Add Downgrade.yml and Fix Minimum Versions

[ChrisRackauckas]

FluxML/Zygote.jl#1583 identified that some of the CUDA.jl minimum versions are incorrect. In particular, it's possible to get versions of GPUCompiler that don't match the requirements of the package.

[ChrisRackauckas]

The Downgrade.yml doesn't work quite right because, well, of course this requires CUDA 😅 But this seems to be the right set of package versions. I can at least isolate this to be an improvement and downgrade CI could come later?

[github-actions]

CUDA.jl Benchmarks

Benchmark suite	Current: 2f847f3	Previous: c05359d	Ratio
latency/precompile	43009370470.5 ns	42922336650.5 ns	1.00
latency/ttfp	7011870092 ns	7015168424 ns	1.00
latency/import	3572091762 ns	3571269514 ns	1.00
integration/volumerhs	9627594 ns	9608723 ns	1.00
integration/byval/slices=1	147087 ns	146920.5 ns	1.00
integration/byval/slices=3	426017 ns	425845 ns	1.00
integration/byval/reference	145114 ns	145020 ns	1.00
integration/byval/slices=2	286435 ns	286380 ns	1.00
integration/cudadevrt	103547 ns	103554 ns	1.00
kernel/indexing	14145 ns	14235 ns	0.99
kernel/indexing_checked	14764 ns	14711 ns	1.00
kernel/occupancy	671.3630573248407 ns	672.5506329113924 ns	1.00
kernel/launch	2125.2 ns	2270.3333333333335 ns	0.94
kernel/rand	17437 ns	14669 ns	1.19
array/reverse/1d	19771 ns	19682 ns	1.00
array/reverse/2d	24668 ns	23613.5 ns	1.04
array/reverse/1d_inplace	10153.333333333334 ns	10461 ns	0.97
array/reverse/2d_inplace	12029 ns	13212 ns	0.91
array/copy	20668 ns	20972 ns	0.99
array/iteration/findall/int	157264 ns	157808 ns	1.00
array/iteration/findall/bool	139314.5 ns	139837 ns	1.00
array/iteration/findfirst/int	161899.5 ns	164937 ns	0.98
array/iteration/findfirst/bool	163430.5 ns	165868 ns	0.99
array/iteration/scalar	70556 ns	73041 ns	0.97
array/iteration/logical	212221.5 ns	214850 ns	0.99
array/iteration/findmin/1d	46278 ns	46704 ns	0.99
array/iteration/findmin/2d	96586 ns	96962.5 ns	1.00
array/reductions/reduce/Int64/1d	44092.5 ns	46033 ns	0.96
array/reductions/reduce/Int64/dims=1	55520.5 ns	55193 ns	1.01
array/reductions/reduce/Int64/dims=2	62743.5 ns	62917 ns	1.00
array/reductions/reduce/Int64/dims=1L	89021 ns	88869 ns	1.00
array/reductions/reduce/Int64/dims=2L	88699 ns	87079 ns	1.02
array/reductions/reduce/Float32/1d	35221 ns	34606 ns	1.02
array/reductions/reduce/Float32/dims=1	51495 ns	43875 ns	1.17
array/reductions/reduce/Float32/dims=2	59721.5 ns	59705 ns	1.00
array/reductions/reduce/Float32/dims=1L	52335 ns	52260 ns	1.00
array/reductions/reduce/Float32/dims=2L	70346 ns	70051.5 ns	1.00
array/reductions/mapreduce/Int64/1d	44753 ns	42671.5 ns	1.05
array/reductions/mapreduce/Int64/dims=1	48290 ns	45980 ns	1.05
array/reductions/mapreduce/Int64/dims=2	62460.5 ns	62143.5 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	89024 ns	88812 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	87076.5 ns	86818 ns	1.00
array/reductions/mapreduce/Float32/1d	35034 ns	34742 ns	1.01
array/reductions/mapreduce/Float32/dims=1	41996 ns	43090.5 ns	0.97
array/reductions/mapreduce/Float32/dims=2	59859 ns	60061 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	52524 ns	52528 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	70104 ns	70191 ns	1.00
array/broadcast	20048 ns	20155 ns	0.99
array/copyto!/gpu_to_gpu	12892 ns	11294 ns	1.14
array/copyto!/cpu_to_gpu	217203 ns	216503 ns	1.00
array/copyto!/gpu_to_cpu	285399 ns	284237 ns	1.00
array/accumulate/Int64/1d	125289.5 ns	125529 ns	1.00
array/accumulate/Int64/dims=1	83781 ns	84037 ns	1.00
array/accumulate/Int64/dims=2	158173.5 ns	159166 ns	0.99
array/accumulate/Int64/dims=1L	1711336.5 ns	1720376 ns	0.99
array/accumulate/Int64/dims=2L	966771.5 ns	968348 ns	1.00
array/accumulate/Float32/1d	109881 ns	109984 ns	1.00
array/accumulate/Float32/dims=1	80990.5 ns	81082 ns	1.00
array/accumulate/Float32/dims=2	147723.5 ns	148760 ns	0.99
array/accumulate/Float32/dims=1L	1619738 ns	1629307.5 ns	0.99
array/accumulate/Float32/dims=2L	698792 ns	701479 ns	1.00
array/construct	1277.6 ns	1287.2 ns	0.99
array/random/randn/Float32	44953.5 ns	44176 ns	1.02
array/random/randn!/Float32	25288.5 ns	24930 ns	1.01
array/random/rand!/Int64	27570 ns	27547 ns	1.00
array/random/rand!/Float32	8640.333333333334 ns	8724.666666666666 ns	0.99
array/random/rand/Int64	30128 ns	30114 ns	1.00
array/random/rand/Float32	13079 ns	13059 ns	1.00
array/permutedims/4d	60556 ns	60761 ns	1.00
array/permutedims/2d	54347 ns	54037 ns	1.01
array/permutedims/3d	55391 ns	54954 ns	1.01
array/sorting/1d	2759605 ns	2756544 ns	1.00
array/sorting/by	3347300.5 ns	3343249 ns	1.00
array/sorting/2d	1082195.5 ns	1080799 ns	1.00
cuda/synchronization/stream/auto	1017 ns	1040.3 ns	0.98
cuda/synchronization/stream/nonblocking	7719 ns	7220 ns	1.07
cuda/synchronization/stream/blocking	829.4268292682926 ns	802.3333333333334 ns	1.03
cuda/synchronization/context/auto	1165.7 ns	1203.5 ns	0.97
cuda/synchronization/context/nonblocking	7199.1 ns	7276.700000000001 ns	0.99
cuda/synchronization/context/blocking	900.1730769230769 ns	900.4347826086956 ns	1.00

This comment was automatically generated by workflow using github-action-benchmark.

[maleadt]

I don't see a Downgrade.yml being added here? And I'm not sure bluntly bumping all compats to the latest version is the way to go here; that'll just arbitrarily restrict compatibility with users' environments. What was the specific issue you ran into? I don't see anything being reported in FluxML/Zygote.jl#1583.

[ChrisRackauckas]

The last commit deleted it because you cannot pass due to the machine not having CUDA ff19390.

Why do you call this "bluntly"? Before this PR you cannot using CUDA with minimum versions, it will fail as shown in the CI logs. Before:

https://github.com/JuliaGPU/CUDA.jl/actions/runs/16853250942/job/47742615017

After:

https://github.com/JuliaGPU/CUDA.jl/actions/runs/16853423430/job/47743015124

Can you be more specific on which minimum versions you think I got wrong? Because according to CI the Project.toml previously allowed minimum versions that would lead to an unusable package and CI says that this now picks versions for which that is not possible. Is there a specific package you think was pushed to far up?

[maleadt]

Why do you call this "bluntly"? Before this PR you cannot using CUDA with minimum versions, it will fail as shown in the CI logs.

I didn't say all the minimum versions were correct (even though I expected them to mostly be), but simply removing them all is pretty blunt, no?

Can you be more specific on which minimum versions you think I got wrong?

Removing all compat versions and bumping them to the latest isn't "wrong", it just unnecessarily restricts installability of the package. So let's flip the question, which existing minimum bound is wrong? We can remove that one.

I do agree that a way to test these would be good; it's a shame https://github.com/julia-actions/julia-downgrade-compat is GH:A specific.

[ChrisRackauckas]

Starting by cutting anything out from earlier than 6 months ago and then refining it based on CI results was a heuristic that got to a resolvable set that was still fairly conservative in many respects. Do you have a better system? I think this set is pretty clearly better because it resolves unlike the previous version. Even if we did get a script that could fully brute force it, I don't think we have the compute resources to fully brute force it right now.

[maleadt]

Do you have a better system?

Yes, use Resolver.jl, just like the GH:A plug-in does. That gives me:

⌃ [621f4979] AbstractFFTs v0.4.0
⌃ [79e6a3ab] Adapt v4.0.0
⌃ [ab4f0b2a] BFloat16s v0.4.0
⌃ [fa961155] CEnum v0.2.0
  [1af6417a] CUDA_Runtime_Discovery v1.0.0
⌃ [a8cc5b0e] Crayons v4.0.0
⌃ [a93c6f00] DataFrames v1.4.0
⌃ [e2ba6199] ExprTools v0.1.0
⌃ [0c68f7d7] GPUArrays v11.2.1
⌃ [61eb1bfa] GPUCompiler v0.27.1
  [096a3bc2] GPUToolbox v0.3.0
  [63c18a36] KernelAbstractions v0.9.38
⌃ [929cbde3] LLVM v9.1.0
  [8b046642] LLVMLoopInfo v1.0.0
⌃ [5da4648a] NVTX v1.0.0
⌃ [21216c6a] Preferences v1.4.0
⌃ [08abe8d2] PrettyTables v2.1.0
⌃ [74087812] Random123 v1.2.0
⌃ [e6cf234a] RandomNumbers v1.5.3
⌃ [189a3867] Reexport v1.0.0
⌃ [ae029012] Requires v1.3.0
⌃ [90137ffa] StaticArrays v1.0.0
⌃ [10745b16] Statistics v1.10.0
  [d1e2174e] CUDA_Compiler_jll v0.2.0+0
  [4ee394cb] CUDA_Driver_jll v13.0.0+0
  [76a88914] CUDA_Runtime_jll v0.19.0+0
  [1e29f10c] demumble_jll v1.3.0+0
  [4af54fe1] LazyArtifacts
  [8f399da3] Libdl
  [37e2e46d] LinearAlgebra
  [56ddb016] Logging
  [de0858da] Printf
  [9a3f8284] Random
  [2f01184e] SparseArrays v1.10.0

[maleadt]

I've created a new PR with a CI action as well.