-
Notifications
You must be signed in to change notification settings - Fork 34
Open
Labels
Description
From previous tests, only the runner machine had issues running some atomics (all reduce for example) when using the vmem allocator. Could be a bad initialization flag. I am not able to reproduce the bug on other GPUs. Please use the vmem allocator, run all unittests and figure out a fix. For reference, consider shallow cloning https://github.com/ROCm/rocm-systems to understand how the low level apis. Your outcomes should be 2 things: 1: minimal repro that doesn't include any iris code that show the bug, 2: ideally a bug fix patch (I suspect the code had some wrong flag when allocating memory or something similar)
Reactions are currently unavailable