From ddecf0a0a20165c1bc4543c7927af052ad7f84bf Mon Sep 17 00:00:00 2001
From: AntonOresten <antonoresten@proton.me>
Date: Wed, 12 Aug 2026 14:32:48 +0200
Subject: [PATCH 1/3] Add cuDNN extension

---
 Project.toml    |  2 ++
 ext/cuDNNExt.jl | 16 ++++++++++++++++
 2 files changed, 18 insertions(+)
 create mode 100644 ext/cuDNNExt.jl

diff --git a/Project.toml b/Project.toml
index 849b560..152820a 100644
--- a/Project.toml
+++ b/Project.toml
@@ -12,9 +12,11 @@ Republic = "27243419-9dde-4721-b67c-fd63626fea7f"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 
 [weakdeps]
+cuDNN = "02a925ec-e4fe-4b08-9a7e-0d78e3d38ccd"
 cuTile = "0dea8319-8c4a-4662-a73d-20234d115b9a"
 
 [extensions]
+cuDNNExt = "cuDNN"
 cuTileExt = "cuTile"
 
 [compat]
diff --git a/ext/cuDNNExt.jl b/ext/cuDNNExt.jl
new file mode 100644
index 0000000..a54272d
--- /dev/null
+++ b/ext/cuDNNExt.jl
@@ -0,0 +1,16 @@
+module cuDNNExt
+
+import BitPacking
+import cuDNN
+
+function cuDNN.checked_array_pointer(t::cuDNN.Tensor, a::BitPacking.NarrowArray)
+    parent(a) isa cuDNN.DenseCuArray || throw(ArgumentError(
+        "binding for $(t.name) needs GPU-backed storage, got $(typeof(parent(a)))"))
+    cuDNN.graph_dtype(eltype(a)) == t.dtype || throw(ArgumentError(
+        "binding for $(t.name) has eltype $(eltype(a)), expected $(t.dtype)"))
+    length(a) == prod(t.dims) || throw(DimensionMismatch(
+        "binding for $(t.name) has $(length(a)) elements, expected $(prod(t.dims))"))
+    return pointer(parent(a))
+end
+
+end

From 67a35892b9b53425f39956a7d955240a83c46b18 Mon Sep 17 00:00:00 2001
From: AntonOresten <antonoresten@proton.me>
Date: Thu, 13 Aug 2026 21:25:43 +0200
Subject: [PATCH 2/3] Update cuTile version to v1; add temporary cuDNN compat
 of v6.2.1

---
 Project.toml | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/Project.toml b/Project.toml
index 152820a..e5efc40 100644
--- a/Project.toml
+++ b/Project.toml
@@ -23,5 +23,6 @@ cuTileExt = "cuTile"
 Adapt = "4.6.1"
 Republic = "2.2.0"
 StaticArrays = "1.9.18"
-cuTile = "0.3.2"
+cuDNN = "6.2.1"
+cuTile = "1.0.0"
 julia = "1.10"

From e7cd78659893c33d8959b68e41bd61aa30268f36 Mon Sep 17 00:00:00 2001
From: AntonOresten <antonoresten@proton.me>
Date: Thu, 13 Aug 2026 21:26:24 +0200
Subject: [PATCH 3/3] Allow cuTile v0.3.2

---
 Project.toml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Project.toml b/Project.toml
index e5efc40..93ed50c 100644
--- a/Project.toml
+++ b/Project.toml
@@ -24,5 +24,5 @@ Adapt = "4.6.1"
 Republic = "2.2.0"
 StaticArrays = "1.9.18"
 cuDNN = "6.2.1"
-cuTile = "1.0.0"
+cuTile = "0.3.2, 1.0.0"
 julia = "1.10"