CUDNN_STATUS_NOT_SUPPORTED error

[SusanL82]

Hello,

When running Kilosort 4 from the GUI for a NP1.0, we get the same error message as described in issue #740.
It looks like the clustering finished, but I am not sure if this means that the results are (in)correct.
Everything also loads fine in phy so far.

Is this something we should worry about? And how could we fix it?

Thanks,
Susan

kilosort4.log

[jacobpennington]

Can you please copy-paste the full error message here, or include a screenshot of it from the GUI? The "CUDNN_STATUS_NOT_SUPPORTED" message is a generic error handle from pytorch that doesn't say much on its own, and they get missed by the log for some reason.

[SusanL82]

The whole error message looks like this:

C:\Users\neurolab.conda\envs\KSort4\lib\site-packages\kilosort\spikedetect.py:136: UserWarning: Plan failed with a cudnnException: CUDNN_BACKEND_EXECUTION_PLAN_DESCRIPTOR: cudnnFinalize Descriptor Failed cudnn_status: CUDNN_STATUS_NOT_SUPPORTED (Triggered internally at C:\cb\pytorch_1000000000000\work\aten\src\ATen\native\cudnn\Conv_v8.cpp:919.)
B = conv1d(X.unsqueeze(1), W, padding=nt//2)

I've attached the complete message window output in the textfile below.

KS4_CUDNNerror_SL.txt

[jacobpennington]

Please try setting max_channel_distance = 16 and then see if the error still shows up when you re-sort. The staggered contact layout you're using resulted in a situation where the default behavior still placed a lot of extra templates in the areas of the probe that were not being recorded from.