Revert "[CI] Bump Level Zero version to 1.20.6"

[sarnex]

Reverts #17778

Same problem

[sarnex]

@npmiller @kbenzie @pbalcer Can you guys help us look into this? 1.20.6 also has no device found.

[npmiller]

I don't have experience with level zero, so I can't help much on this unfortunately

[pbalcer]

@npmiller @kbenzie @pbalcer Can you guys help us look into this? 1.20.6 also has no device found.

I ran this exact combination of driver/loader/igc and it worked fine. Admittedly I built it all from source.

@nrspruit any ideas?

[sarnex]

ok ill try to repro locally and do a deep dive, might need help from you guys as i dont know anything about level zero :P

[pbalcer]

ok ill try to repro locally and do a deep dive, might need help from you guys as i dont know anything about level zero :P

We have driver build automated in benchmark scripts.

You can test this like so:
./main.py ~/workdir --compute-runtime 25.09.32961.7 --build-igc --filter api_overhead_benchmark_sycl --sycl /path/to/llvm/build

This will first build the entire compute-runtime from source, with matching IGC, gmmlib and level-zero versions, and then run a simple test to verify whether it worked (there's a --verbose option if you want to see the commands it executes).

I just ran it and it worked.

[sarnex]

thanks, let me try that on the runner inside docker to see if the problem lies there somewhere

[sarnex]

@pbalcer @nrspruit I found the problem, and it's way simplier than I imagined.

The L0 releases no longer have any binaries. We use the deb binaries to install L0. Since there are no binaries the install fails and thus we have no L0 so no devices are found.

@nrspruit Is it intended that the newer L0 releases have no binaries?

Last good example (1.20.2) here
Bad example here

If this is intentional I'll change our scripts to build L0 from source.

[sarnex]

@nrspruit Can you comment on the above finding? Thx

[sarnex]

Fixed in #18016, thx all!