Skip to content

Make svd_trunc pullback GPU friendly #271

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

Open
kshyatt wants to merge 1 commit into
base: main
Choose a base branch
from
+3 −2
Open

Make svd_trunc pullback GPU friendly #271

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
5 changes: 3 additions & 2 deletions src/pullbacks/svd.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -241,8 +241,9 @@ function svd_trunc_pullback!(
Y₀ᴴ = iszerotangent(ΔV₊ᴴ) ? zero(Vᴴ) : ldiv!(Diagonal(S), ΔV₊ᴴ)
US = mul!(ΔAV, U, Smat) # recycle ΔAV
AP = mul!(copy(A), US, Vᴴ, -1, 1)
AP ./= S[end]
S⁻¹ = S[end] ./ S
minS = minimum(S)

@lkdvos lkdvos Aug 17, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Does last also work? Can we avoid blocking/syncing by using a view of the last element?

@kshyatt kshyatt Aug 18, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

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

I don't think so 😬 let me try and report back

@kshyatt kshyatt Aug 18, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

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

I think the problem is any of these require copying the last element back to the CPU, unless we write a custom kernel/map here which I'm kind of loathe to do...

AP ./= minS
S⁻¹ = minS ./ S
X₁ = rmul!(AP * Y₀ᴴ', Diagonal(S⁻¹))
X₁ .+= X₀
Y₁ᴴ = lmul!(Diagonal(S⁻¹), X₀' * AP)
Expand Down
Loading