)

I guess by accident in 2.3.0 the stride was such that it allowed for the view.
I think on a lot of operations we implicitly assume that they will produce contiguous tensors. It seems that the doc does not state this, but I would bet that even if this was the intention on a lot of ops, it would not be stated.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for reviewing and testing!

return x

Expand Down
Loading