Rybka Chess Community Forum
Topic The Rybka Lounge / Computer Chess / LeelaChess crashing in UCI go nodes
By turbojuice1122 Date 2019-06-09 14:44
I am having trouble getting Leela Chess to work on my computer.  I used DriverBooster to update all of the drivers and restarted.  I have downloaded the cuda version and the weights from one of the networks into the same folder on my SSD, and when running lc0 and typing go nodes 100, it says,

Found pb network file: C:\LC0/weights.pb
Creating backend [cudnn]...

Then the window closes after a few seconds.  If I restart it and do the same thing, it closes after only about 1 second.

My computer has a 6-core Core i7-8700K at 3.7 GHz, 64 GB RAM, and an 8 GB NVIDIA GeForce GTX 1080

Any ideas?
By Master Om Date 2019-06-09 16:21
try to update CUDA files. Something like that Nvidia SDK files. See what happens.
By turbojuice1122 Date 2019-06-09 18:04
Thanks, I downloaded and installed the CUDA toolkit from the NVIDIA website, but it still doesn't work.  (It takes longer to crash this time, though, I think.)

Should I download and install NVIDIA's SDK manager?  It seems that I need a special account in order to do this.  Is this typical?
By MrKris Date 2019-06-09 21:29

https://github.com/LeelaChessZero/lc0/wiki/Getting-Started

Towards the bottom:
" Common problems
lc0.exe exits without printing anything
Double check you have cublas64_100.dll, cudart64_100.dll, and cudnn64_7.dll in the same directory as lc0.exe
"
(maybe the toolkit did not put them in the right place)

" Debug
For easier debug, verify install by running from command line (cmd.exe)
cd to directory with lc0.exe.
run lc0.exe --logfile=log.txt
Most recent weights file in that directory will be used automatically.
"
and see what the file log.txt says.
By turbojuice1122 Date 2019-06-09 23:25

I do all this, and in the log file, I get the following:

CUDA error: CUDA driver version is insufficient for CUDA runtime version

But...I just updated my CUDA drivers, and I made sure that my GPU is operating in maximum performance mode.
By Master Om Date 2019-06-10 18:44
Try The CPU only version and see. I think will work.
By turbojuice1122 Date 2019-06-10 22:38
Yes, this works, but it's very slow: 500 nodes per second.
By Peter Grayson Date 2019-06-10 10:02

> I am having trouble getting Leela Chess to work on my computer.  I used DriverBooster to update all of the drivers and restarted.  I have downloaded the cuda version and the weights from one of the networks into the same folder on my SSD, and when running lc0 and typing go nodes 100, it says,
>
> Found pb network file: C:\LC0/weights.pb
> Creating backend [cudnn]...
>
> Then the window closes after a few seconds.  If I restart it and do the same thing, it closes after only about 1 second.
>
> My computer has a 6-core Core i7-8700K at 3.7 GHz, 64 GB RAM, and an 8 GB NVIDIA GeForce GTX 1080

I have the same problem with my i7-8700 + 6Gb GTX 1060. Latest nVidia CUDA drivers installed plus other nVidia stuff. I also installed the recommended Microsoft development package. Wasted a lot of time on something that should be so simple to do but concluded it just was not woth the effort and wasting more time. In comparison to settting up an alpha-beta uci engine, for LC0 the user friendliness gets a zero rating!
By Uly Date 2019-06-10 16:57
Just checking, can you run the CPU-only version? I also wasted a lot of time trying to get Leela to work on my GPU, and it seems it was impossible, but the CPU-only version did work (I posted my thoughts about it in Talkchess, but the TLDR is that Leela at Depth 8, which takes 10 minutes to reach looks worse than TSCP at Depth 8, which takes a second, such a disappointment but I might buy a capable GPU just out of curiosity...)
By turbojuice1122 Date 2019-06-10 22:40
Yes, that works.  But the thing is, I very definitely have a capable GPU--there just seems to be something screwy with the program itself.
By Master Om Date 2019-06-11 16:42
now try the open GL version. See if it works.
By turbojuice1122 Date 2019-06-12 00:11
You mean the opencl version?  I tried this, and this works.

I'm a bit illiterate with this stuff, in spite of a little Googling.  What exactly is the opencl version?  I see it stated as Open Cloud, but other than that, I really don't know.
By Labyrinth Date 2019-06-12 16:55
Open CL (Open Computing Language), is basically CUDA but open source since CUDA is proprietary Nvidia and is made to only work on Nvidia hardware (there are ways around this though).

Both are software frameworks that enable the use of a GPU's processing power for unusual tasks.
By Master Om Date 2019-06-12 18:46
Me illiterate too. But see i found a work around.
By Lukas Cimiotti Date 2019-06-12 13:06 Upvotes 2
I'm running lc0 on a bunch of computers using CUDA on RTX 2080 and 2080Ti.
You don't have to install any NVidia stuff but the drivers.
Update Windows to version 1809, get all updates. Install GeForce Experience. Update the driver using this tool.
For me this works - I hope it will work for you, too.
By turbojuice1122 Date 2019-06-12 23:36
Thanks--though I can't get version 1809.  It makes me get version 1903, and upon attempted installation, says that I have hardware on the machine that isn't ready for this version.  So, I seem to be stuck with version 1803.
By Peter Grayson Date 2019-06-13 14:08 Upvotes 1

> Thanks--though I can't get version 1809.  It makes me get version 1903, and upon attempted installation, says that I have hardware on the machine that isn't ready for this version.  So, I seem to be stuck with version 1803.

I had the same message for my two, 10 year old Intel Q9550 machines when I ran the W10 compatibility tester. However, on the basis I had two machines running Windows 64 7 Pro that would become redundant in the not too distant future I decided to run the automatic W10 update on one of them to see what happened. Windows 10 64 Pro installed and is running fine and currently on version 1809. During installation there was a message regarding an incompatible graphics card but after installation I updated the Graphics card drivers and it switched to using the graphics card from Windows generic graphics drivers. The only comment I would make is to disconnect from the Internet once the installer has downloaded W10 because on my Dell dual Xeon machine it tried to run W10 updates during the installation from W7 Pro with the consequence the installation failed and it dropped back to W7. Just wasted time but it does seem to give that fall back safeguard if it fails to install properly.

I've just updated the dual Xeon to W10 Pro 1903 and no issues so waiting for W10 1903 update on the two Q9550 machines now.

Peter
By Peter Grayson Date 2019-06-14 21:54 Upvotes 1
With Windows 10 1903 update completed on the i7-8700 machine, I installed the nVidia 10.1 development pack and noted the graphics drivers were much later than the W10 installed drivers that reported they were up to date. Since installing the drivers from this package the LC0 CUDA engine is now working. Presently the performance on the nVidia GTX 1060 + 6Gb card looks to be similar to Brainfish 100619 64 BMI2 running on the i7-8700 CPU so a good match for comarison purposes.

Peter
By turbojuice1122 Date 2019-06-12 23:56 Upvotes 2
Update: the driver update with GeForce instead of Driver Booster was apparently the key.  The CUDA program now runs normally, and I'm getting 7 knps on the opening position.

Thanks, Lukas!
By Uly Date 2019-06-13 14:30
Hey, please keep up updated about what you think of Leela. I'm considering getting a decent GPU to run it (you were complaining of only getting 500 nps? I get 1 nps!) but the more data the merrier.
By Lukas Cimiotti Date 2019-06-13 17:10 Upvotes 2
If you can afford it try to get a latest generation NVidia GPU - GTX 1650 and up or RTX. These GPUs support fp16 calculations which makes them much faster and much more efficient. A RTX 2080 Ti gets ~ 60 kn/s, a RTX 2080 ~ 45 kn/s.
I made a lot of tests: lc0 on a RTX 2080 is roughly as strong as Stockfish 10 running on 10 cores on a modern Intel CPU @ 4 GHz.
By Uly Date 2019-06-13 21:00 Edited 2019-06-13 21:03
Thanks for the suggestions.

I am looking for something more like a... GeForce 710? At least I could buy it soon, while with the GTX 1650 I'd need to wait like 8 months to buy it or something...

How "indecent" would it be? I'm "upgrading" from a 4CPU Q8200 @2.33Ghz, so the question is if a 710 would be worth it, or I'd rather wait (and hopefully in the future I don't wait again lest I just wait and wait for something better.)
By Labyrinth Date 2019-06-14 02:29
So Geforce 710 is a low end card released January 2016, I'm surprised it's even a consideration. Even a geforce 1030 is like, 3x as fast, and I can't imagine that running Leela well. Even this is >4x as fast, although not sure about that seller.
By Uly Date 2019-06-14 10:08
I have problems getting stuff by mail, so I'm limited to buying what I can directly go to a shop and get. The 710 is appealing because I can go to a shop and buy it, and it's cheap. The 1030 costs almost twice as much, and at that price range it might be better to just wait for the 1650 (the 1650 is just the equivalent of US\$50 pricier than the 1030?).

But yeah, I'm ruling out the 710, seems it's not cost-efficient
By turbojuice1122 Date 2019-06-15 12:22
I actually have a GTX 1080, which is getting on the order of 5-10 knps, depending on the position, but according to the benchmarks, the RTX 2060 is slightly faster and a lot cheaper.
By Vegan Date 2019-06-23 01:53
The GTX 1080 is a powerful card but its designed for 3D graphics not AI

The RTX 2060 is the lowest cost card and it is not that good for games by comparison
By Labyrinth Date 2019-06-23 20:39

>The RTX 2060 is the lowest cost card and it is not that good for games by comparison

In comparison to what? 2080 Ti? An RTX 2060 is certainly 'good for games'.
By turbojuice1122 Date 2019-06-15 12:25
Have you done tests at longer time controls?  I have the impression that Leela benefits more from longer time controls/analysis times than do the mainstream engines--similar to how a human does.
By Lukas Cimiotti Date 2019-06-16 08:20
I only did tests at a time control of 4 minutes + 2sec. increment.
There is one problem with longer games: leela keeps the whole search tree in memory. It so uses a huge amount of memory. I upgraded my computers to 64 GB RAM.
By Master Om Date 2019-06-15 12:30
I made a lot of tests: lc0 on a RTX 2080 is roughly as strong as Stockfish 10 running on 10 cores on a modern Intel CPU @ 4 GHz.

Sir what about Lc0 running on GTX 1050 Ti ?
By Lukas Cimiotti Date 2019-06-17 11:31
It will work, but performance won't be great.
By Master Om Date 2019-06-17 17:31
Topic The Rybka Lounge / Computer Chess / LeelaChess crashing in UCI go nodes