[CI] Enable v2 adapter testing on BMG on linux

[igchor]

No description provided.

[aelovikov-intel]

@sarnex , @uditagarwal97 , do we have enough runners capacity in CI for that?

[uditagarwal97]

@sarnex , @uditagarwal97 , do we have enough runners capacity in CI for that?

Here's the time taken by E2E jobs before and after this PR:
PVC: 6m 24s --> 14m 31s (https://github.com/intel/llvm/actions/runs/15544819604/job/43765014831?pr=18357)
BMG: 4m 26s --> 7m 14s

Overhead on PVC seems significant to me. IIRC, we also had one PVC runner also used for build. @sarnex have you seen PVC being a bottleneck recently? I think we should be good with BMG, although we have only one Linux BMG runner, I haven't experienced BMG being a bottleneck recently.

[sarnex]

PVC GPU hasn't been a blocker recently, but I agree the more than doubling of the job time is worrying.

[aelovikov-intel]

PVC GPU hasn't been a blocker recently, but I agree the more than doubling of the job time is worrying.

That seems to be due to the timeout, so not representative. That said, timeout in itself is a problem.

[sarnex]

Ah sorry, it would be nice to see some results we think are representative of normal runs.

[igchor]

Yes, I'm investigating the timeouts right now - it seems to be related to copy offload (so likely a driver issue). For now, perhaps we can enable the testing just for BMG?

[aelovikov-intel]

For now, perhaps we can enable the testing just for BMG?

👍

[igchor]

@uditagarwal97 @sarnex @aelovikov-intel I changed this PR to only enable testing on BMG. Here's a separate PR for PVC: #18909 Interestingely, on PVC 'Run E2E tests' takes almost exactly the same amount of time with v2 and without it (there's only a few seconds difference which seems pretty random) once I disable the time-outing tests.

[sarnex]

Test d:\github\_work\llvm\llvm\build-e2e\reduction\output\reduction_ctor.cpp.tmp.out hung!

is gonna be disabled right?

change lgtm otherwise

[igchor]

Test d:\github\_work\llvm\llvm\build-e2e\reduction\output\reduction_ctor.cpp.tmp.out hung!

is gonna be disabled right?

change lgtm otherwise

Actually, I only enabled the v2 testing on linux so this failure seems unrelated - I think I saw it failing on other PRs as well. I updated the PR title to make it clear it's only on linux for now.

[igchor]

@aelovikov-intel any other comments/concerns? If no, I think this should be ready for merge