KataGo Distributed Training and new networks

[And]

@PC_Screen
thank you! it works, but KataGo throws an error on the second move (in Sabaki and on the command line). probably not enough video memory.

Code: Select all

Uncaught exception: CL_OUT_OF_RESOURCES <possibly ran out of GPU memory?>

EDIT yes, there is not enough video memory - 7x7 works
could you make an eigenavx2 version?

[hope366]

Thanks for uploading the engine files.
I tried "katago.exe" which was put on Google Drive.
I've tried with Lizzie and LizzieYzy, but I get an error and the engine doesn't work. The error message said something like "If you have ucrtbased.dll, it will work", so I got it from the internet and put it in the folder where katago.exe is located.

I wasn't expecting much, but this worked.
It's working with Lizzie and LizzieYzy, but it's noticeably slower compared to 40 blocks.
It is about 60 to 70% of the speed of 40 blocks.
I thought it was a problem related to FP16, so I tried changing the FP16 setting in the configuration file from "auto" to "true", but there was no change in speed.
Any good ideas?

[And]

I managed to run KataGo 18b on 19x19, and the speed has increased significantly! how I did it:
1 started tunning with a 10b network
2 renamed the resulting file to "tune8...x19_y19_c384_mv11.txt "
and that's it!

EDIT will it work for you? ...

[PC_Screen]

Katago Eigen, same process, drop it inside the katago 1.11 eigen folder and replace the old katago.exe.

https://drive.google.com/file/d/18XJHc8 ... sp=sharing

As for the speed being lower, try running the benchmark in the command prompt (./katago.exe benchmark -model <NEURALNET>.bin.gz) and adjust the number of search threads based on it

[And]

KataGo (18b, playouts 1) - CS Zero 9d

[hope366]

>EDIT will it work for you? ...

I tried that method, but unfortunately I didn't see any speed improvement.
I also ran a benchmark test and it's slower than 40 blocks due to false FP16 relationships.

CUDA version or TensorRT version will enable FP16 and speed will be faster.
The katago.exe provided by @PC_Screen seems to be the OpenCL version.

[And]

second game KataGo (18b, playouts 1) - CS Zero 9d

[And]

Similarly replaced the tuning file for the 40b network, the time to move 40b is 20% longer than that of 18b (playouts 1, gt 610)

[And]

KataGo 18b, playouts 10 - 40b, playouts 10

[And]

second game KataGo 18b, playouts 10 - 40b, playouts 10
(40b-s118)

[hope366]

If anyone has a CUDA version of KataGo for the new architecture, could you please upload it?
I tried to compile using cmake before, but it didn't work in my environment.

[And]

and the version of eigenavx2 for Windows please!

[And]

KataGo 18b, playouts 10 - 60b, playouts 10

[And]

Katago Eigen, same process, drop it inside the katago 1.11 eigen folder and replace the old katago.exe.

https://drive.google.com/file/d/18XJHc8 ... sp=sharing

As for the speed being lower, try running the benchmark in the command prompt (./katago.exe benchmark -model <NEURALNET>.bin.gz) and adjust the number of search threads based on it

many thanks! it works, with different networks, and 18b works. but the speed compared to eigen, eigenavx2 1.11.0 is tens times slower. something's wrong.

visits: 1
10b eigen 1.11 Time taken: 0.281
10b your file Time taken: 15.21

60b eigen 1.11 Time taken: 2.605
60b your file Time taken: 160.166

18b your file Time taken: 56.941

[And]

second game KataGo 18b, playouts 10 - 60b, playouts 10
(60b-s665)