From 156a51a8f33f8e028009ac47c4887b077962a993 Mon Sep 17 00:00:00 2001
From: Tim Besard <tim.besard@gmail.com>
Date: Fri, 15 May 2026 10:57:14 +0200
Subject: [PATCH] cuBLAS: extend norm/norm2 to strided vector views.

Generalizes the BLAS-optimized `norm`/`norm2` methods from `DenseCuArray`
to `StridedCuVecOrDenseMat`, so 1D strided subarray views also dispatch
to `nrm2`. Multi-dim non-contiguous views go through the sum-based
fallback in GPUArrays (which now dispatches on `AnyGPUArray`).

Resolves JuliaGPU/CUDA.jl#2280, replaces #2302.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 CUDACore/Project.toml          |  2 +-
 lib/cublas/src/linalg.jl       |  6 +++---
 lib/cublas/test/level1/core.jl | 13 +++++++++++++
 3 files changed, 17 insertions(+), 4 deletions(-)

diff --git a/CUDACore/Project.toml b/CUDACore/Project.toml
index ad44b0203f..bab293aba3 100644
--- a/CUDACore/Project.toml
+++ b/CUDACore/Project.toml
@@ -52,7 +52,7 @@ CUDA_Runtime_jll = "0.22"
 ChainRulesCore = "1"
 EnzymeCore = "0.8.2"
 ExprTools = "0.1"
-GPUArrays = "11.5"
+GPUArrays = "11.5.4"
 GPUCompiler = "1.10"
 GPUToolbox = "1.1"
 KernelAbstractions = "0.9.38"
diff --git a/lib/cublas/src/linalg.jl b/lib/cublas/src/linalg.jl
index 60cdc49a2c..c138c2ce23 100644
--- a/lib/cublas/src/linalg.jl
+++ b/lib/cublas/src/linalg.jl
@@ -209,16 +209,16 @@ function LinearAlgebra.:(*)(transx::Transpose{<:Any,<:StridedCuVector{T}},
     return dotu(n, x, y)
 end
 
-function LinearAlgebra.norm(x::DenseCuArray{<:Union{Float16, ComplexF16, CublasFloat}},
+function LinearAlgebra.norm(x::StridedCuVecOrDenseMat{<:Union{Float16, ComplexF16, CublasFloat}},
                             p::Real=2)
     if p == 2
         return nrm2(x)
     else
-        return invoke(norm, Tuple{AbstractGPUArray, Real}, x, p)
+        return invoke(norm, Tuple{AnyGPUArray, Real}, x, p)
     end
 end
 LinearAlgebra.norm(x::Diagonal{T, <:StridedCuVector{T}}, p::Real=2) where {T<:Union{Float16, ComplexF16, CublasFloat}} = norm(x.diag, p)
-LinearAlgebra.norm2(x::DenseCuArray{<:Union{Float16, ComplexF16, CublasFloat}}) = nrm2(x)
+LinearAlgebra.norm2(x::StridedCuVecOrDenseMat{<:Union{Float16, ComplexF16, CublasFloat}}) = nrm2(x)
 
 LinearAlgebra.BLAS.asum(x::StridedCuArray{<:CublasFloat}) = asum(length(x), x)
 
diff --git a/lib/cublas/test/level1/core.jl b/lib/cublas/test/level1/core.jl
index 22e6daec51..0fafe5de92 100644
--- a/lib/cublas/test/level1/core.jl
+++ b/lib/cublas/test/level1/core.jl
@@ -76,6 +76,19 @@ using LinearAlgebra
         @test norm(dDx, 2) ≈ norm(Dx, 2)
         @test norm(dDx, Inf) ≈ norm(Dx, Inf)
     end
+
+    @testset "norm of strided views" begin # JuliaGPU/CUDA.jl#2280
+        # 1D contiguous view: should hit the cuBLAS nrm2 fast path.
+        x = rand(T, m)
+        dx = CuArray(x)
+        @test norm(@view(dx[2:end-1]), 2) ≈ norm(@view(x[2:end-1]), 2)
+        # Multi-dim non-contiguous view: must avoid scalar iteration.
+        y = rand(T, 10, 10)
+        dy = CuArray(y)
+        @test norm(@view(dy[2:end-1, 2:end-1]), 1) ≈ norm(@view(y[2:end-1, 2:end-1]), 1)
+        @test norm(@view(dy[2:end-1, 2:end-1]), 2) ≈ norm(@view(y[2:end-1, 2:end-1]), 2)
+        @test norm(@view(dy[2:end-1, 2:end-1]), Inf) ≈ norm(@view(y[2:end-1, 2:end-1]), Inf)
+    end
 end
 
 @testset for T in [Float16, ComplexF16]