Skip to content

Fix BatchNorm crash under Mooncake in eval mode #740

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
AstitvaAggarwal wants to merge 2 commits into
base: master
Choose a base branch
from
+11 −1
Open

Fix BatchNorm crash under Mooncake in eval mode #740

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bf330b9
Only construct a live BatchNorm cache when training=true
AstitvaAggarwal Jul 22, 2026
033864c
Respect a caller-supplied BatchNorm cache instead of overriding it
AstitvaAggarwal Jul 22, 2026
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
12 changes: 11 additions & 1 deletion ext/NNlibMooncakeCUDAExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,7 +65,17 @@ function rrule!!(
prv = primal(running_var)
pm = primal(momentum)

fwd_cache = _BNFwdCache()
# cudnnBNForward! only populates a passed-in cache when training=true; in eval mode
# it's left unset, and cudnnBNBackward! reads it anyway, crashing on the stale
# `nothing`. Only build a live cache when training is actually on, and respect a
# caller-supplied cache (e.g. via Flux.BatchNorm's own `cache` argument) instead of
# silently replacing it with our own.
fwd_cache = if get(pkw, :training, true)
caller_cache = get(pkw, :cache, nothing)
caller_cache === nothing ? _BNFwdCache() : caller_cache
else
nothing
end
y = batchnorm(pg, pb, px, prm, prv, pm; pkw..., cache=fwd_cache)
dy_out = zero(y)
zero_kw = zero_rdata(pkw)
Expand Down
Loading