Valentin Churavy

The way to do it is to not use AD over the entire function, but rather split your function into two parts and use AD o KY over the later...

I agree with Erik. That code piece is the most suspect. One must synchronize the CUDA stream before using CUDA memory in an MPI call. The allocation of the copy...

``` julia: /opt/x86_64-linux-gnu/x86_64-linux-gnu/sys-root/usr/local/include/llvm/IR/Instructions.h:1190: void llvm::ICmpInst::AssertOK(): Assertion `getOperand(0)->getType() == getOperand(1)->getType() && "Both operands to ICmp instruction are not of the same type!"' failed. [3341] signal (6.-6): Aborted in expression starting at...

This MWE confuses me. What is `Flux.gradient` and why does it take Enzyme data-structures?

I think this was fixed in 1.9, but for 1.8 https://github.com/JuliaPackaging/JLLWrappers.jl/pull/35#discussion_r779956552

This should be equivalent to https://github.com/llvm/llvm-project/blob/7cbf1a2591520c2491aa35339f227775f4d3adf6/llvm/examples/OrcV2Examples/LLJITWithCustomObjectLinkingLayer/LLJITWithCustomObjectLinkingLayer.cpp#L47 so I am surprised that we are hitting an assert.

We might be constructing a different TM, but otherwise things should be the same

https://github.com/ROCm/clr/blob/204d35d16ef5c2c1ea1a4bb25442908a306c857a/rocclr/os/os_posix.cpp#L301-L323 What is `versioninfo()` and `AMDGPU.versioninfo()`?

> Right. That segfaults even when I am using Julia single threaded. Yeah that's #767 :/

What is: ``` using AMDGPU import Libdl foreach(println, Libdl.dllist()) ``` > Ah, oops. Duplicate then No on my Archlinux system, versioninfo crashes, but other things work.