Ben Vanik
Ben Vanik
I just hit a flake in sort on the MI250, which may be related. https://github.com/iree-org/iree/actions/runs/14038781847/job/39303988986?pr=20355 ``` 122/176 Test #733: iree/tests/e2e/stablehlo_ops/check_rocm_hip_stream_sort.mlir .....................................***Failed 0.53 sec [==========] Running 4 tests from 1 test...
(I suspect by disabling the topk test we now are just hitting the next test that has the same root bug - maybe sort is easier to diagnose than the...
I'm not saying I'm for or against any proposal here, only voicing that I agree what's in the API today does not fully satisfy workflows with dynamic data moving across...
is this a duplicate of #16025? (want to make sure they're the same effort - probably worth picking one issue to use?)
Nice reproducer!
Root cause is probably L183 in ReferencePartitioning.cpp, doing the check on AsyncBarrierOp. I suspect we want to remove the AsyncBarrierOp carveout there - it should go with the consumer, not...
if you remove a field it breaks the files - that is what the deprecated attribute solves
We'll always need to keep the field until we make a breaking format change. Give it a release or two of always being 0 and then we can remove support...
Let's not break the format - keep it deprecated for now.
(if you want to error on older files, then have the executable loader in the runtime code error out if the field is present and non-zero)