Error network format body_with_headformat is not supported by cudNN backend

[JohnDijkstra]

When attempting to load a T1 network (e.g t1-512×15×8) into LC0 0.30.0 with latest cudaNN package, running on a GTX 1050ti GPU, I am getting this error:

Error network format body_with_headformat is not supported by cudNN backend.

Any idea how to fix this?

[mooskagh]

Hi JohnDijkstra,

The cudnn backend hasn't been updated for quite long time. For newer networks, use CUDA version rather than cudNN.

[JohnDijkstra]

Great, thanks! I'll try that in stead. John Dijkstra

…

________________________________ From: Alexander Lyashuk ***@***.***> Sent: Saturday, July 29, 2023 3:02:04 PM To: LeelaChessZero/lc0 ***@***.***> Cc: JohnDijkstra ***@***.***>; Author ***@***.***> Subject: Re: [LeelaChessZero/lc0] Error network format body_with_headformat is not supported by cudNN backend (Discussion #1904) Hi JohnDijkstra, The cudnn backend hasn't been updated for quite long time. For newer networks, use CUDA version rather than cudNN. — Reply to this email directly, view it on GitHub<#1904 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AXUJFVQN6TIXN4EE5TEWDFLXST3TZANCNFSM6AAAAAA2W4HBBA>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[artemmiyy]

Hey JohnDijkstra,

Were you able to fix the issue? If so, how did you use CUDA instead of cudNN? Please let me know.

Regards,
Artemii

[skorokithakis]

./lc0 --backend=cuda-auto

[mooskagh]

Were you able to fix the issue? If so, how did you use CUDA instead of cudNN? Please let me know.

Hi Artemii,

If you are running Lc0 under Windows, just download the version that has cuda in the name from here https://github.com/LeelaChessZero/lc0/releases (in the "Assets" section), rather than cudnn.

[artemmiyy]

I appreciate the fast reply. I'm running Linux and still not sure about how to fix the error, I have both backends (cuDNN and CUDA) installed with two GTX 1080 ti GPUs. Do you happen to know how to fix the error on Linux?

[mooskagh]

I appreciate the fast reply. I'm running Linux and still not sure about how to fix the error, I have both backends (cuDNN and CUDA) installed with two GTX 1080 ti GPUs. Do you happen to know how to fix the error on Linux?

First make sure that your binary is indeed compiled both with CUDA and CudNN backends. To do that, run ./lc0 --help and look at the list of possible values for the --backend parameter. cuda-auto should be one of them.

Then, to select that backend, either run lc0 with --backend=cuda-auto parameter, or set Backend UCI parameter to cuda-auto, or create file ~/.config/lc0/lc0.config and write backend=cuda-auto there.

For futher questions, I'd recommend to ask in the #help channel of our Discord chat at http://lc0.org/chat

[artemmiyy]

That worked – thank you!