[devmotion]

Fixes #865.

Problem

#855 removed rand!(::AbstractRNG, ::AbstractArray) because for Random's RNGs it forwards to the sampler rules. GPUArrays and CUDA, however, implement rand!(rng, A) directly without a sampler, and forward all their out-of-place rand(rng, T, dims...), rand(rng, dims...) methods to it. Zygote therefore differentiates into their kernels.

Change

Restore only @non_differentiable rand!(::AbstractRNG, ::AbstractArray), the lowest-level function that all GPU paths reach. Its output depends only on eltype(A), and the previous values of A are overwritten.

Not added, compared with the proposal in #865:

• rand(::AbstractRNG, ::Type, ...), rand(::AbstractRNG, ::Integer...): forwarders to rand! in GPUArrays and CUDA.
• rand(::AbstractRNG, ::Dims): forwarder as well; moreover, in Base rand(rng, (5, 7)) samples from the tuple.
• rand!(::AbstractArray): Base and CUDA forward it to the 2-arg rand!.
• rand!(::AbstractRNG, ::AbstractArray, ::Type) would not help: GPU packages define the 2-arg method, so it is never reached.

Zygote on JLArrays (GPUArrays reference backend)

	main	this PR
rand!(GPUArrays.RNG{JLArray}(...), A), rand!(Xoshiro(1), A)	error	✅
rand(GPUArrays.RNG{JLArray}(...), Float32, 3), same with (3,), rand(rng, 3), rand(rng, (3,))	error	✅
CPU: rand(rng, [x]), x .* rand!(rng, zeros(3))	✅	✅ (unchanged)

Not verified

No CUDA available. rand!(CUDA.default_rng(), A) hits the rule directly. rand!(A) and rand(CUDA.default_rng(), Float32, 3) reach it after library_rng() and CuArray{T}(undef, dims), which, judging from the errors in #865, Zygote already traces fine. @CarloLucibello, could you re-run your table on this branch?

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.65%. Comparing base (a6107e9) to head (21093c8).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #866   +/-   ##
=======================================
  Coverage   96.65%   96.65%           
=======================================
  Files          28       28           
  Lines        3614     3615    +1     
=======================================
+ Hits         3493     3494    +1     
  Misses        121      121           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.