@qnixsynapse have you been able to replicate the issue? I've only seen it on Windows and I haven't been able to pinpoint the issue on Linux. I've seen multiple exceptions thrown when setting the tmp_buf using memcpy after the malloc_device in the reorder.

I'm asking because I am not certain if we may want to set the reorder value to disabled just for Windows builds. What do you think?

Edit: Could be a false hint though, as it's not always crashing there.

@qnixsynapse have you been able to replicate the issue? I've only seen it on Windows and I haven't been able to pinpoint the issue on Linux.

I'm asking because I am not certain if we may want to set the reorder value to disabled just for Windows builds. What do you think?

Edit: Could be a false hint though, as it's not always crashing there.

You are correct that this doesn't happen on Linux which I confirmed. But I feel for now it should be disabled by default until we resolve all the issues and implement for every quantization type.

My main concern is setting tensor extras in the init_tensor function. This PR disables that when reorder optimization is disabled.

It seems @sgeor255 found a fix in #13109 , we can merge this and once the fix is approved we can make sure it's re-enabled.

I see. Okay.