From f692e007bf168f4b89a9ddd6c51b7600b61a4f61 Mon Sep 17 00:00:00 2001
From: "copilot-swe-agent[bot]" <198982749+Copilot@users.noreply.github.com>
Date: Thu, 20 Aug 2026 17:47:52 +0000
Subject: [PATCH 1/4] Initial plan


From 6af44cfeeb53cbdfef7f40b2191273237c78c465 Mon Sep 17 00:00:00 2001
From: "copilot-swe-agent[bot]" <198982749+Copilot@users.noreply.github.com>
Date: Thu, 20 Aug 2026 17:54:26 +0000
Subject: [PATCH 2/4] Fix derived CuTracedArray construction

Co-authored-by: ptiede <5565461+ptiede@users.noreply.github.com>
---
 ext/ReactantCUDAExt.jl   | 15 +++++++++++----
 test/integration/cuda.jl | 19 +++++++++++++++++++
 2 files changed, 30 insertions(+), 4 deletions(-)

diff --git a/ext/ReactantCUDAExt.jl b/ext/ReactantCUDAExt.jl
index 0b2cd80243..7f868bbcec 100644
--- a/ext/ReactantCUDAExt.jl
+++ b/ext/ReactantCUDAExt.jl
@@ -59,6 +59,9 @@ struct CuTracedArray{T,N,A,Size} <: DenseArray{T,N}
         ptr = Base.reinterpret(Core.LLVMPtr{T,CUDA.AS.Global}, Base.pointer_from_objref(xs))
         return new(ptr)
     end
+    function CuTracedArray{T,N,A,Size}(ptr::Core.LLVMPtr{T,A}) where {T,N,A,Size}
+        return new(ptr)
+    end
 end
 
 Reactant.use_overlayed_version(::CuTracedArray) = true
@@ -562,15 +565,13 @@ function Base.reinterpret(::Type{T}, a::CuTracedArray{S,N,A}) where {T,S,N,A}
     err === nothing || throw(err)
 
     if sizeof(T) == sizeof(S) # fast case
-        return CuTracedArray{T,N,A}(
-            reinterpret(Core.LLVMPtr{T,A}, a.ptr), size(a), a.maxsize
-        )
+        return _derived_array(a, T, size(a))
     end
 
     isize = size(a)
     size1 = div(isize[1] * sizeof(S), sizeof(T))
     osize = tuple(size1, Base.tail(isize)...)
-    return CuTracedArray{T,N,A}(reinterpret(Core.LLVMPtr{T,A}, a.ptr), osize, a.maxsize)
+    return _derived_array(a, T, osize)
 end
 
 ## reshape
@@ -589,6 +590,12 @@ function Base.reshape(a::CuTracedArray{T,M,A}, dims::NTuple{N,Int}) where {T,N,M
     return _derived_array(a, T, dims)
 end
 
+@inline function _derived_array(
+    a::CuTracedArray{<:Any,<:Any,A}, ::Type{T}, osize::Dims{N}
+) where {T,N,A}
+    return CuTracedArray{T,N,A,osize}(reinterpret(Core.LLVMPtr{T,A}, a.ptr))
+end
+
 struct ReactantKernelAdaptor end
 
 function Adapt.adapt_storage(to::ReactantKernelAdaptor, p::CUDA.CuPtr)
diff --git a/test/integration/cuda.jl b/test/integration/cuda.jl
index 9ab811a74f..09dad6de6e 100644
--- a/test/integration/cuda.jl
+++ b/test/integration/cuda.jl
@@ -114,6 +114,25 @@ function sin!(x, y)
     return nothing
 end
 
+function reshape_kernel!(out, x)
+    xr = reshape(x, 2, 2)
+    @inbounds out[1] = xr[1, 2] + xr[2, 1]
+    return nothing
+end
+
+function reshape!(out, x)
+    @cuda blocks = 1 threads = 1 reshape_kernel!(out, x)
+    return nothing
+end
+
+@testset "Reshape Kernel" begin
+    x = Reactant.to_rarray([1.0, 2.0, 3.0, 4.0])
+    out = Reactant.to_rarray(zeros(Float64, 1))
+
+    @jit reshape!(out, x)
+    @test Array(out) == [5.0]
+end
+
 @testset "Sin Kernel" begin
     oA = collect(Float64, 1:1:64)
     A = Reactant.to_rarray(oA)

From 090bc74025243f2225e787ff4bdfaea5a8639c1b Mon Sep 17 00:00:00 2001
From: "copilot-swe-agent[bot]" <198982749+Copilot@users.noreply.github.com>
Date: Thu, 20 Aug 2026 17:55:00 +0000
Subject: [PATCH 3/4] Tighten CUDA reshape regression test

Co-authored-by: ptiede <5565461+ptiede@users.noreply.github.com>
---
 test/integration/cuda.jl | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/test/integration/cuda.jl b/test/integration/cuda.jl
index 09dad6de6e..8336bc13f1 100644
--- a/test/integration/cuda.jl
+++ b/test/integration/cuda.jl
@@ -116,7 +116,7 @@ end
 
 function reshape_kernel!(out, x)
     xr = reshape(x, 2, 2)
-    @inbounds out[1] = xr[1, 2] + xr[2, 1]
+    @inbounds out[1] = xr[1, 2]
     return nothing
 end
 
@@ -130,7 +130,7 @@ end
     out = Reactant.to_rarray(zeros(Float64, 1))
 
     @jit reshape!(out, x)
-    @test Array(out) == [5.0]
+    @test Array(out) == [3.0]
 end
 
 @testset "Sin Kernel" begin

From adc7e6b6cfb7398be071a43ce3275673cec1c2aa Mon Sep 17 00:00:00 2001
From: "copilot-swe-agent[bot]" <198982749+Copilot@users.noreply.github.com>
Date: Thu, 20 Aug 2026 18:07:43 +0000
Subject: [PATCH 4/4] Use CUDA _derived_array when available

Co-authored-by: ptiede <5565461+ptiede@users.noreply.github.com>
---
 ext/ReactantCUDAExt.jl | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/ext/ReactantCUDAExt.jl b/ext/ReactantCUDAExt.jl
index 7f868bbcec..49162c60e4 100644
--- a/ext/ReactantCUDAExt.jl
+++ b/ext/ReactantCUDAExt.jl
@@ -15,7 +15,12 @@ using Reactant.Ops: @opcall
 using Enzyme
 using Adapt: Adapt, adapt
 using CUDA: CUDA, CuDim, DenseCuArray, unsafe_cached_load
-# Compatibility for CUDA v5 and v6
+@static if isdefined(CUDA, :_derived_array)
+    # Extend CUDA's internal derived-array helper so traced arrays follow the
+    # same reshape/reinterpret reconstruction path when it is available; the
+    # CuTracedArray-specific method is defined alongside reshape below.
+    import CUDA: _derived_array
+end
 
 const CUVERSION = isdefined(CUDA, :CUDACore) ? 6 : 5
 if CUVERSION == 6