How to run/obtain LC0-cudnn?

[Albert Silver]

I noticed in the benchmarks a number of entries using lc0-cudnn. How can I use it or get it?

Thanks,

Albert

[Alexander Lyashuk]

Hi, 

To use it, you have to install CUDA and cuDNN from Nvidia site. (Downloading cuDNN requires registration on their site).

After that you can either build you version from sources (instructions https://github.com/glinscott/leela-chess/blob/next/lc0/README.md)

or use binaries that I've built:

https://crem.xyz/lc0/

Note: versions for Windows require CUDA v9.0 (not v9.1).

Version for Linux is for v9.1.

It's possible to build it for CUDA v9.1 for windows, I just didn't have time to set up the environment.

It may contain bugs. As of yesterday, it was weaker than "main" lczero on the same number of nodes, but today hopefully the reason for that is fixed, tests to confirm/reject that are currently ongoing.

One particular thing that may affect strength (in either way) is "virtual loss bug emulation", which can be enabled using --virtual-loss-bug command line flag or "Emulate virtual loss bug" UCI parameter.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/6a80ff31-086f-48cf-84fb-6ed8b67378af%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Albert Silver]

Hi Alexander,

Many thanks. I actually used 9.1.... and renamed the 2 DLLs to 90 where needed. Worked fine. :-)

The speedup is amazing. I went from 2200 NPS after a minute to 12400 NPS after 30 seconds (9350 after ply 26, but that was too fast). I will run a match to test. Previous NN202 was ~2900 CCRL on a GTX1060 (with modest i5-2500K)

Albert

[Edward Panek]

Using CUDA 9.1 but having problems displaying PV in Arena. Thinking about trying 9.0 and your compiles. Any strength diff between 9.1 and 9.0?

[Albert Silver]

Add this to the commandline: -w weights.txt

It seems to start searching for the weights file each time if you do not.

Albert

[Nay Lin Tun]

I am stacked in installing cuda 9.0 or 9.1, I installed visual studio 2015 and 2017 already. Any advice? My card 1060 GTX, windows 10, 64 bit. 

[Nay Lin Tun]

I installed latest version of nvidia driver already. 

[Alexander Lyashuk]

If you don't want to build it yourself but just use prebuilt binaries, uncheck "Visual Studio Integration" checkbox during CUDA installation.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/eca881fe-35d3-467a-88fb-be338ae7227f%40googlegroups.com.

[Edward Panek]

I just installed the individual files in the engine directory and it worked

[Alexander Lyashuk]

Ah yes, forgot to mention that unlike the ordinary lczero.exe, it doesn't support gzipped weight files yet, so you have to unzip them.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/58fa4d1c-30af-4f0f-8468-f5fcff0a115d%40googlegroups.com.

[Nay Lin Tun]

would you mind sharing your working files ? 

[Alexander Lyashuk]

Note that distribution of cuDNN is not allowed (that's common understanding of their license), everyone should register on the Nvidia website and get their own copy

1. CUDA 9.0 is here.
   

2. cuDNN for CUDA v9.0 is here (you have to register and fill a survey)

So please don't publish packages of CUDA/cuDNN dlls on this forum (or anywhere, especially when it's clear that it's related to lczero project).

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/d9cde708-8e03-4805-920e-6f3bb2ca138d%40googlegroups.com.

[solitarycandle]

According to the cuDNN distribution of applications that are linked to cuDNN is possible. So we could distribute executables for all platforms with the required .dll, .so and license. Redistribution of the entire toolkit or bundling it with a hardware solution is not allowed.

3.1. CUDNN DISTRIBUTION

Subject to the terms of the SLA and this Supplement, NVIDIA hereby grants you a non-exclusive, non-transferable license during the applicable license term unless earlier terminated pursuant to the SLA, to distribute the runtime files (.so, .h) when delivered to you as part of the cuDNN Licensed Software in source code form or binary form (but not when provided to you as part of a hardware product), subject to such distribution being solely in binary form to your licensees (“Customers”) only as a component of your own software products having additional material functionality beyond the redistributable cuDNN Licensed Software (each, a “Licensee Application"). Subject to the terms and conditions of the SLA and this Supplement, you may further authorize Customers to redistribute the cuDNN Licensed Software as incorporated into a Licensee Application, solely in binary form, provided, however, that you shall require in your agreements with your Customers that their distributions be on terms at least as restrictive as those applicable for your use of the cuDNN Licensed Software within a Licensee Application. The expiration or termination of your licenses to the cuDNN Licensed Software under the SLA and this Supplement will not affect your previous cuDNN distributions in compliance with the SLA and this Supplement.

The above distributions are subject to the following:

1. all distributions by you or your distribution channels must be consistent with the terms of the AGREEMENT;
2. the distributed cuDNN License Software must include valid copyright notices indicating NVIDIA’s ownership of the Licensed Software and (if permitted) modifications; and
3. you must enter into enforceable agreements that pass down terms consistent with the terms set forth in the AGREEMENT for use of the distributable cuDNN License Software, including (without limitation) terms relating to the license grant and license restrictions, confidentiality and protection of NVIDIA’s Intellectual Property Rights in and to the cuDNN Licensed Software.

[Albert Silver]

I'm running into what must be a bug. In matches of ten minutes, it starts at a high NPS, but after 10-20 seconds, the average NPS is well under 2 (GTX1060), so somewhere it is hitting a wall. The 'hashtable' or whatever it is, shows 100% from the first 2-3 moves and remains at 100% throughout. I don't know if this is normal, so am just mentioning it.

Albert

On Tuesday, May 1, 2018 at 2:59:12 PM UTC-3, Alexander Lyashuk wrote:

[Albert Silver]

I should point out that I reset the cPUCT to 0.85 like the unmodified LCZero and turned off the Smart Pruning, in order to compare similar settings to the normal 0.7 and see how it performed.

[Albert Silver]

Sorry, one last thing: the task manager is showing only about 10% of the GPU used. Is that normal? Even with a GTX1060? 

[tun nay]

Ok, thanks. I got through installation of cuda 9.0.

[Alexander Lyashuk]

This version has position evaluation values from -1 to 1 internally (like in AlphaZero), unlike the ordinary lczero which has it from 0 to 1 (heritage from leela zero, which took it from AlphaGo Zero). With that difference, equivalent value of Cpuct in lc0 should be two times larger than Cpuct in lczero.

So what is 0.85 in lczero, is 1.7 in lc0.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/7464582d-f919-4448-9a33-d0ad474ebcab%40googlegroups.com.

[Graham]

On Wednesday, 2 May 2018 14:44:21 UTC+12, Albert Silver wrote:

Sorry, one last thing: the task manager is showing only about 10% of the GPU used. Is that normal? Even with a GTX1060?

Same for me. With LC0 GPU my GTX970 was generally showing around 50% used (don't know why it never went higher: maybe because the GPU is also controlling my monitor?). With the CuDNN version it starts off on the first moves at around 30-40% but then drops to around 11%. At the same time the NPS goes back to the 2K level that I saw with the GPU version. I'm using Win10 with CUDA 9.0.

I noticed there are also a lot of differences in the UCI params compared to version 0.7. For example, I found that setting Slowmover=120 in the UCI settings seemed to lead to better time management but that option has gone in CuDNN. Also in Arena the output from LC0 is saying "info depth 2 seldepth 21..." for example, so Arena shows the search depth as 2 all the time.

Would it be possible to have a CDUNN version that is just a compile of the standard 0.7 client?

[Frank Müller]

Also at least on my slow gpu the cuda version needs much too long at startup. Actually it needs so long to load the NN and probably the cudnn.dll
(at least ram usage suggests it) that cutechess-cli at least cancels all games with it before it could make its first move.

Nps would be around twice as fast here as the standard OpenBlas version.

[Alexander Lyashuk]

Maybe it's weights file autodiscovery that is slow? Some people said so.

When weights file is not passed (vie command line flag or uci parameter), the engine scans whole subdirectory at where the .exe located and for every file larger than 30M checks whether it looks like network file or not.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/dc7cc317-bb5b-41fb-9d22-570ec6f67b4c%40googlegroups.com.

[Frank Müller]

I don't think so. By adding the NN as param the process is as slow as before.
It just slowly increases its ram usage up to around 600MB before it shows its first thinking line.
IMO this is the loading of several dlls plus NN (cudnn.dll alone is around 322MB)

My estimation of loading time is between 15 and 25sec here!
Of course it is possible that the main reason is my very slow, but CUDA enabled gpu.
It is a simple GT 710.

[Albert Silver]

Have the weights been trained that way too? Otherwise it seems this might throw it off, no? I stopped an early match as the build is badly underperforming.

[Alexander Lyashuk]

Network computes value from -1 .. 1, it's the "old" lczero implementation that rescaled it to 0 .. 1 immediately after the computation is done.

It's weird that the lc0 build is underperforming, people reported it finally being the same strength as lczero, if "virtual loss bug emulation"  is on.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/9a08d85b-67db-4129-9858-f1de689aebe0%40googlegroups.com.

[Albert Silver]

"virtual loss bug emulation" It was not on for me as it was off by default and I did not know what it was for. Still, I was rather expecting a strength increase with the speedups, not an effort to perform the same. How well tested are each change aside from this? 

[Alexander Lyashuk]

It performs the same per number of nodes visited, not per time. (So, per time it should be somewhat stronger).

It's not tested well so that may still be weaker, but it shouldn't be easily visible (like obvious blunders, consistently losing more games that lczero, etc), but rather needs stats from at least 10-20 games.

If you see moves which is obviously wrong which old lczero can find, or a game which looks like too weak and for which it seems that lczero would win it, I'll be happy to debug if you send me a PGN.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/33d358a2-1727-4709-9bb0-e3fa7b433bad%40googlegroups.com.

[Albert Silver]

Sorry, I was not clear. Please forgive me if this is old news: in the classic chess engine development cycle, each and every change, however small, whether in evaluation, algorithm or just code cleaning, is subject to massive testing to be sure it is either beneficial or at worse, not detrimental. For example: Rook value is changed by 0.05? 50 Tens of thousands of games will be played between the before and after versions to compare results, until 95% confidence. 

Until about 10 years ago, it was not done this way, and a single author, Vasik Rajlich who imposed this scientific method, dominated the computer chess world like no other. Nowadays it is accepted that this is the only way to work for maximum progress, since it only takes a single half-broken idea to cripple or negate all the rest.

I saw that this build includes many changes compared to the current 0.7 build. So I was asking how well tested each of these changes was.

Albert

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+unsubscribe@googlegroups.com.

[Albert Silver]

I forgot to mention that he did this with tens of thousands of ultra fast games played in a second or two total. It had been previously thought that this was not 'deep enough' to test the value of an idea, his games were played at 3-4 plies maximum, but since then, this way of thinking has been proven erroneous. 

Albert

[Alexander Lyashuk]

At this point both lczero-0.7 and lc0-cudnn are in very early stage and changes are either done with very rough ballpark tests, or even using just intuition.

That's true for both lczero-0.7 and lc0-cudnn. For lc0-cudnn the main requirement so far is actually algorithm equivalence to lczero (so that switch would be possible) so the question to test is currently not "how strong it is" but rather "how close it to 0.7",

and for 0.7 aggressive parameters tuning/optimization is also discouraged at this point, as network strength still growth, and having other variables confuses what the source of the skill change is.

Also there's no computational capacity to run thousands of tests games. That's especially true as optimal parameters may differ from network to network, and with few new networks per day, there's no chance it can be realistically tested.

I'm sure that when network reaches it's limit, the approach to tune the algorithm/network will be much more well defined and all changes will be more thoroughly tested.

[Albert Silver]

On Wednesday, May 2, 2018 at 9:45:28 AM UTC-3, Alexander Lyashuk wrote:

At this point both lczero-0.7 and lc0-cudnn are in very early stage and changes are either done with very rough ballpark tests, or even using just intuition.

That's true for both lczero-0.7 and lc0-cudnn. For lc0-cudnn the main requirement so far is actually algorithm equivalence to lczero (so that switch would be possible) so the question to test is currently not "how strong it is" but rather "how close it to 0.7",

and for 0.7 aggressive parameters tuning/optimization is also discouraged at this point, as network strength still growth, and having other variables confuses what the source of the skill change is.

Also there's no computational capacity to run thousands of tests games.

Why not? You can run 50-100 thousand games in a day easily with a single modest PC, More in fact, since only one core is needed. This is more than enough to test a parameter. Don't underestimate how detrimental one problem change or two can cause. You could add 15 function, brilliant that add 100 Elo, and two bad ones that seemed logical, that negate it all or worse. Right now, cuDNN presumably triples or quadruples the NPS and depth by a factor of 2 plies or more. In classical Alpha-Beta engines, with a branching factor of 2, that is about 100 Elo, yet it is barely performing at the same level as 0.7. This screams there is something wrong. Can you build a 0.7 with cuDNN, and no other change, to test? This would at least create a baseline for reference.

 

That's especially true as optimal parameters may differ from network to network, and with few new networks per day, there's no chance it can be realistically tested.

It is extremely unlikely that a parameter will differ from network to network. The parameters are not to the evaluation, which is what the network does. Plus there are tools for specifically finetuning a single parameter automatically.

[Alexander Lyashuk]

On Wed, May 2, 2018 at 3:14 PM Albert Silver <silver...@gmail.com> wrote:

On Wednesday, May 2, 2018 at 9:45:28 AM UTC-3, Alexander Lyashuk wrote:

At this point both lczero-0.7 and lc0-cudnn are in very early stage and changes are either done with very rough ballpark tests, or even using just intuition.

That's true for both lczero-0.7 and lc0-cudnn. For lc0-cudnn the main requirement so far is actually algorithm equivalence to lczero (so that switch would be possible) so the question to test is currently not "how strong it is" but rather "how close it to 0.7",

and for 0.7 aggressive parameters tuning/optimization is also discouraged at this point, as network strength still growth, and having other variables confuses what the source of the skill change is.

Also there's no computational capacity to run thousands of tests games.

Why not? You can run 50-100 thousand games in a day easily with a single modest PC, More in fact, since only one core is needed.

As you can see on a table on the main page, http://lczero.org/, "a single modest PC" can only generate ~500 games per day on average, on 800 nodes per move, which is already arguably too few to reliably evaluate MCTS.

if average PC could generate 100'000 games, total count for 580 users would be not 200'000.

And even for 500 games a day, we don't have a spare machine to run tests on it.

Also the main developer of LCZero and GitHub repository owner is the author of Fishtest, the framework to test Stockfish changes. I actually expect him to have the best intuition when it's the time to start proper code change evaluation (when time testing worth more than time developing), so far he didn't say we should start doing that.

 

This is more than enough to test a parameter. Don't underestimate how detrimental one problem change or two can cause. You could add 15 function, brilliant that add 100 Elo, and two bad ones that seemed logical, that negate it all or worse. Right now, cuDNN presumably triples or quadruples the NPS and depth by a factor of 2 plies or more. In classical Alpha-Beta engines, with a branching factor of 2, that is about 100 Elo, yet it is barely performing at the same level as 0.7. This screams there is something wrong. Can you build a 0.7 with cuDNN, and no other change, to test? This would at least create a baseline for reference.

No, cudnn version and 0.7 are completely different and there's no easy way to use cuDNN implementation in 0.7.

 

 

That's especially true as optimal parameters may differ from network to network, and with few new networks per day, there's no chance it can be realistically tested.

It is extremely unlikely that a parameter will differ from network to network. The parameters are not to the evaluation, which is what the network does. Plus there are tools for specifically finetuning a single parameter automatically.

In this case you cannot trust your intuition and say something is extremely unlikely.

Take for example FPU reduction. If policy network is trained well enough to see tactical complications, you can increase it trusting the network more. If network doesn't see some tactics well, it's better to reduce it so that search is shallower but thorough.

And from network to network what NN see does change, and (rough ballpark) tests shown that this parameter has to be tuned for every network individually.

 

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/75050553-abb8-44be-8664-1bf96ed0fc18%40googlegroups.com.

[Albert Silver]

Ok, then I guess the conclusion is that there is no compelling reason to use cuDNN if there is no gain other than benchmarks.

[Gian-Carlo Pascutto]

On 02-05-18 16:30:

>> Can you build a 0.7 with cuDNN, and no other change, to test? This
>> would at least create a baseline for reference.
>
> No, cudnn version and 0.7 are completely different and there's no
> easy way to use cuDNN implementation in 0.7.

Well, just using cuDNN won't be too difficult, someone already did a
port of the original client:
https://github.com/Ttl/leela-zero/commit/11a254cf06a5577abcf6de97fdc2d72de90d0d62

But note that without aggressive batching, performance didn't seem
anything special compared to OpenCL:
https://github.com/gcp/leela-zero/issues/875#issuecomment-367328496

Recovering the batching implementation is more work as it predates
multi-GPU and Winograd support:
https://github.com/gcp/leela-zero/pull/526

Batching is going to be rather more important for chess, where there are
only 1/6th the pixels to compute per network compared to Go.

I'm bit curious what the NPS performance of lc0 looks like if batching
and parallel search / lookahead to gather networks for the batching is
disabled.

Also, disabling those may help understand where the strength loss at
fixed nodes is coming from.

--
GCP

[Alexander Lyashuk]

cuDNN version on batch size 1 is 25% slower than lczero variant.

Comparing how search in lc0 vs lczero differs in two versions while running with batchsize=1 is what I've been doing for a week at least.

Reasons of the difference that I managed to find are:

• "virtual loss bug" wasn't there in lc0 version
  
• "smart pruning" wasn't implemented
• openCl returns slightly different results of computation than tf/cuDNN for a same net.

After "virtual loss bug" and smart pruning was implemented (although I really hesitated to port a bug), it seems (from rough tests) that the skill with same number of nodes is now the same.

Also, if I disable smart pruning in lczero and "disable" virtual loss bug, they also seem to be roughly equal.

When running head to head in time controlled tournament, tf/cudnn version wins opencl without any questions (someone in discord posted W-L-D 44-1-55 for cudnn vs opencl in 5 sec/move game).

But when playing against other engine, it seems only a little bit stronger. Good network is more important than number of nodes in search.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/ef53718e-2f9c-ff7c-1926-aee0e0605c7a%40sjeng.org.

[Tadeusz R]

Hi Alexander,

Could you please attach required *.DLLs to run your lc0-win-cudnn version?

As solitarycandle wrote before, it is perfectly OK to include required *.dlls to your program.

On Tuesday, 1 May 2018 19:59:12 UTC+2, Alexander Lyashuk wrote:

Hi, 

To use it, you have to install CUDA and cuDNN from Nvidia site. (Downloading cuDNN requires registration on their site).

After that you can either build you version from sources (instructions https://github.com/glinscott/leela-chess/blob/next/lc0/README.md)

or use binaries that I've built:

https://crem.xyz/lc0/

Note: versions for Windows require CUDA v9.0 (not v9.1).

Version for Linux is for v9.1.

It's possible to build it for CUDA v9.1 for windows, I just didn't have time to set up the environment.

It may contain bugs. As of yesterday, it was weaker than "main" lczero on the same number of nodes, but today hopefully the reason for that is fixed, tests to confirm/reject that are currently ongoing.

One particular thing that may affect strength (in either way) is "virtual loss bug emulation", which can be enabled using --virtual-loss-bug command line flag or "Emulate virtual loss bug" UCI parameter.

On Tue, May 1, 2018 at 5:51 PM Albert Silver <silver...@gmail.com> wrote:

I noticed in the benchmarks a number of entries using lc0-cudnn. How can I use it or get it?

Thanks,

Albert

--

You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/6a80ff31-086f-48cf-84fb-6ed8b67378af%40googlegroups.com.

[Alexander Lyashuk]

No, I'm afraid not, I need to look into the license further. When I downloaded cuDNN, I don't remember seeing that claim in the license, although I read it.

And ./lc0 currently doesn't contain "valid copyright notices indicating NVIDIA’s ownership of the Licensed Software".

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/a5a2d1ab-91d9-44c0-acc0-39194083e39a%40googlegroups.com.

[Tadeusz R]

OK, I understand, thanks for reply.

I'm just not getting e-mail from nvidia for account confirmation... checked spam, re-requested email, nothing...

Already installed CUDA dev. , but no luck with cuDNN... 
Missing just one cudnn64_7.dll... :(

[Kilder Urrutia]

Excuse me, could someone give a guide on how to get CUDNN leela running? I already have CUDA and CUDNN installed and I am at a loss on what to do now. Trying to install the .exe to Arena or LucasChess just crashes. Many thanks in advance.

[Albert Silver]

How did you install or obtain CUDA without registering successfully with nVidia?

[Tadeusz R]

Does not require registration:

https://developer.nvidia.com/cuda-toolkit

[Tadeusz R]

Can I compare performance of lczero.exe vs lc0-win-cudnn.exe?

I run "go infinite" on both but I do not know when to stop with lc0-win-cudnn.exe

lczero.exe output at depth 26:

info depth 26 nodes 133253 nps 323 score cp 8 time 412718 

lc0-win-cudnn.exe output:

info depth 2 seldepth 26 time 9132   nodes 10700  score cp 7 hashfull 45 nps 1171

info depth 2 seldepth 27 time 11590  nodes 13863  score cp 7 hashfull 58 nps 1196

info depth 2 seldepth 28 time 16209  nodes 19841  score cp 7 hashfull 81 nps 1224

info depth 2 seldepth 29 time 21366  nodes 26493  score cp 7 hashfull 107 nps 1239

info depth 2 seldepth 30 time 22681  nodes 28258  score cp 7 hashfull 114 nps 1245

info depth 2 seldepth 31 time 34317  nodes 44542  score cp 7 hashfull 175 nps 1297

info depth 2 seldepth 32 time 131922 nodes 191133 score cp 8 hashfull 675 nps 1448

info depth 2 seldepth 33 time 131986 nodes 191186 score cp 8 hashfull 675 nps 1448

So is it nps 323 vs nps 1448?
That would be amazing!

[Albert Silver]

The benchmark is for 26 depth, but the real proof is in the pudding: how well it performs in games. 

Albert

[Jesse Jordache]

It doesn't work for me - for one thing I wasn't sure what to download in Alexander's link - I just downloaded the  most recent .7 file, and unzipped it where lczero is.

Just saying.  I'm going to wait for an official "all clear - just like the original, only faster!" before I put real effort into getting it working.

[Graham]

On Thursday, 3 May 2018 12:20:42 UTC+12, Jesse Jordache wrote:

It doesn't work for me - for one thing I wasn't sure what to download in Alexander's link - I just downloaded the  most recent .7 file, and unzipped it where lczero is.

Just saying.  I'm going to wait for an official "all clear - just like the original, only faster!" before I put real effort into getting it working.

What caught me out trying to use it in Arena was that it can't process compressed weights files (i.e. as they are when downloaded from the Networks list). Having downloaded the network required rename it to add a ".gz" file extension and then you can open it with your unzip program to expand it.

Then, provided you have cublas64_90.dll, cudart64_90.dll and cudnn64_7.dll in the program directory, it should run OK. You can even uninstall NVidia CUDA from your PC as long as you leave those DLLs there. (It's a real shame you need to register with NVidia, download 1.63GB and run all the installation stuff when it's only those three files you need).

Having said that, at the moment it doesn't look like the CudNN version is as good as the normal 0.7 version so you might be sensible waiting. But I'm not sure if there will ever be an "official" CUDA version as I think CUDA is proprietary to NVidia and only on their later GPUs.

[Nabil Danial]

You can actually extract those dlls from the installation file using 7zip without having the need of installing both CUDA and cuDNN themselves. You can download CUDA 9.1 and rename both cublas64_91.dll and cudart64_91.dll to cublas64_90.dll and cudart64_90.dll with no issues while gaining some performance boost (need some tests to back this one up). Those 3 dlls only use 176 MB of disk space after UPX compressed and can easily be distributed alongside lc0.  

[Nay Lin Tun]

Thank you very much guys. I can successfully installed my cuDNN version of Leela. Wow , the speed increased roughly four times!! 

P.S I am just amateur with very limited level of C++. It was too much challenging for me. :D 

[Nay Lin Tun]

Hi Guys, 

First of all, thanks very much for everyone who helped me in installing this. As I am just amateur with very limited level of C++, it was very difficult for me to understand how things work. I dont know how things magically worked but I am just sharing this to other people and also please correct me on unnecessary steps.  

Steps 1. 
1. Upgrade latest Nvidia driver 
2. Download and install visual studio 2015 (not sure this is a requitement for all machines), visual studio 2017 did not work for me 
3. Download CUDA 9.0 (9.1 did not work for me), 
Install CUDA in optional mode and uncheck visual studio integration on installation 
3. Download cudNN 9.0 
4. Unzip and paste the necessary files exactly according to nvidia instruction. 
https://ibb.co/eGFgon 
5. copy binaries from this link https://crem.xyz/lc0/ and unzip into ur installed directory. eg. C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\bin 

6. Download and install the latest version of leela into the same directory as in step 5. (they two must be together in the same directory as cuDNN version of leela needs some dll files , I think) 
7. Download weights and install in Arena. 

https://ibb.co/myORon 

The speed increased about 4 times, Wow, worth doing. 

[Nay Lin Tun]

Is there any tuning method for this version? 

[Graham]

I think Visual Studio is only required if your are going to recompile Leela yourself. I didn't install it. For my machine, running in Arena, I ended up with the following directory:

D:\Arena\Engines\LCzeroCudNN

containing:

202
202.gz
cublas64_90.dll
cudart64_90.dll
cudnn64_7.dll
lc0-win-cudnn.exe
lc0-win-cudnn.jpg

where 202 is the unzipped file from 202.gz and the JPG logo file is downloaded from TCEC (right click, Save Image As....). That's all it seems to need and everything else I uninstalled or deleted.

[Albert Silver]

Hi Alexander,

I have been testing the last May 1st version, comparing results and anything of note. I enabled Virtual Loss Bug this time. Three significant issues:

1) it is playing way too fast, and I have no idea how to control this. 

2) It loses on time plenty of times, but not because it is short of time. In my observations, this happens always when it is a draw and it is about to repeat for the third time. Instead of repeating, it freezes and lets all its time use up, which can mean minutes without making a move.

3) The startup time for the first move is amazingly slow. For my ten-minute games, this is not fatal, but it is annoying and strange. And yes, I configured it to know which weights.txt file to use.

For the moment, these are the most obvious issues. Thanks for the great work.

Albert

[Albert Silver]

An example, after 35 moves in a 10+0 game, it has only used 2m30s.

Albert

[frakty]

This may be a stupid question, but can't opencl be used similarly as cudnn with larger batch sizes? Or is the motivation behind cudnn, that at larger batchsizes it gives much better speed (if so, by how much?)

Also 4x increase in nodes count for official lc0 engine on any reasonable timecontrol gives at least +100 elo (on short tc even much more)

So node speed definitely is important for performance - except the nps must be reported in the official lc0 equivalent node count. I assume that lc0cudnn searches different tree / counts nodes differently. I wonder what speedup would be possible if the goal would be to run the exact search tree as the official lc0 (just using larger batch sizes).  Can you please elaborate on this?

[Jens Hartmann]

I can confirm these issues.

[Alexander Lyashuk]

Thanks for your suggestions, answers are inline.

On Thu, May 3, 2018 at 10:04 PM Albert Silver <silver...@gmail.com> wrote:

Hi Alexander,

I have been testing the last May 1st version, comparing results and anything of note. I enabled Virtual Loss Bug this time. Three significant issues:

1) it is playing way too fast, and I have no idea how to control this. 

Will fix it. With no time addition (e.g. after 40 moves) it budgets the remaining time equally for the next 50 moves.

With smart prunings, the engine often sees that there's no point of thinking all that time and returns earlier. That way it uses less time that it budgets.

 

2) It loses on time plenty of times, but not because it is short of time. In my observations, this happens always when it is a draw and it is about to repeat for the third time. Instead of repeating, it freezes and lets all its time use up, which can mean minutes without making a move.

There was a bug in smart pruning implementation, which caused deadlock. It should be fixed in the latest build from yesterday.

Could anyone please check and confirm that that doesn't happen anymore?

 

3) The startup time for the first move is amazingly slow. For my ten-minute games, this is not fatal, but it is annoying and strange. And yes, I configured it to know which weights.txt file to use.

Startup time consists of weight loading time and GPU initialization time. There's hardly anything that can be done with the latter, but I'll optimize the first.

 

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/b4ec60c1-92a3-4991-ac7c-bdfc6a1965f5%40googlegroups.com.

[Alexander Lyashuk]

OpenCL can indeed be used to implement batch computation. The problem is that OpenCL is much more low-level than cudnn/tensorflow, and it's not easy to implement that functionality in openCL without previous experience in openCL.

Therefore it's not implemented so far, and if anyone would like to volunteer to do that, that would be really appreciated.

Cudnn lc0 currently tries to replicate official lczero's tree search as close as possible, so the only known difference is difference in computation backends.

Computation backends themselves return slightly different results from NN evaluation due to pecularities in implementations. All backends (cudnn, opencl, cpu and opencl) have 1 to 5% differences in their output (the largest 5% difference noticed is CPU vs openCL).

So, slight differences in backend evaluations it the only known reason why trees may differ between implementations, otherwise they are equivalent. Also the visited nodes count is implemented the same way (again, unless I missed something, but so far no differences are known).

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/0dd28e5f-5b47-4eb7-85a6-151789b1f62b%40googlegroups.com.

[Albert Silver]

Answers below

On Friday, May 4, 2018 at 7:24:19 AM UTC-3, Alexander Lyashuk wrote:

Thanks for your suggestions, answers are inline.

On Thu, May 3, 2018 at 10:04 PM Albert Silver <silver...@gmail.com> wrote:

Hi Alexander,

I have been testing the last May 1st version, comparing results and anything of note. I enabled Virtual Loss Bug this time. Three significant issues:

1) it is playing way too fast, and I have no idea how to control this. 

Will fix it. With no time addition (e.g. after 40 moves) it budgets the remaining time equally for the next 50 moves.

With smart prunings, the engine often sees that there's no point of thinking all that time and returns earlier. That way it uses less time that it budgets.

More often 'thinks' there is no point. However, there is no way using up less than 1/4 of total time in the first 35 moves is proper time management.

 

2) It loses on time plenty of times, but not because it is short of time. In my observations, this happens always when it is a draw and it is about to repeat for the third time. Instead of repeating, it freezes and lets all its time use up, which can mean minutes without making a move.

There was a bug in smart pruning implementation, which caused deadlock. It should be fixed in the latest build from yesterday.

Could anyone please check and confirm that that doesn't happen anymore?

I'm running a match and will report after I have some results.

 

 

3) The startup time for the first move is amazingly slow. For my ten-minute games, this is not fatal, but it is annoying and strange. And yes, I configured it to know which weights.txt file to use.

Startup time consists of weight loading time and GPU initialization time. There's hardly anything that can be done with the latter, but I'll optimize the first.

Be that as it may, taking a full 20 seconds to initialize is excessive.

[Graham]

Hi Alexander,

The version from 20180505 seems to start much faster than the earlier one I was using (20180501). On a "go infinite" command that took about 7 seconds from saying "Found network file" to outputting the first linew and now it's down to one to two seconds. Thanks!

I know you're busy working on the processing of the program and this is a really trivial issue, but on your next release could you try to fix the output of depth? It's currently showing lines like:

info depth 1 seldepth 11 time 87 nodes 165 score cp 14 hashfull 0 nps 1896 pv e2e4 c7c5 c2c3 e7e6 d2d4 a7a6 d4c5 f8c5 f1d3 b8c6 g1f3
info depth 2 seldepth 11 time 95 nodes 195 score cp 14 hashfull 1 nps 2052 pv e2e4 c7c5 c2c3 e7e6 d2d4 a7a6 d4c5 f8c5 f1d3 b8c6 g1f3

so always shows the depth in Arena as either 1 or 2, when clearly it has searched much further than that.

Thanks again for your work

[Jesse Jordache]

No, it still doesn't work exactly for me.  I get an engine palette with everything blank except for a window that says "depth 2/2x" with the 2x increasing.

It's possible my gpu is too old - it's a GTX 960m.

It's at least well-behaved in its not-workingness.  The problem I had before was that I hadn't put the three dlls in the cb14 directory.

[Albert Silver]

As far as I know, the DLLs are supposed to go in the directory where the Leela executable is.

--
You received this message because you are subscribed to a topic in the Google Groups "LCZero" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lczero/JplcAeITux0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to lczero+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/f1d39aab-0c77-4cff-ba06-5df7dec4d439%40googlegroups.com.

[Jesse Jordache]

Same difference.  The old way of getting Leela to work with Chessbase was to put the Leela executable there, and I just left it that way, so they're all bundled together.

I learned something though - regular old lczero.exe can run off of a uncompressed weights file, so to try cudaNN, I just need to comment out one script line and uncomment out another, without having two weights.txt.s.

On Saturday, May 5, 2018 at 9:10:04 PM UTC-4, Albert Silver wrote:

As far as I know, the DLLs are supposed to go in the directory where the Leela executable is.

On Sat, May 5, 2018 at 9:33 PM, Jesse Jordache <youtwis...@gmail.com> wrote:

No, it still doesn't work exactly for me.  I get an engine palette with everything blank except for a window that says "depth 2/2x" with the 2x increasing.

It's possible my gpu is too old - it's a GTX 960m.

It's at least well-behaved in its not-workingness.  The problem I had before was that I hadn't put the three dlls in the cb14 directory.

On Saturday, May 5, 2018 at 6:30:24 PM UTC-4, Graham wrote:

Hi Alexander,

The version from 20180505 seems to start much faster than the earlier one I was using (20180501). On a "go infinite" command that took about 7 seconds from saying "Found network file" to outputting the first linew and now it's down to one to two seconds. Thanks!

I know you're busy working on the processing of the program and this is a really trivial issue, but on your next release could you try to fix the output of depth? It's currently showing lines like:

info depth 1 seldepth 11 time 87 nodes 165 score cp 14 hashfull 0 nps 1896 pv e2e4 c7c5 c2c3 e7e6 d2d4 a7a6 d4c5 f8c5 f1d3 b8c6 g1f3
info depth 2 seldepth 11 time 95 nodes 195 score cp 14 hashfull 1 nps 2052 pv e2e4 c7c5 c2c3 e7e6 d2d4 a7a6 d4c5 f8c5 f1d3 b8c6 g1f3

so always shows the depth in Arena as either 1 or 2, when clearly it has searched much further than that.

Thanks again for your work

--
You received this message because you are subscribed to a topic in the Google Groups "LCZero" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lczero/JplcAeITux0/unsubscribe.

To unsubscribe from this group and all its topics, send an email to lczero+un...@googlegroups.com.

[Alexander Lyashuk]

Depth parameter shows "full depth", i.e. "depth 2" means that there is at least one node at depth 3 which is not yet evaluated.

I think that the "full depth" is what engine is supposed to show, and I thought it would be interesting to see the difference between lczero and classical engines, as for lczero tree is very sparse.

Now it turned out that full depth is almost always 2 which is kind of boring, so maybe it worth showing some other kind of depth, maybe average depth. The depth to show must be easy to compute though, it doesn't seem worthwhile to slow down computation just to show more interesting number.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/e98f3297-ff12-406c-9d61-9f0d0405425c%40googlegroups.com.

[Graham]

Oh, I see.

You are currently showing a "seldepth" next to the depth and that seems to be counting up nicely. Perhaps use that number for the depth too? At least then it will look like it's searching progressively deeper (even if it's not!), and it won't require any additional work as you already have that number.

[Vivek Moroni]

Thank you for posting.  I do have some questions:

a) during the installation process, step 4, the last NVIDIA instruction (no.5) is to "include cudnn.lib in "your Visual Studio project".  Sorry - what project?  I installed Visual Studio 30 sec. ago only for the purpose of doing this setup.  I have no Visual Studio projects to open. 

b) step 5 from your instructions: when you say "copy the binaries from..." - I see a directory with executables; do I select the latest?  more than one?

Appreciate your help with these questions, thanks.

VM

[Jesse Jordache]

Albert, what is your execution script/line?  Do you use a wrapper, i.e, "inbetween" or something else?

I really want to get this right.

[Albert Silver]

I'm not sure I understand the question. You mean how do I use lc0-cudnn? I just install it as a UCI engine. I do have both CUDA and cuDNN installed beforehand. After, I also installed the latest Nvidia drivers, since CUDA comes with an older set.

On Mon, May 7, 2018 at 5:31 PM, Jesse Jordache <youtwis...@gmail.com> wrote:

Albert, what is your execution script/line?  Do you use a wrapper, i.e, "inbetween" or something else?

I really want to get this right.

--

You received this message because you are subscribed to a topic in the Google Groups "LCZero" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lczero/JplcAeITux0/unsubscribe.

To unsubscribe from this group and all its topics, send an email to lczero+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/a120a149-f84f-4238-9a18-768f9d014c80%40googlegroups.com.

[Jesse Jordache]

Thanks, that answers my question: I was wondering if you were doing something extra..

Blah.  I just get a climbing depth count, but no moves or evaluation.

On Monday, May 7, 2018 at 5:22:28 PM UTC-4, Albert Silver wrote:

I'm not sure I understand the question. You mean how do I use lc0-cudnn? I just install it as a UCI engine. I do have both CUDA and cuDNN installed beforehand. After, I also installed the latest Nvidia drivers, since CUDA comes with an older set.

On Mon, May 7, 2018 at 5:31 PM, Jesse Jordache <youtwis...@gmail.com> wrote:

Albert, what is your execution script/line?  Do you use a wrapper, i.e, "inbetween" or something else?

I really want to get this right.

--
You received this message because you are subscribed to a topic in the Google Groups "LCZero" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lczero/JplcAeITux0/unsubscribe.

To unsubscribe from this group and all its topics, send an email to lczero+un...@googlegroups.com.

[Graham]

The easiest way to check it's actually able to run is to open a command prompt, go to the LCzero directory and run the EXE file. It should show you the "LC0 built date..." message. Then type in "go infinite" (without the quotes!) and press enter. It should show you "Found network file...", "Creating backend..." and then the search results. Press Ctrl-C to stop it.

If it does that then your problem is the setup in Arena. To check that, go to the Engines > Manage > LCzero page and click on "Start this engine now". Then press F4 and it will show you the communications with the engine. That might show you what is wrong.

[Jesse Jordache]

Argh that's so frustrating.  Because it works, and much faster than doing the same thing with lczero.  And yet I can't get it to display moves in Chessbase 14 - I know it's thinking at least, because it shows me estimated node depth, but nothing else.  Still, thanks for the diagnostic tip.

lczero: "info depth 22 nodes 13006 nps 386 score cp 13 time 33704 pv e2e4 c7c5 g1f3 d7d6 f1b5 b8d7 e1g1 g8f6 f1e1 e7e6 c2c3 f8e7 d2d4 e8g8 e4e5 f6d5 c3c4 d5c7 e5d6 e7d6 c1g5"

lc0-win-cudNN-etc: "info depth 2 seldepth 22 time 1661 nodes 1648 score cp 16 hashfull 7 nps 992 pv e2e4 c7c5 g1f3 d7d6 d2d4 c5d4 f3d4 g8f6 b1c3 b8c6 c1g5 g7g6 g5f6 e7f6"

That's triple the node speed, just about.

[Jesse Jordache]

It works in Arena.  Which I'm totally not used to.  I want my databases :(.

[Stefan Pohl]

Yes, I installed all 3 needed .dll, the unpacked weights.txt-file and LC0 Cuda works in console-mode ("go infinite"). And in LittleBlitzerGUI. But in FritzGUI not (13 and 16): Sometimes, one move is played, but never a second move... and if I want to use LC0 Cuda in an Engine-Tournament, it crashes immediatly or doesnt start playing. And it is impossible to unload the engine and start another (Wasp, Komodo or something like that). I have to shut down the FritzGUI with the TaskManager...

[Alexander Lyashuk]

There is "Do debug logging into file" UCI option (or alternatively --debuglog=filename command line flag). Could you set it to some file name/path? Then it will generate a file with log, could you send it to me to take a look what's happening there?

Thanks!

--

You received this message because you are subscribed to the Google Groups "LCZero" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/8cc0bbc6-a810-4658-b9b7-9d7a487982b2%40googlegroups.com.

[Jesse Jordache]

The version I have (may 5) only has a textbox for a logfile, without an option of whether logging should be on.  Anyway, I can't get it to output a logfile, no matter if I enclose the filename in quotes, put in a path, use linux style directory divider "/", use a relative or absolute path, in any combination.
 
Chessbase/Fritz doesn't accept UCI arguments, so I can't try it the other way.  At any rate it works fine from the command line; also from Arena; just not from anything chessbase related.  Oh, the directory I was trying to write to has "Everyone" write privileges, so you can rule that out if it crosses your mind.

[Albert Silver]

Just type in any name.txt in the blank space. It will save it to the folder where Leela is.

[Jesse Jordache]

It doesn't.

[evalon32]

On Sunday, May 6, 2018 at 2:59:15 AM UTC-4, Alexander Lyashuk wrote:

Depth parameter shows "full depth", i.e. "depth 2" means that there is at least one node at depth 3 which is not yet evaluated.

I think that the "full depth" is what engine is supposed to show, and I thought it would be interesting to see the difference between lczero and classical engines, as for lczero tree is very sparse.

Come to think of it, this "full depth" concept seems too strict even for classical engines. After all, alpha-beta search doesn't evaluate every node at depth k, either; only naive minimax would.

[Stefan Pohl]

The Cuda-Leela is very buggy, sorry. In LittleBlitzerGUI it runs (in Fritz not), but lost 9 games on time (of 70 games, with 6'+2.5'' timecontrol) and lost 2 games by making illegal moves! (I used the version of May,8)
Here the log of the illegal moves:


--Start Board-----------------------------------------------------------
bR =  0x0000000000000000 wR =  0x0000000000000000 Castling =     0x30
bB =  0x4000000000000000 wB =  0x0000000000000000 EPSquare =     -1
bN =  0x0000000000000000 wN =  0x0000000000000000 White Mat =     0
bQ =  0x0000000000000000 wQ =  0x0000000000000000 Black Mat =     0
bK =  0x0000000008000000 wK =  0x0000000000000200 50 Moves =     100
bP =  0x0000000000000100 wP =  0x0000000000000000 Side To Move = W
bA =  0x4000000008000100 wA =  0x0000000000000200 Half Move =     0
FEN = 1b6/8/8/8/4k3/8/6Kp/8 w - - 100 0
   a b c d e f g h
   ---------------
8|   b             |8  63 62 61 60 59 58 57 56    a8 b8 c8 d8 e8 f8 g8 h8
7|                 |7  55 54 53 52 51 50 49 48    a7 b7 c7 d7 e7 f7 g7 h7
6|                 |6  47 46 45 44 43 42 41 40    a6 b6 c6 d6 e6 f6 g6 h6
5|                 |5  39 38 37 36 35 34 33 32    a5 b5 c5 d5 e5 f5 g5 h5
4|         k       |4  31 30 29 28 27 26 25 24    a4 b4 c4 d4 e4 f4 g4 h4
3|                 |3  23 22 21 20 19 18 17 16    a3 b3 c3 d3 e3 f3 g3 h3
2|             K p |2  15 14 13 12 11 10 09 08    a2 b2 c2 d2 e2 f2 g2 h2
1|                 |1  07 06 05 04 03 02 01 00    a1 b1 c1 d1 e1 f1 g1 h1
   ---------------
   a b c d e f g h
--End Board-------------------------------------------------------------
Starting Position: rnbqkbnr/pppp1ppp/8/4p3/2P5/8/PP1PPPPP/RNBQKBNR w KQkq e6 0 0
Movelist:  g2g3 g8f6 f1g2 c7c6 g1f3 d7d6 b1c3 f8e7 e2e4 c6c5 a1b1 b8c6 a2a3 e8g8 d2d3 c6d4 b2b4 b7b6 h2h3 a7a6 e1g1 c8d7 f3d4 c5d4 c3e2 b6b5 c4b5 a6b5 f2f4 h7h6 c1b2 a8c8 d1d2 d8b6 b1c1 d7e6 g1h2 f6d7 e2g1 b6b7 g1f3 e7f6 f4f5 e6b3 d2e2 d7b6 h3h4 b6a4 b2a1 c8c1 f1c1 f8c8 e2e1 c8c6 g3g4 f6d8 g4g5 c6c1 e1c1 b7c7 c1c7 d8c7 a1d4 e5d4 f3d4 b3d1 d4b5 c7b8 h2g1 a4b2 d3d4 d1a4 g2f1 a4c2 b5c3 b8a7 c3b5 a7b6 g1f2 c2e4 f2e3 e4f5 g5h6 g7h6 b5d6 f5e6 e3e4 g8f8 d4d5 e6d7 f1b5 f8e7 b5d7 e7d6 d7e8 d6e7 e8b5 b6f2 h4h5 f2e1 e4f5 e1c3 f5e4 c3e1 e4f5 e1c3 f5e4 c3d2 e4e5 f7f6 e5f5 b2d1 d5d6 e7d6 f5f6 d2g5 f6g6 d1e3 b5e8 e3c2 a3a4 c2b4 g6f5 b4d5 a4a5 d6c5 e8d7 d5f6 d7c8 f6h5 f5g4 h5f6 g4f5 c5b5 a5a6 b5b6 f5g6 f6d5 g6h5 d5c7 c8f5 c7a6 f5e4 a6c5 e4a8 c5d7 h5g4 d7f6 g4h3 b6c5 h3g3 c5d4 a8b7 d4e3 b7f3 g5h4 g3g2 e3f4 f3b7 f6h7 b7c8 h7g5 c8d7 h4g3 d7e8 g5e4 g2h3 e4f6 e8g6 f4g5 g6f7 g3c7 f7e6 f6h5 e6f7 h5f4 h3g3 f4d3 g3h3 d3f4 h3g3 c7b8 g3h2 g5g4 f7e8 f4d3 h2g2 d3f4 g2h2 g4f3 e8f7 f3f2 h2h1 f2g3 f7e8 f4d3 e8h5 d3f2 h1g1 b8a7 h5g6 f2g4 g1h1 g4f2 h1g1 a7b6 g1f1 f2g4 g6h5 g4f6 h5g6 g3f3 f1e1 b6e3 e1f1 f6d5 g6h5 f3g3 h5f7 d5f4 f1e1 g3f3 e1d1 f4d3 d1c2 d3e5 f7h5 f3g3 h5e8 g3f4 c2b3 e3d2 b3c2 d2b4 c2b2 f4g5 b2b3 b4d6 b3c3 e5g6 c3d2 h6h5 d2e3 d6c5 e3f3 h5h4 e8g6 g5g6 f3g4 c5e7 g4h3 g6h5 h3h2 h5g4 h2g2 e7b4 g2h1 g4h3 h1g1 b4c3 g1h1 c3f6 h1g1 f6d4 g1h1 d4e5 h1g1 e5g3 g1h1 h3g4 h1g2 g3d6 g2h1 g4h3 h1g1 d6b4 g1h1 h3g4 h1g2 b4a5 g2h2 a5c7 h2h1 g4f3 h1g1 c7a5 g1h2 f3g4 h2g2 a5e1 g2h1 g4h3 h1g1 e1d2 g1h1 d2c3 h1g1 c3e5 g1h1 e5b2 h1g1 b2f6 g1h1 f6e7 h1g1 h3g3 g1h1 e7b4 h1g1 b4c3 g1h1 c3b2 h1g1 g3h3 g1h1 b2a3 h1g1 a3d6 g1h1 h3g4 h1g2 g4h5 g2h3 d6g3 h3g2 h5g4 g2h1 g4h3 h1g1 g3f4 g1h1 f4d2 h1g1 d2a5 g1h1 h3g4 h1h2 a5c7 h2h1 c7f4 h1g2 f4g5 g2h2 h4h3 h2h1 g4g3 h1g1 g5d2 g1h1 d2c3 h1g1 c3a5 g1h1 a5c7 h1g1 g3f3 g1h1 f3e3 h1g1 e3d4 g1h1 d4e3 h1g1 e3e4 g1h1 e4f3 h1g1 c7d6 g1h1 f3g3 h1g1 g3g4 g1h1 g4f3 h1g1 f3e2 g1h1 e2d3 h1g1 d3c4 g1h1 c4c5 h1g1 c5d4 g1h1 d6b8 h1g1 d4e3 g1h1 e3d3 h1g1 d3c4 g1h1 b8d6 h1g1 c4d3 g1h1 d3d4 h1g1 d4c5 g1h1 c5b4 h1g1 b4c5 g1h1 c5d4 h1g1 d4e3 g1h1 d6b8 h1g1 e3f3 g1h1 f3g3 h1g1 g3g4 g1h1 b8g3 h1g1 g4f4 g1h1 f4e3 h1g1 e3f3 g1h1 g3f4 h1g1 f3g3 g1h1 f4e5 h1g1 e5d6 g1h1 g3g4 h1g1 d6b4 g1h1 b4a5 h1h2 a5c7 h2g1 h3h2 g1h1 g4f5 h1g2 f5e4 g2h1 e4d5 h1g2 d5c4 g2h1 c4b4 h1g2 b4c5 g2h1 c7d6 h1g2 c5b6 g2h1 b6b5 h1g2 b5c4 g2h1 c4b4 h1g2 b4c5 g2h1 c5b6 h1g2 b6c5 g2h1 c5c4 h1g2 c4b4 g2h1 b4c3 h1g2 c3d4 g2h1 d4c5 h1g2 c5d4 g2h1 d4c4 h1g2 d6b8 g2h1 c4b4 h1g2 b8f4 g2h1 b4c5 h1g2 c5d4 g2h1 d4c3 h1g2 f4c7 g2h1 c3b2 h1g2 b2c3 g2h1 c3b3 h1g2 b3b2 g2h1 b2c3 h1g2 c3b4 g2h1 b4a5 h1g2 c7f4 g2h1 a5b4 h1g2 b4c5 g2h1 f4e5 h1g2 c5d6 g2h1 d6c5 h1g2 c5d4 g2h1 d4c3 h1g2 c3d4 g2h1 d4d3 h1g2 e5c7 g2h1 c7b8 h1g2 d3e4 g2h1 e4d3 h1g2 d3e4
Engine Output: bestmove a1a1
Test engine output with:
position fen rnbqkbnr/pppp1ppp/8/4p3/2P5/8/PP1PPPPP/RNBQKBNR w KQkq e6 0 0 moves g2g3 g8f6 f1g2 c7c6 g1f3 d7d6 b1c3 f8e7 e2e4 c6c5 a1b1 b8c6 a2a3 e8g8 d2d3 c6d4 b2b4 b7b6 h2h3 a7a6 e1g1 c8d7 f3d4 c5d4 c3e2 b6b5 c4b5 a6b5 f2f4 h7h6 c1b2 a8c8 d1d2 d8b6 b1c1 d7e6 g1h2 f6d7 e2g1 b6b7 g1f3 e7f6 f4f5 e6b3 d2e2 d7b6 h3h4 b6a4 b2a1 c8c1 f1c1 f8c8 e2e1 c8c6 g3g4 f6d8 g4g5 c6c1 e1c1 b7c7 c1c7 d8c7 a1d4 e5d4 f3d4 b3d1 d4b5 c7b8 h2g1 a4b2 d3d4 d1a4 g2f1 a4c2 b5c3 b8a7 c3b5 a7b6 g1f2 c2e4 f2e3 e4f5 g5h6 g7h6 b5d6 f5e6 e3e4 g8f8 d4d5 e6d7 f1b5 f8e7 b5d7 e7d6 d7e8 d6e7 e8b5 b6f2 h4h5 f2e1 e4f5 e1c3 f5e4 c3e1 e4f5 e1c3 f5e4 c3d2 e4e5 f7f6 e5f5 b2d1 d5d6 e7d6 f5f6 d2g5 f6g6 d1e3 b5e8 e3c2 a3a4 c2b4 g6f5 b4d5 a4a5 d6c5 e8d7 d5f6 d7c8 f6h5 f5g4 h5f6 g4f5 c5b5 a5a6 b5b6 f5g6 f6d5 g6h5 d5c7 c8f5 c7a6 f5e4 a6c5 e4a8 c5d7 h5g4 d7f6 g4h3 b6c5 h3g3 c5d4 a8b7 d4e3 b7f3 g5h4 g3g2 e3f4 f3b7 f6h7 b7c8 h7g5 c8d7 h4g3 d7e8 g5e4 g2h3 e4f6 e8g6 f4g5 g6f7 g3c7 f7e6 f6h5 e6f7 h5f4 h3g3 f4d3 g3h3 d3f4 h3g3 c7b8 g3h2 g5g4 f7e8 f4d3 h2g2 d3f4 g2h2 g4f3 e8f7 f3f2 h2h1 f2g3 f7e8 f4d3 e8h5 d3f2 h1g1 b8a7 h5g6 f2g4 g1h1 g4f2 h1g1 a7b6 g1f1 f2g4 g6h5 g4f6 h5g6 g3f3 f1e1 b6e3 e1f1 f6d5 g6h5 f3g3 h5f7 d5f4 f1e1 g3f3 e1d1 f4d3 d1c2 d3e5 f7h5 f3g3 h5e8 g3f4 c2b3 e3d2 b3c2 d2b4 c2b2 f4g5 b2b3 b4d6 b3c3 e5g6 c3d2 h6h5 d2e3 d6c5 e3f3 h5h4 e8g6 g5g6 f3g4 c5e7 g4h3 g6h5 h3h2 h5g4 h2g2 e7b4 g2h1 g4h3 h1g1 b4c3 g1h1 c3f6 h1g1 f6d4 g1h1 d4e5 h1g1 e5g3 g1h1 h3g4 h1g2 g3d6 g2h1 g4h3 h1g1 d6b4 g1h1 h3g4 h1g2 b4a5 g2h2 a5c7 h2h1 g4f3 h1g1 c7a5 g1h2 f3g4 h2g2 a5e1 g2h1 g4h3 h1g1 e1d2 g1h1 d2c3 h1g1 c3e5 g1h1 e5b2 h1g1 b2f6 g1h1 f6e7 h1g1 h3g3 g1h1 e7b4 h1g1 b4c3 g1h1 c3b2 h1g1 g3h3 g1h1 b2a3 h1g1 a3d6 g1h1 h3g4 h1g2 g4h5 g2h3 d6g3 h3g2 h5g4 g2h1 g4h3 h1g1 g3f4 g1h1 f4d2 h1g1 d2a5 g1h1 h3g4 h1h2 a5c7 h2h1 c7f4 h1g2 f4g5 g2h2 h4h3 h2h1 g4g3 h1g1 g5d2 g1h1 d2c3 h1g1 c3a5 g1h1 a5c7 h1g1 g3f3 g1h1 f3e3 h1g1 e3d4 g1h1 d4e3 h1g1 e3e4 g1h1 e4f3 h1g1 c7d6 g1h1 f3g3 h1g1 g3g4 g1h1 g4f3 h1g1 f3e2 g1h1 e2d3 h1g1 d3c4 g1h1 c4c5 h1g1 c5d4 g1h1 d6b8 h1g1 d4e3 g1h1 e3d3 h1g1 d3c4 g1h1 b8d6 h1g1 c4d3 g1h1 d3d4 h1g1 d4c5 g1h1 c5b4 h1g1 b4c5 g1h1 c5d4 h1g1 d4e3 g1h1 d6b8 h1g1 e3f3 g1h1 f3g3 h1g1 g3g4 g1h1 b8g3 h1g1 g4f4 g1h1 f4e3 h1g1 e3f3 g1h1 g3f4 h1g1 f3g3 g1h1 f4e5 h1g1 e5d6 g1h1 g3g4 h1g1 d6b4 g1h1 b4a5 h1h2 a5c7 h2g1 h3h2 g1h1 g4f5 h1g2 f5e4 g2h1 e4d5 h1g2 d5c4 g2h1 c4b4 h1g2 b4c5 g2h1 c7d6 h1g2 c5b6 g2h1 b6b5 h1g2 b5c4 g2h1 c4b4 h1g2 b4c5 g2h1 c5b6 h1g2 b6c5 g2h1 c5c4 h1g2 c4b4 g2h1 b4c3 h1g2 c3d4 g2h1 d4c5 h1g2 c5d4 g2h1 d4c4 h1g2 d6b8 g2h1 c4b4 h1g2 b8f4 g2h1 b4c5 h1g2 c5d4 g2h1 d4c3 h1g2 f4c7 g2h1 c3b2 h1g2 b2c3 g2h1 c3b3 h1g2 b3b2 g2h1 b2c3 h1g2 c3b4 g2h1 b4a5 h1g2 c7f4 g2h1 a5b4 h1g2 b4c5 g2h1 f4e5 h1g2 c5d6 g2h1 d6c5 h1g2 c5d4 g2h1 d4c3 h1g2 c3d4 g2h1 d4d3 h1g2 e5c7 g2h1 c7b8 h1g2 d3e4 g2h1 e4d3 h1g2 d3e4



--Start Board-----------------------------------------------------------
bR =  0x0000000000000000 wR =  0x0000000000000000 Castling =     0x30
bB =  0x2000000000000000 wB =  0x0040000000000000 EPSquare =     -1
bN =  0x0000000000000000 wN =  0x0000000000000000 White Mat =     0
bQ =  0x0000000000000000 wQ =  0x0000000000000000 Black Mat =     0
bK =  0x0000400000000000 wK =  0x0000000000800000 50 Moves =     0
bP =  0x0000000000000000 wP =  0x0000000000000000 Side To Move = W
bA =  0x2000400000000000 wA =  0x0040000000800000 Half Move =     0
FEN = 2b5/1B6/1k6/8/8/K7/8/8 w - - 0 0
   a b c d e f g h
   ---------------
8|     b           |8  63 62 61 60 59 58 57 56    a8 b8 c8 d8 e8 f8 g8 h8
7|   B             |7  55 54 53 52 51 50 49 48    a7 b7 c7 d7 e7 f7 g7 h7
6|   k             |6  47 46 45 44 43 42 41 40    a6 b6 c6 d6 e6 f6 g6 h6
5|                 |5  39 38 37 36 35 34 33 32    a5 b5 c5 d5 e5 f5 g5 h5
4|                 |4  31 30 29 28 27 26 25 24    a4 b4 c4 d4 e4 f4 g4 h4
3| K               |3  23 22 21 20 19 18 17 16    a3 b3 c3 d3 e3 f3 g3 h3
2|                 |2  15 14 13 12 11 10 09 08    a2 b2 c2 d2 e2 f2 g2 h2
1|                 |1  07 06 05 04 03 02 01 00    a1 b1 c1 d1 e1 f1 g1 h1
   ---------------
   a b c d e f g h
--End Board-------------------------------------------------------------
Starting Position: rnbqkbnr/pppp1ppp/4p3/8/2P5/8/PP1PPPPP/RNBQKBNR w KQkq - 0 0
Movelist:  g1f3 d7d5 d2d4 g8f6 b1c3 d5c4 e2e4 f8b4 c1g5 c7c5 f1c4 c5d4 f3d4 d8c7 d1b3 c7a5 e1g1 b4c3 g5f6 g7f6 b2c3 e8g8 a1d1 g8h8 c4e2 a7a6 c3c4 b8d7 b3h3 f8g8 e2h5 g8g7 d4b3 a5e5 h3h4 e5g5 h4g5 f6g5 c4c5 d7e5 d1d8 g7g8 f1d1 a6a5 d8g8 h8g8 d1d8 g8g7 b3d4 a5a4 h5e2 e5c6 d4c6 b7c6 e4e5 a4a3 f2f3 c8b7 d8a8 b7a8 g2g3 f7f6 f3f4 g5f4 g3f4 f6e5 f4e5 g7g6 e2d3 g6g5 d3h7 g5f4 g1f2 f4e5 h2h4 e5f4 h4h5 a8b7 h7g8 f4g5 g8e6 g5h5 f2e3 h5g5 e3d4 g5f6 e6g4 b7a6 g4f3 a6b7 f3h1 f6e6 h1e4 e6f7 d4e5 f7e7 e5d4 e7f7 d4c4 f7e6 c4b4 e6e5 e4g6 e5d4 g6e8 b7c8 e8c6 c8e6 c6f3 e6a2 c5c6 a2e6 b4a3 d4c5 f3h1 e6c4 h1e4 c4b5 c6c7 b5d7 e4b7 c5b6 c7c8q d7c8
Engine Output: bestmove a1a1
Test engine output with:
position fen rnbqkbnr/pppp1ppp/4p3/8/2P5/8/PP1PPPPP/RNBQKBNR w KQkq - 0 0 moves g1f3 d7d5 d2d4 g8f6 b1c3 d5c4 e2e4 f8b4 c1g5 c7c5 f1c4 c5d4 f3d4 d8c7 d1b3 c7a5 e1g1 b4c3 g5f6 g7f6 b2c3 e8g8 a1d1 g8h8 c4e2 a7a6 c3c4 b8d7 b3h3 f8g8 e2h5 g8g7 d4b3 a5e5 h3h4 e5g5 h4g5 f6g5 c4c5 d7e5 d1d8 g7g8 f1d1 a6a5 d8g8 h8g8 d1d8 g8g7 b3d4 a5a4 h5e2 e5c6 d4c6 b7c6 e4e5 a4a3 f2f3 c8b7 d8a8 b7a8 g2g3 f7f6 f3f4 g5f4 g3f4 f6e5 f4e5 g7g6 e2d3 g6g5 d3h7 g5f4 g1f2 f4e5 h2h4 e5f4 h4h5 a8b7 h7g8 f4g5 g8e6 g5h5 f2e3 h5g5 e3d4 g5f6 e6g4 b7a6 g4f3 a6b7 f3h1 f6e6 h1e4 e6f7 d4e5 f7e7 e5d4 e7f7 d4c4 f7e6 c4b4 e6e5 e4g6 e5d4 g6e8 b7c8 e8c6 c8e6 c6f3 e6a2 c5c6 a2e6 b4a3 d4c5 f3h1 e6c4 h1e4 c4b5 c6c7 b5d7 e4b7 c5b6 c7c8q d7c8

[Alexander Lyashuk]

That's fixed in the latest version.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/1971d587-f216-46e3-9e18-aa1b84d7087b%40googlegroups.com.

[Albert Silver]

Yes, the May 8 version was very buggy, more than you can imagine, due to MS in fact, but is solved in the last version. 

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+unsubscribe@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/1971d587-f216-46e3-9e18-aa1b84d7087b%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

You received this message because you are subscribed to a topic in the Google Groups "LCZero" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lczero/JplcAeITux0/unsubscribe.

To unsubscribe from this group and all its topics, send an email to lczero+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/CACGjms%2Bnm8nv-M66Zp6mQ6%2BNu9qfuKDDMfJkYF%2ByEuAS%3DMrRow%40mail.gmail.com.

[Enrico Caruso]

There are two releases from the 9/5. One .00 and one .01. Which is better? :)

[Albert Silver]

The more recent one.

On Thu, May 10, 2018 at 12:33 PM, Enrico Caruso <rosapo...@gmail.com> wrote:

There are two releases from the 9/5. One .00 and one .01. Which is better? :)

--

You received this message because you are subscribed to a topic in the Google Groups "LCZero" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lczero/JplcAeITux0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to lczero+unsubscribe@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/15bc2e02-1c71-4c1f-974c-b70b5fc97357%40googlegroups.com.

[Stefan Pohl]

Even the latest version from today does not run in FritzGUI. Plays one or two moves, then all is over. With or without polyglot-Adapter. With or without uci as argument.
Not running in analyze-mode or in an engine-tournament.

[svoi s]

How  long have you not updated your Fritz?
четверг, 10 мая 2018 г., 19:18:36 UTC+3 пользователь Stefan Pohl написал:

[Stefan Pohl]

Correction. The latest version seems to work. But the GUI sent a ponder-command, which doesnt work. I will play some games and report here.
Sad, that there is no output, when Cuda-Leela is thinking. Only, after it moves, there is Output.

[Stefan Pohl]

And Leela-Cuda only works in a Fritz engine-tournament, when I use the polyglot-adapter (without uci-argument). Otherwise it crashes immediately.

[123]

Stefan Pohl:

And Leela-Cuda only works in a Fritz engine-tournament, when I use the polyglot-adapter (without uci-argument). Otherwise it crashes immediately.

 Use Aquarium 2018 gui, it works perfectly with ...v0.7, v0.8, v0.9...

[svoi s]

re-download, please
fixed
четверг, 10 мая 2018 г., 20:41:35 UTC+3 пользователь 123 написал:

[Stefan Pohl]

NOT fixed. New version crashes immediately in a Fritz Engine-tournament, without using polyglot-adapter.

[Stefan Pohl]

And still no information displayed, while it is thinking. Only after moving. In FritzGUI.

[mjlitola]

For me it installed on Fritz 14 ok. First lc0 was doing timeouts, but after I rebooted the system, it started working. It will be interesting to do deep analysis with latest SF build, when lc0 gets +3300-400 ELO range.

[Jesse Jordache]

Oh I should try that.  I've just waited for the moves to show up; I don't think I've moved while it was thinking.

[kirill57]

I am not sure is this a bug or a feature, but every time I increase my Cpuct my node counts also increases,

[Jesse Jordache]

Feature I think.  Less decision making, pruning, etc.

The thing is, a lot of the nodes that make up that number are useless.  Like, hard core useless - the results of moves that would never in a million years be a good idea.

[kirill57]

I wander is anywhere description what a recent changes in the lco updates?

[Alexander Lyashuk]

As far as I remember, it's only one bug fixed (analyze mode in arena and other GUIs work better now).

And today also build for CUDA 9.2 appeared.

On Sat, May 19, 2018 at 11:07 PM kirill57 <cyri...@gmail.com> wrote:

I wander is anywhere description what a recent changes in the lco updates?

--

You received this message because you are subscribed to the Google Groups "LCZero" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/88a537a4-7fed-4303-baf8-b03c445a5cdc%40googlegroups.com.

[Kostas Oreopoulos]

Is there any reference to what each UCI option does? (temperature, for example, Cpuct MCTS option)

[Alexander Lyashuk]

I started to write this https://github.com/glinscott/leela-chess/blob/next/lc0/FLAGS.md yesterday.

On Mon, May 28, 2018 at 3:33 PM Kostas Oreopoulos <kostas.o...@gmail.com> wrote:

Is there any reference to what each UCI option does? (temperature, for example, Cpuct MCTS option)

--

You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/36aa206a-934f-481e-bae7-334cf616317c%40googlegroups.com.

[Kirill Oseledets]

I am also curious what "smart pruning" do?

On Mon, May 28, 2018 at 8:39 AM, Alexander Lyashuk <moos...@gmail.com> wrote:

I started to write this https://github.com/glinscott/leela-chess/blob/next/lc0/FLAGS.md yesterday.

On Mon, May 28, 2018 at 3:33 PM Kostas Oreopoulos <kostas.o...@gmail.com> wrote:

Is there any reference to what each UCI option does? (temperature, for example, Cpuct MCTS option)

--
You received this message because you are subscribed to the Google Groups "LCZero" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+unsubscribe@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/36aa206a-934f-481e-bae7-334cf616317c%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

You received this message because you are subscribed to a topic in the Google Groups "LCZero" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lczero/JplcAeITux0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to lczero+unsubscribe@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/CACGjms%2Bk0n0qnHeoKke4QRP-MHwXGQcbAQLKcu9QbyrnanBLfQ%40mail.gmail.com.