When using Leela with last Network family T81-817216, the engine when downloaded get disconnected. However, Leela is working with old networks such as T79.
I also notice the size of the networks since T81 increased in size.
Please, fix it. Thanks!
When using Leela with last Network family T81-817216, the engine when downloaded get disconnected. However, Leela is working with old networks such as T79.
I also notice the size of the networks since T81 increased in size.
Please, fix it. Thanks!
Yeah I mentioned this back in July, still not working.
T81 and T82 also fails
So sorry about this! I’m seeing the error message on my end: “Network format NETWORK_ATTENTIONBODY_WITH_HEADFORMAT is not supported by CuDNN backend.”
After some digging, this seems to be the result of me using an outdated backend for lc0. I was using cudnn-fp16, but this post suggests that I should instead be using cuda-fp16:
T81 and T822 appear to be working now with the cuda-fp16 backend.
Apologies for leaving this broken for so long!
Shame on you…lol…but nice to see you on the ball as always