Skip to content

Bump minimum version of CUDA and cuTENSOR #404

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
kshyatt merged 10 commits into from
Apr 27, 2026
Merged

Bump minimum version of CUDA and cuTENSOR #404

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
e882b8e
Bump minimum version of CUDA and cuTENSOR
kshyatt Apr 17, 2026
20fbbdc
Some import fixes
kshyatt Apr 21, 2026
66c1bc6
Sources for Projects
kshyatt Apr 21, 2026
b7825aa
Fix rand/randn
kshyatt Apr 22, 2026
2a34113
Cleanup
kshyatt Apr 24, 2026
b514220
Update Project.toml
kshyatt Apr 24, 2026
df502ab
Merge branch 'main' into ksh/cuda_bump
kshyatt Apr 24, 2026
9af9c4c
Merge branch 'main' into ksh/cuda_bump
lkdvos Apr 27, 2026
62632fb
Fix import
Apr 25, 2026
566adbd
Another module path
Apr 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 0 additions & 2 deletions .buildkite/pipeline.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,8 +14,6 @@ steps:
agents:
queue: "juliagpu"
cuda: "*"
commands: |
unset LD_LIBRARY_PATH
if: build.message !~ /\[skip tests\]/
timeout_in_minutes: 90
matrix:
Expand Down
8 changes: 4 additions & 4 deletions Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,14 +41,14 @@ projects = ["test", "docs"]
[compat]
Adapt = "4"
AMDGPU = "2"
CUDA = "5.9"
CUDA = "6"
ChainRulesCore = "1"
Dictionaries = "0.4"
FiniteDifferences = "0.12"
LRUCache = "1.0.2"
LinearAlgebra = "1"
MatrixAlgebraKit = "0.6.5"
Mooncake = "0.5"
MatrixAlgebraKit = "0.6.6"
Mooncake = "0.5.27"
OhMyThreads = "0.8.0"
Printf = "1"
Random = "1"
Expand All @@ -58,5 +58,5 @@ TensorKitSectors = "0.3.7"
TensorOperations = "5.1"
TupleTools = "1.5"
VectorInterface = "0.4.8, 0.5"
cuTENSOR = "2"
cuTENSOR = "6"
julia = "1.10"
6 changes: 3 additions & 3 deletions ext/TensorKitCUDAExt/TensorKitCUDAExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,11 +1,11 @@
module TensorKitCUDAExt

using CUDA, CUDA.CUBLAS, CUDA.CUSOLVER, LinearAlgebra
using CUDA, CUDA.cuBLAS, CUDA.cuSOLVER, CUDA.cuRAND, LinearAlgebra
using CUDA: @allowscalar
using cuTENSOR: cuTENSOR
import CUDA.cuRAND: rand as curand, rand! as curand!, randn as curandn, randn! as curandn!
using Strided: StridedViews
import CUDA: rand as curand, rand! as curand!, randn as curandn, randn! as curandn!
using CUDA.KernelAbstractions: @kernel, @index, get_backend
using CUDA.CUDACore.KernelAbstractions: @kernel, @index, get_backend

using TensorKit
using TensorKit.Factorizations
Expand Down
2 changes: 1 addition & 1 deletion ext/TensorKitCUDAExt/cutensormap.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -170,5 +170,5 @@ for f in (:sqrt, :log, :asin, :acos, :acosh, :atanh, :acoth)
end

function TensorKit._add_transform_multi!(tdst::CuTensorMap, tsrc, p, (U, structs_dst, structs_src)::Tuple{<:Array, TD, TS}, buffers, alpha, beta, backend...) where {TD, TS}
return TensorKit._add_transform_multi!(tdst, tsrc, p, (CUDA.Adapt.adapt(CuArray, U), structs_dst, structs_src), buffers, alpha, beta, backend...)
return TensorKit._add_transform_multi!(tdst, tsrc, p, (CUDA.CUDACore.Adapt.adapt(CuArray, U), structs_dst, structs_src), buffers, alpha, beta, backend...)
end
102 changes: 51 additions & 51 deletions test/cuda/factorizations.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
using Adapt, CUDA, cuTENSOR
using Adapt, CUDA, CUDA.cuRAND, cuTENSOR
using Test, TestExtras
using TensorKit
using LinearAlgebra: LinearAlgebra
Expand All @@ -25,10 +25,10 @@ for V in spacelist
@testset "QR decomposition" begin
for T in eltypes,
t in (
CUDA.rand(T, W, W), CUDA.rand(T, W, W)',
CUDA.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'), CUDA.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)')',
CUDA.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5)), CUDA.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
DiagonalTensorMap(CUDA.rand(T, reduceddim(V1)), V1),
cuRAND.rand(T, W, W), cuRAND.rand(T, W, W)',
cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'), cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)')',
cuRAND.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5)), cuRAND.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
DiagonalTensorMap(cuRAND.rand(T, reduceddim(V1)), V1),
)

Q, R = @constinferred qr_full(t)
Expand All @@ -54,7 +54,7 @@ for V in spacelist

# empty tensor
for T in eltypes
t = CUDA.rand(T, V1 ⊗ V2, zerospace(V1))
t = cuRAND.rand(T, V1 ⊗ V2, zerospace(V1))

Q, R = @constinferred qr_full(t)
@test Q * R ≈ t
Expand All @@ -80,10 +80,10 @@ for V in spacelist
@testset "LQ decomposition" begin
for T in eltypes,
t in (
CUDA.rand(T, W, W), CUDA.rand(T, W, W)',
CUDA.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'), CUDA.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)')',
CUDA.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5)), CUDA.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(CUDA.rand(T, reduceddim(V1)), V1),
cuRAND.rand(T, W, W), cuRAND.rand(T, W, W)',
cuRAND.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'), cuRAND.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)')',
cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5)), cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(cuRAND.rand(T, reduceddim(V1)), V1),
)

L, Q = @constinferred lq_full(t)
Expand All @@ -105,7 +105,7 @@ for V in spacelist

for T in eltypes
# empty tensor
t = CUDA.rand(T, zerospace(V1), V1 ⊗ V2)
t = cuRAND.rand(T, zerospace(V1), V1 ⊗ V2)

L, Q = @constinferred lq_full(t)
@test L * Q ≈ t
Expand All @@ -131,10 +131,10 @@ for V in spacelist
@testset "Polar decomposition" begin
@testset for T in eltypes,
t in (
CUDA.rand(T, W, W),
CUDA.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'),
CUDA.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
DiagonalTensorMap(CUDA.rand(T, reduceddim(V1)), V1),
cuRAND.rand(T, W, W),
cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'),
cuRAND.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
DiagonalTensorMap(cuRAND.rand(T, reduceddim(V1)), V1),
)

@assert domain(t) ≾ codomain(t)
Expand All @@ -150,10 +150,10 @@ for V in spacelist

@testset for T in eltypes,
t in (
CUDA.rand(T, W, W),
CUDA.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'),
CUDA.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(CUDA.rand(T, reduceddim(V1)), V1),
cuRAND.rand(T, W, W),
cuRAND.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'),
cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(cuRAND.rand(T, reduceddim(V1)), V1),
)

@assert codomain(t) ≾ domain(t)
Expand All @@ -171,10 +171,10 @@ for V in spacelist
@testset "SVD" begin
for T in eltypes,
t in (
CUDA.rand(T, W, W), CUDA.rand(T, W, W)',
CUDA.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'), CUDA.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
CUDA.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'), CUDA.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(CUDA.rand(T, reduceddim(V1)), V1),
cuRAND.rand(T, W, W), cuRAND.rand(T, W, W)',
cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'), cuRAND.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
cuRAND.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'), cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(cuRAND.rand(T, reduceddim(V1)), V1),
)

u, s, vᴴ = @constinferred svd_full(t)
Expand Down Expand Up @@ -222,8 +222,8 @@ for V in spacelist

# empty tensor
for T in eltypes, t in (
CUDA.rand(T, W, zerospace(V1)),
CUDA.rand(T, zerospace(V1), W),
cuRAND.rand(T, W, zerospace(V1)),
cuRAND.rand(T, zerospace(V1), W),
)
U, S, Vᴴ = @constinferred svd_full(t)
@test U * S * Vᴴ ≈ t
Expand All @@ -239,10 +239,10 @@ for V in spacelist
@testset "truncated SVD" begin
for T in eltypes,
t in (
CUDA.randn(T, W, W), CUDA.randn(T, W, W)',
CUDA.randn(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'), CUDA.randn(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
CUDA.randn(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'), CUDA.randn(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(CUDA.randn(T, reduceddim(V1)), V1),
cuRAND.randn(T, W, W), cuRAND.randn(T, W, W)',
cuRAND.randn(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'), cuRAND.randn(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
cuRAND.randn(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'), cuRAND.randn(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(cuRAND.randn(T, reduceddim(V1)), V1),
)

@constinferred normalize!(t)
Expand Down Expand Up @@ -307,10 +307,10 @@ for V in spacelist
@testset "Eigenvalue decomposition" begin
for T in eltypes,
t in (
CUDA.rand(T, V1, V1),
CUDA.rand(T, W, W),
CUDA.rand(T, W, W)',
# DiagonalTensorMap(CUDA.rand(T, reduceddim(V1)), V1),
cuRAND.rand(T, V1, V1),
cuRAND.rand(T, W, W),
cuRAND.rand(T, W, W)',
# DiagonalTensorMap(cuRAND.rand(T, reduceddim(V1)), V1),
)

d, v = @constinferred eig_full(t)
Expand Down Expand Up @@ -367,15 +367,15 @@ for V in spacelist
@testset "Condition number and rank" begin
for T in eltypes,
t in (
CUDA.rand(T, W, W), CUDA.rand(T, W, W)',
CUDA.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'), CUDA.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
CUDA.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'), CUDA.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(CUDA.rand(T, reduceddim(V1)), V1),
cuRAND.rand(T, W, W), cuRAND.rand(T, W, W)',
cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'), cuRAND.rand(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
cuRAND.rand(T, (V1 ⊗ V2), (V3 ⊗ V4 ⊗ V5)'), cuRAND.rand(T, (V1 ⊗ V2 ⊗ V3)', (V4 ⊗ V5))',
DiagonalTensorMap(cuRAND.rand(T, reduceddim(V1)), V1),
)

d1, d2 = dim(codomain(t)), dim(domain(t))
r = rank(t)
@test r == min(d1, d2)
@test r ≈ min(d1, d2)
@test typeof(r) == typeof(d1)
M = left_null(t)
@test @constinferred(rank(M)) + r ≈ d1
Expand All @@ -387,15 +387,15 @@ for V in spacelist
@test @constinferred(cond(u)) ≈ one(real(T))
@test @constinferred(rank(u)) == dim(V1 ⊗ V2)

t = CUDA.rand(T, zerospace(V1), W)
t = cuRAND.rand(T, zerospace(V1), W)
@test rank(t) == 0
t2 = CUDA.rand(T, zerospace(V1) * zerospace(V2), zerospace(V1) * zerospace(V2))
t2 = cuRAND.rand(T, zerospace(V1) * zerospace(V2), zerospace(V1) * zerospace(V2))
@test rank(t2) == 0
@test cond(t2) == 0.0
end
for T in eltypes, t in (
CUDA.rand(T, W, W),
CUDA.rand(T, W, W)',
cuRAND.rand(T, W, W),
cuRAND.rand(T, W, W)',
)
project_hermitian!(t)
vals = @constinferred LinearAlgebra.eigvals(t)
Expand All @@ -408,10 +408,10 @@ for V in spacelist
@testset "Hermitian projections" begin
for T in eltypes,
t in (
CUDA.rand(T, V1, V1),
CUDA.rand(T, W, W),
CUDA.rand(T, W, W)',
DiagonalTensorMap(CUDA.rand(T, reduceddim(V1)), V1),
cuRAND.rand(T, V1, V1),
cuRAND.rand(T, W, W),
cuRAND.rand(T, W, W)',
DiagonalTensorMap(cuRAND.rand(T, reduceddim(V1)), V1),
)
normalize!(t)
noisefactor = eps(real(T))^(3 / 4)
Expand Down Expand Up @@ -441,10 +441,10 @@ for V in spacelist
@testset "Isometric projections" begin
for T in eltypes,
t in (
CUDA.randn(T, W, W),
CUDA.randn(T, W, W)',
CUDA.randn(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'),
CUDA.randn(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
cuRAND.randn(T, W, W),
cuRAND.randn(T, W, W)',
cuRAND.randn(T, (V1 ⊗ V2 ⊗ V3), (V4 ⊗ V5)'),
cuRAND.randn(T, (V1 ⊗ V2)', (V3 ⊗ V4 ⊗ V5))',
)
t2 = project_isometric(t)
@test isisometric(t2)
Expand All @@ -459,7 +459,7 @@ for V in spacelist

# test that t2 is closer to A then any other isometry
for k in 1:10
δt = CUDA.randn!(similar(t))
δt = cuRAND.randn!(similar(t))
t3 = project_isometric(t + δt / 100)
@test norm(t - t3) > norm(t - t2)
end
Expand Down
Loading
Loading