Googulator
Googulator
> > Hi [@Googulator](https://github.com/Googulator), CK currently only supports FA for MI-series cards; for example, https://github.com/ROCm/flash-attention has forward and backward attention with a CK backend for MI200 and MI300, but not...
Currently testing this fix in live-bootstrap.
Confirming that this fixes the issue.
Found a workaround: by loading the amdgpu kernel module with the option "sched_policy=2", the power draw stays low both in llama.cpp and the attached examples. Unfortunately, this causes a slight...
Debugging this further, it seems that excessive power usage starts when the offending operation (memory write or stream creation) creates a new HW queue. On RDNA3, this always uses MES,...
Thank you, GPU_MAX_HW_QUEUES=1 is indeed a viable workaround with fewer side effects than disabling hardware scheduling altogether.
The problem is only seen if you first boot Fiwix, and then kexec from Fiwix to Linux. Kexec'd Linux will then see an incorrect memory map. The issue was seen...
[kernels.zip](https://github.com/mikaku/Fiwix/files/14266248/kernels.zip) This contains the Fiwix and Linux kernels, as well as the kexec loaders for them, as compiled in live-bootstrap. These were captured from one of my bare metal test...
They are, but the syntax is slightly different in Linux than in Fiwix: `console=ttyS0`
Easiest way to reproduce is probably using live-bootstrap. Apply this patch to rootfs.py to get a serial log: ``` diff --git a/rootfs.py b/rootfs.py index c31d5a1..5d7df2c 100755 --- a/rootfs.py +++ b/rootfs.py...