Skip prepare_state in GC finalizers

src/tls.jl

@@ -189,7 +189,11 @@ function priority!(f::Function, p::Symbol)
     end
 end
 
-@inline function prepare_state(state = task_local_state!())
+@inline function prepare_state(state = nothing)
+    # Finalizers run on whichever task GC interrupted and pick their own context
+    # with `HIP.context!`, so leave that task's state and sticky error alone.
+    GC.in_finalizer() && return
+    isnothing(state) && (state = task_local_state!())
     HIP.clear_last_error() # Drain any sticky HIP error left by a prior kernel failure
     hip_ctx = Ref{HIP.hipCtx_t}()
     HIP.hipCtxGetCurrent(hip_ctx)

test/core/tls.jl

@@ -15,6 +15,19 @@ using AMDGPU: ROCArray, HIPDevice, HIPStream
     @test AMDGPU.device(x) ≡ d1
 end
 
+@testset "GC finalizers leave task-local state alone" begin
+    # GC runs finalizers on whichever task it interrupts. Freeing an array makes
+    # HIP calls, which must not create state on that task.
+    weak_array() = WeakRef(ROCArray{Float32}(undef, 16))
+    w = weak_array()
+    no_state, collected = fetch(@async begin
+        GC.gc(true)
+        (AMDGPU.task_local_state() ≡ nothing, w.value ≡ nothing)
+    end)
+    @test collected
+    @test no_state
+end
+
 @testset "Stream" begin
     s1 = @inferred AMDGPU.stream()
     @test s1 isa AMDGPU.HIPStream