[Issue]: change_current_allocator failed

[hebais]

Problem Description

when import iris, I have encountered this error
RuntimeError: Can't swap an already initialized allocator, since torch could not change allocator once it is already initialized.
I have commented this torch.cuda.memory.change_current_allocator and it seems the example still works correctly.

So what might be the side effect if I commented this hip allocator ? Actually I did not quite understand why it becomes fine grained after switch to hip Allocator. Thanks

Operating System

22.04.5 LTS (Jammy Jellyfish)

CPU

Intel(R) Xeon(R) Platinum 8458P

GPU

MI300X

ROCm Version

ROCm 6.3

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[mawad-amd]

Hi @hebais, to summarize:

• Fine-grained memory enables fine-grained coherence, allowing communication at the system-scope level while a kernel is running.
• Coarse-grained memory only guarantees coherence at kernel boundaries (via hipDeviceSynchronize, hipStreamSynchronize, or similar).

On MI300X and MI350, both behave the same, so your change does not affect correctness. I would like to actually remove the fine-grained allocator. If you are open to creating a PR with your changes, I would be happy to review it.

See more details here.