[TEST_DEBUG] Enable test device assert #5395
Conversation
dev-tomek
changed the title
|
There is no updates from pytorch/pytorch#142135 for a while, do you know what's the ETA for that? |
To my knowledge, sometime next year. I'll send you an internal ticket for that. |
|
|
||
| if should_fail: | ||
| abort_or_runtime_error = ( | ||
| result.returncode == 1 or # RuntimeError |
There was a problem hiding this comment.
In what cases do we get a RuntimeError? At first glance, I would expect all errors of SIGABRT type.
There was a problem hiding this comment.
If this is needed for other devices, then you can leave the old implementation for devices that are not XPU: if not is_xpu() and add a separate branch for us with what you wrote, perhaps this can simplify the code. This is optional because I don't know if it will result in less code.
There was a problem hiding this comment.
maybe easier to resolve merge conflicts
| mask_str = "None" if mask is None else str(mask) | ||
| opt_flag_str = "None" if opt_flag is None else str(opt_flag) | ||
|
|
||
| result = subprocess.run([ |
There was a problem hiding this comment.
You can set environment variables directly when calling subprocess.run, for example for TRITON_DEBUG. This way you can pass fewer parameters into test_debug_kernels.py.
Comes from #2755.
This PR is a temporary fix leveraging subprocess for catching the abort signal, enabling test_device_assert until pytorch/pytorch#142135 is resolved.