diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2ae90818b..2c8edfd7d 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -37,6 +37,7 @@ Entries link to the pull request that introduced them.
 
 **Fixed**
 - Fixed `Enzyme.gradient` failing when differentiating `GCNConv`, `SGConv` and `TAGConv` on `:dense`/`:sparse` adjacency graphs: their adjacency-matrix fallbacks now convert via the Enzyme-differentiable `GNNGraphs._to_coo_graph` instead of the keyword `GNNGraph` constructor. Requires GNNGraphs ≥ 1.5.2 and, for Enzyme, Enzyme ≥ 0.13.197 ([#703]).
+- `GATConv` and `GATv2Conv` skip attention dropout when `dropout == 0`, which also makes them differentiable with Mooncake on CUDA ([#XXX]).
 
 ## GraphNeuralNetworks.jl — Unreleased (towards 1.1.1)
 
@@ -238,4 +239,5 @@ Lux implementations of the graph convolutional, pooling, and temporal layers
 [#704]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/704
 [#707]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/707
 [#706]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/706
+[#709]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/709
 [FluxML/Zygote.jl#1662]: https://github.com/FluxML/Zygote.jl/issues/1662
diff --git a/GNNlib/src/layers/conv.jl b/GNNlib/src/layers/conv.jl
index ab40f13f2..fbdf6cc5d 100644
--- a/GNNlib/src/layers/conv.jl
+++ b/GNNlib/src/layers/conv.jl
@@ -136,7 +136,8 @@ function gat_conv(l, g::AbstractGNNGraph, x, e::Union{Nothing, AbstractMatrix} =
     message = Fix1(gat_message, l)
     m = apply_edges(message, g, Wxi, Wxj, e)
     α = softmax_edge_neighbors(g, m.logα)
-    α = dropout(α, l.dropout)
+    # Skip at p == 0: NNlib.dropout fetches the (CUDA) RNG before checking p, which Mooncake cannot trace.
+    iszero(l.dropout) || (α = dropout(α, l.dropout))
     β = α .* m.Wxj
     x = aggregate_neighbors(g, +, β)
 
@@ -188,7 +189,7 @@ function gatv2_conv(l, g::AbstractGNNGraph, x, e::Union{Nothing, AbstractMatrix}
     message = Fix1(gatv2_message, l)
     m = apply_edges(message, g, Wxi, Wxj, e)
     α = softmax_edge_neighbors(g, m.logα)
-    α = dropout(α, l.dropout)
+    iszero(l.dropout) || (α = dropout(α, l.dropout))
     β = α .* m.Wxj
     x = aggregate_neighbors(g, +, β)