Stream: helpdesk (published)

Topic: New resolver errors


view this post on Zulip James Wrigley (Nov 16 2026 at 11:17):

I'm a bit confused about this error I'm getting with the new resolver in 1.14 in a project with both CUDA.jl and DifferentialEquations.jl:

pkg> add CUDA@6.5.1
   Resolving package versions...
ERROR: Unsatisfiable requirements detected — 1 conflict:

Conflict 1: CUDA and DifferentialEquations
  • DifferentialEquations constrains CUDA to ≤6.5.0 (through CUDACore,
    OrdinaryDiffEq, Reexport, KernelAbstractions, LinearSolve, NonlinearSolve
    and OrdinaryDiffEqRosenbrock)
  • your compat restricts CUDA to 6.5.1
  Fix it by any one of:
    1. relax your compat on CUDA
       → allows: CUDA 6.4.2, DifferentialEquations 8.1.1
    2. drop dependency CUDA
       → allows: DifferentialEquations 8.1.1
  Note: dropping dependency DifferentialEquations would not help unless you
    also dropped dependency OrdinaryDiffEq.

In this project the only compat lines I have are:

[compat]
CUDA = "6"
Revise = "3"

And I went through the packages listed in the conflict error message, but unless I'm missing something none of them declare any compat with CUDA at all :thinking: Sooo I'm a bit confused. Does anyone know how to find the offending package?

This is the output of st --outdated:

pkg> st --outdated
Status `/path/to/Project.toml`
⌃ [052768ef] CUDA v6.4.2 (<v6.5.1)
⌅ [bd0ed864] CUDACore v6.4.2 (<v6.5.1): CUDA, CUDATools, CUPTI, NVML, cuBLAS, cuFFT, cuRAND, cuSOLVER, cuSPARSE
⌅ [9ec180c6] CUDATools v6.4.2 (<v6.5.1): CUDA
⌅ [63c18a36] KernelAbstractions v0.9.47 (<v0.10.7): CUDACore, cuSPARSE
⌃ [b20c7882] LinearInterpolations v0.1.4 (<v0.1.6)
⌅ [bac558e1] OrderedCollections v1.8.2 (<v2.0.3): DataFramesMeta, PProf, Parameters
⌅ [182d3088] cuBLAS v6.4.2 (<v6.5.1): CUDA, cuSOLVER
⌅ [533571aa] cuFFT v6.4.2 (<v6.5.1): CUDA
⌅ [20fd9a0b] cuRAND v6.4.2 (<v6.5.1): CUDA
⌅ [887afef0] cuSOLVER v6.4.2 (<v6.5.1): CUDA
⌅ [b26da814] cuSPARSE v6.4.2 (<v6.5.1): CUDA, cuSOLVER

view this post on Zulip Gunnar Farnebäck (Nov 16 2026 at 13:05):

For example LinearSolve has a weak dependency on CUDA, although not restricted in the way indicated here.

I think the real problem is that CUDACore 6.5.1 has KernelAbstractions compat of 0.10 but some other of your indirect dependencies holds it back to 0.9.

view this post on Zulip Gunnar Farnebäck (Nov 16 2026 at 13:08):

Then there's also something that's not completely right with the error messages:

(jl_Adq52L) pkg> add OrdinaryDiffEqRosenbrock KernelAbstractions@0.10.7
   Resolving package versions...
ERROR: Unsatisfiable requirements detected — 1 conflict:

Conflict 1: KernelAbstractions and OrdinaryDiffEqRosenbrock
  • OrdinaryDiffEqRosenbrock requires LinearSolve ≥2.30.2
  • LinearSolve ≥2.30.2 constrains KernelAbstractions to ≤0.10.6
  • your compat restricts KernelAbstractions to 0.10.7

(jl_Adq52L) pkg> add OrdinaryDiffEqRosenbrock KernelAbstractions@0.10.6
   Resolving package versions...
ERROR: Unsatisfiable requirements detected — 1 conflict:

Conflict 1: KernelAbstractions and OrdinaryDiffEqRosenbrock
  • OrdinaryDiffEqRosenbrock requires LinearSolve ≥2.30.2
  • LinearSolve ≥2.30.2 constrains KernelAbstractions to ≤0.10.5, 0.10.7
  • your compat restricts KernelAbstractions to 0.10.6

(jl_Adq52L) pkg> add OrdinaryDiffEqRosenbrock KernelAbstractions@0.10.5
   Resolving package versions...
ERROR: Unsatisfiable requirements detected — 1 conflict:

Conflict 1: KernelAbstractions and OrdinaryDiffEqRosenbrock
  • OrdinaryDiffEqRosenbrock requires LinearSolve ≥2.30.2
  • LinearSolve ≥2.30.2 constrains KernelAbstractions to ≤0.10.4, ≥0.10.6
  • your compat restricts KernelAbstractions to 0.10.5

view this post on Zulip James Wrigley (Nov 16 2026 at 13:27):

I'm having a look through the downstream packages of KA, but nothing's jumping out at me so far:

pkg> why KernelAbstractions
  KernelAbstractions
  ├─← CUDACore
  │   ├─← CUDA
  │   ├─← CUDATools
  │   │   └─← CUDA
  │   ├─← CUPTI
  │   │   └─← CUDATools (*)
  │   ├─← NVML
  │   │   └─← CUDATools (*)
  │   ├─← cuBLAS
  │   │   ├─← CUDA
  │   │   └─← cuSOLVER
  │   │       └─← CUDA
  │   ├─← cuFFT
  │   │   └─← CUDA
  │   ├─← cuRAND
  │   │   └─← CUDA
  │   ├─← cuSOLVER (*)
  │   └─← cuSPARSE
  │       ├─← CUDA
  │       └─← cuSOLVER (*)
  ├─← GPUArrays
  │   ├─← CUDACore (*)
  │   ├─← cuBLAS (*)
  │   └─← cuSPARSE (*)
  └─← cuSPARSE (*)

view this post on Zulip James Wrigley (Nov 16 2026 at 13:29):

But I agree that KA is suspect since CUDA.jl 6.5 was the first version that required v0.10.

view this post on Zulip James Wrigley (Nov 16 2026 at 13:35):

Ah it probably is LinearSolve, this PR didn't make it into a release yet: https://github.com/SciML/LinearSolve.jl/pull/1404

view this post on Zulip James Wrigley (Nov 16 2026 at 13:37):

Maybe there should be a why --include-indirect option.

view this post on Zulip Gunnar Farnebäck (Nov 16 2026 at 13:51):

Then there's also something that's not completely right with the error messages:

https://github.com/JuliaLang/Resolver.jl/issues/136


Last updated: Nov 21 2026 at 12:52 UTC