Add diag argument to copytri!

[eschnett]

Julia 1.13 added a diag::Bool argument to copytri!. We need to mimic that so that we can match the signature when overloading for GPUArrays.

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic master) to apply these changes.

Click here to view the suggested changes.

diff --git a/src/host/linalg.jl b/src/host/linalg.jl
index ea4c89c..f5ebbc4 100644
--- a/src/host/linalg.jl
+++ b/src/host/linalg.jl
@@ -114,8 +114,8 @@ function LinearAlgebra.copytri!(A::AbstractGPUMatrix, uplo::AbstractChar, conjug
             I = @index(Global, Cartesian)
             i, j = Tuple(I)
             if j + diag > i
-                @inbounds _A[j,i] = conj(_A[i,j])
-          end
+                @inbounds _A[j, i] = conj(_A[i, j])
+            end
         end
         U_conj!(get_backend(A))(A; ndrange = size(A))
     elseif uplo == 'U' && !conjugate
@@ -123,8 +123,8 @@ function LinearAlgebra.copytri!(A::AbstractGPUMatrix, uplo::AbstractChar, conjug
             I = @index(Global, Cartesian)
             i, j = Tuple(I)
             if j + diag > i
-                @inbounds _A[j,i] = _A[i,j]
-          end
+                @inbounds _A[j, i] = _A[i, j]
+            end
         end
         U_noconj!(get_backend(A))(A; ndrange = size(A))
     elseif uplo == 'L' && conjugate
@@ -132,7 +132,7 @@ function LinearAlgebra.copytri!(A::AbstractGPUMatrix, uplo::AbstractChar, conjug
             I = @index(Global, Cartesian)
             i, j = Tuple(I)
             if j + diag > i
-                @inbounds _A[i,j] = conj(_A[j,i])
+                @inbounds _A[i, j] = conj(_A[j, i])
             end
         end
         L_conj!(get_backend(A))(A; ndrange = size(A))
@@ -141,14 +141,14 @@ function LinearAlgebra.copytri!(A::AbstractGPUMatrix, uplo::AbstractChar, conjug
             I = @index(Global, Cartesian)
             i, j = Tuple(I)
             if j + diag > i
-                @inbounds _A[i,j] = _A[j,i]
+                @inbounds _A[i, j] = _A[j, i]
             end
         end
         L_noconj!(get_backend(A))(A; ndrange = size(A))
     else
         throw(ArgumentError("uplo argument must be 'U' (upper) or 'L' (lower), got $uplo"))
     end
-    A
+    return A
 end
 
 ## copy a triangular part of a matrix to another matrix

[eschnett]

I think runic's indentation is wrong. Look at these lines:

        I = @index(Global, Cartesian)
        i, j = Tuple(I)
            if j + diag > i
          @inbounds _A[j,i] = conj(_A[i,j])
        end

The if statement is spuriously indented too far.

[maleadt]

Yeah feel free to ignore Runic...

[maleadt]

Can you add 1.13-nightly to the GH:A configuration too?

Also, this seems to break CUDA.jl etc tests on <1.13:

Testing finished in 40 minutes, 23 seconds, 298 milliseconds
2026-01-19 22:24:29 CEST
	Worker 3 failed running test gpuarrays/linalg/core:
2026-01-19 22:24:29 CEST
	Some tests did not pass: 420 passed, 0 failed, 8 errored, 0 broken.
2026-01-19 22:24:29 CEST
	gpuarrays/linalg/core: Error During Test at /var/lib/buildkite-agent/builds/gpuci-17/julialang/gpuarrays-dot-jl/test/testsuite/linalg.jl:86
2026-01-19 22:24:30 CEST
	  Test threw exception
2026-01-19 22:24:30 CEST
	  Expression: compare((x->begin
2026-01-19 22:24:30 CEST
	            #= /var/lib/buildkite-agent/builds/gpuci-17/julialang/gpuarrays-dot-jl/test/testsuite/linalg.jl:86 =#
2026-01-19 22:24:30 CEST
	            LinearAlgebra.copytri!(x, uplo, conjugate, diag)
2026-01-19 22:24:30 CEST
	        end), AT, a)
2026-01-19 22:24:30 CEST
	  KernelException: exception thrown during kernel execution on device Quadro RTX 5000

[eschnett]

We're going to need to ignore the Buildkite error at the moment. It fails because CUDA.jl isn't working with Julia 1.13. That's a known problem, and a prerequisite for making it work is making GPUArrays.jl work with Julia 1.13...

[eschnett]

This is running tests on the macOS-latest image. This image changed from x86_64 to aarch64 recently. Are you aware of this? Would this be using an aarch64 Julia, or is this running an x86_64 Julia on aarch64 via Rosetta? It's running very slowly, that's sure.

[eschnett]

... it's using julia/1.12.4/aarch64, all is fine.

[maleadt]

CUDA.jl on 1.12 failing here but not on master, https://buildkite.com/julialang/gpuarrays-dot-jl/builds/1565, is interesting. I thought I fixed JuliaGPU/CUDA.jl#2946 in JuliaGPU/CUDA.jl#3016, maybe we need another quirk?

[eschnett]

The Julia 1.12 error is

�_bk;t=1768928863646�Worker 4 failed running test gpuarrays/linalg/norm:
�_bk;t=1768928863720��[91mSome tests did not pass: 664 passed, 0 failed, 32 errored, 0 broken.�[39m
�_bk;t=1768928863754�gpuarrays/linalg/norm: �[91m�[1mError During Test�[22m�[39m at �[39m�[1m/var/lib/buildkite-agent/builds/gpuci-15/julialang/gpuarrays-dot-jl/test/testsuite/linalg.jl:525�[22m
�_bk;t=1768928864289�  Test threw exception
�_bk;t=1768928864294�  Expression: compare(norm, AT, arr, Ref(p))
�_bk;t=1768928864312�  Failed to compile PTX code (ptxas exited with code 255)
�_bk;t=1768928864312�  Invocation arguments: --generate-line-info --verbose --gpu-name sm_75 --output-file /tmp/jl_7apmpGDCDA.cubin /tmp/jl_nE3ul05szU.ptx
�_bk;t=1768928864312�  ptxas /tmp/jl_nE3ul05szU.ptx, line 671; error   : Modifier '.NaN' requires .target sm_80 or higher
�_bk;t=1768928864312�  ptxas /tmp/jl_nE3ul05szU.ptx, line 671; error   : Feature 'max.f16 or max.f16x2' requires .target sm_80 or higher

It seems that some part of the infrastructure generates code for a newer GPU than it should.

[eschnett]

I think I see what is happening. The CI machines have different GPUs:

• NVIDIA A100-PCIE-40GB MIG 1g.5gb (sm_80, 4.713 GiB / 4.750 GiB available)
• Quadro RTX 5000 (sm_75, 15.550 GiB / 16.000 GiB available)

The first supports sm_80 and the tests succeed, the second doesn't and the tests fail. This appears unrelated to my changes.

The problem might be in the file CUDA/src/device/intrinsics/math.jl in the function @device_override @inline function Base.max(x::Float32, y::Float32).

[eschnett]

Ping? I think this PR is ready. The Julia 1.12 problem seems unrelated.

[eschnett]

I want to make CUDA.jl work with Julia 1.13. One prerequisite is to make GPUArrays.jl work with Julia 1.13. Can you review this PR?

[maleadt]

I'll SSH into the GPU CI machine asap (tomorrow or Thursday) to diagnose the exact failure and get this merged.

[maleadt]

Re-started CI after merging JuliaGPU/CUDA.jl#3025.

[maleadt]

Thanks!