`LLVM_jll=86de99a1-58d6-5da7-8064-bd56ce2e322c` depends on `libLLVM_jll=8f36deef-c2a5-5394-99ed-8e07531fb29a`, but no such entry exists in the manifest.

[giordano]

Since a couple of days ago, we're seeing this (or related) error(s) on CI in many projects:

`LLVM_jll=86de99a1-58d6-5da7-8064-bd56ce2e322c` depends on `libLLVM_jll=8f36deef-c2a5-5394-99ed-8e07531fb29a`, but no such entry exists in the manifest.

This happens also on julia nightly: https://github.com/JuliaGPU/GPUCompiler.jl/actions/runs/19709438203/job/56465806309?pr=745#step:5:125.

(note, this is the same error as #4536, but that issue was about the error message itself).

CC @vchuravy.

[IanButterworth]

I believe it's also been seen on 1.12.1, 1.12.2, and 1.13.0-alpha1, is that correct?

[giordano]

Definitely 1.12.1, 1.12.2, not sure about v1.13.

[IanButterworth]

And clearing the depot fixed it? (I assume CI there reuses/caches the depot between runs)

[giordano]

That's what happened in Oceananigans (they have self-hosted runners, so full control of the machine), I don't think that was attempted for GPUCompiler (but they use github-hosted runners, I guess they'd need to delete the cache, if any)

[giordano]

Actually, based on https://github.com/JuliaGPU/GPUCompiler.jl/actions/runs/19709438203/job/56465806309?pr=745#step:4:127 the cache wasn't hit at all, so the depot should be completely new?