When I launch the app now it's say I'm already logged u_u and also my bot lost to zen11.4 lol it's impossible normally. When you don't have the good commands put in the txt file it's not the same. That's why i would prefer testing directly my bot against another user instead of using this complicated website…

Yes, perhaps should rather focus on that then.


On a related note, Leela doesn't seem to utilize resources very well on Windows. Right now I'm running Leela_Test, and CPU load rarely exceeds 70% at peak, GPU load rarely exceeds 80%. RAM usage is at about 250 MB and GPU memory at 474 MB (including OS and other apps). Also, it doesn't seem to ponder by default, so I wonder why the --noponder param is provided.

Also testing Leela right now.
CPU user far less (~70%) than when I run Python script from github to analyze games with Leela. In that case CPU usage is 100%. And this is strange.
Video also never user fully, right now - usually 60%.

Leela parameters can be found in read.me file in leela folder.
But not much can be used since time is limited.

Seems like the CPU-only version uses the CPU at 100% all the time. It's the OpenCL version that seems bottlenecked by the GPU, while not completely using all of it.

What does --noponder do? Havne't find any clear explanation anywhere? Does it make leela stronger or weaker?

Yeah, leela works much better on linux, which is really annoying.

Leela will normally think during the opponent's time. You can turn that off using --noponder. Makes her weaker.

Leela does ponder by default, it might be the order the time_left and genmove commands are sent to Leela by the cgos script, I'm not sure. It seems to turn her pondering off.

As someone mentioned it ealier - CPU/GPU is off during opponent time, even if there is no option 'noponder', so seems it simply doesn't work.

Got Leela setup on my Ubuntu 16.04 boot now. CPU/GPU utilization reaches 100%, but memory footprint is still surprisingly small. It's below 80 MB in RAM and ~180 MB in GPU memory.

As for the pondering: That's a function of genmove versus kgs-genmove_cleanup. Simple genmove keeps pondering after move generation, while kgs-genmove_cleanup stops as soon as the move is generated.

Also, looking to compare Leela's internal benchmarks.

Ubuntu 16.04
benchmark
200000 games in 8.02 seconds -> 24937 g/s (3117 g/s per thread)
Avg Len: 418.62 Score: -2.561980
netbench
2000 predictions in 3.99 seconds -> 501 p/s
10000 evaluations in 2.73 seconds -> 3663 p/s

Windows 10
benchmark
200000 games in 13.02 seconds -> 15360 g/s (1920 g/s per thread)
Avg Len: 418.63 Score: -2.444100
netbench
2000 predictions in 4.19 seconds -> 477 p/s
10000 evaluations in 5.59 seconds -> 1788 p/s

That's quite a huge gap.

Are you using VM for Ubuntu?
I tried VMWare Workstation, but as I understand it simply doesn't support to use GPU and provide only virtual one(

Also seems with provided time settings from server - it still doesn't use time effectively. For longest games with 300-320 moves - still 2.5-3+ min left. And most of games finished when still 30% left. No idea how tune it to use a bit more time. Right now trying to edit cgosGtp.tcl but doubt it will help(

Еще похоже Лиила по прежнему очень часто "зевает" некотоыре достаточно очевидные хода
Постоянно наблюдается ситуация, когда Лиила думает что впереди (например 60%), а потом после 1-2 ходов соперника винрейт падает до 20-30%.

I have a dual boot instead of a VM.

I don't think the time management can be much improved in a sudden death time system. Just too risky to time out at some point.

Can't run Ray 4.20.4 bot on server (locally everything is ok).
On genmove - engine crashed.
Tried both genmove and kgs command. Tried remove/add parameters - no effect.
May be someone face such issue already?

What kind of server and what kind of OS? Does the engine initialize at the beginning without a problem, before attempting a genmove?

Win10 x64, CGOS.
Yes, initialization is ok. Game started and only after genmove it fails. Error printed by 479 line of function 'genmove'.

Oh, now I know what you mean. That's something I wrote about in my first post. Rayn, when using the GPU, puts out debug info that conflicts with the CGOS script. I use a modified Python script as CGOS interface, but it should be possible to modify the TCL script to do the same.

Right now I'm testing a tcl script for Leela that allows pondering though, but I will look into it afterwards.

May I ask if you can share that modified python script please?

I tried run Leela with pondering (with original genmove) and faced issue in the game only once - some reason opponent (ray) passed in the middle of game and leela passed too. Somehow they counted game in the middle and opponent won) Even though he was dead)
But can't say that strength increased as it should... probably something was wrong. Bot stuck around 3400.
Almost all games lost because of wrong judgement of group statuses.

In Tromp Taylor rules every stone is alive. Dead stones of your opponent should be killed before you pass. This is why Leela needs the kgs-genmove_cleanup command which will behave that way.

The appended tcl script should allow Leela to ponder on CGOS. It basically uses genmove until that generates a "pass" for the first time, at which point it switches over to kgs-genmove_cleanup (where the non-pondering shouldn't matter anymore as this shouldn't be performance-critical), before resetting when the next game starts.