Upload network to server?

[Nyquixt]

So, I tried to set up the LZ server locally. I managed to get everything done (I hope) except for the uploading network part. What I did was:

• gzip my network txt file (txt -> gz)
• change its name to best-network.gz
• use curl -F '[email protected]' -F 'training_count=0' -F 'key=auth_key' http://localhost:8080/submit-network
• copy best-network.gz into leela-zero-server/network/ directory
• change const Qstring server_url = "https://zero.sjeng.org/" to "http://localhost:8080/"

...and still got this error about mismatching in hash.
BTW, is a code of 52 I got from curl bad?
Thanks in advance!

[roy7]

Sounds similar to the problem in #217 but @Superkidd didn't actually say how he fixed it in the end.

[roy7]

BTW you don't need to rename to best-network.gz and it doesn't matter if you do. The server names each uploaded network with the hash, and stores it under networks/. For a net to become 'best' and promoted by server, you need to have it do a match and then the match successfully pass/promote the challenger. For the very first setup of the server you need to manually copy a network into networks/ and name it best-network.gz.

[Nyquixt]

@roy7 I did what you told me too. Yet, the calculated hash even with a different network uploaded is still 3ba11708... Also, I followed a more detailed instruction on how to set up the server on https://github.com/sai-dev/sai-server and still got the error. Is there anything else to change beside the QString url_server in autogtp? Plus, attached is the server log when I curled the network.

[roy7]

Would you look at the download network with cat or a text editor and see if looks like a valid net? What's its file size? I'm guessing it isn't actually downloading a network so you don't see the hash changing. I'm suspecting it's downloading an error message or a null file or something?

The server error about trying to access 'count' of a null variable, is because you have no games at all. Once a game exists I guess that goes away. Try changing line 611 from "set.training_count = totalgames.count;" to "set.training_count = (totalgames ? totalgames.count || 0);"

That should avoid crashing on upload of network with zero games played, and thus get the game into the database...

[roy7]

I've updated master branch to have everything in next branch, which could possibly have an effect. Some old next changes were related to gzip_hash I think.

Yes, all the code that adds "hash_gzip_hash" field in the selfplay packet from server (missing in your log above) was only in next branch not master branch. I've merged all of next into master now.

[roy7]

@gcp When autogtp downloads a network, and it uses a URL like networks/hash.gz, we have a setting in the web server that intercepts the networks/ path don't we? Because I'm not seeing a handler anywhere in the server code to serve binary files requested under the networks/ end point. Doh?

[roy7]

This I think would be the key?

sai-dev@26c2ab8

[roy7]

SAI team fixed this bug last year, who knew? sai-dev@02f7f10

[gcp]

When autogtp downloads a network, and it uses a URL like networks/hash.gz, we have a setting in the web server that intercepts the networks/ path don't we?

Yes that's handled directly by the webserver without involving node.js. Improves the memory usage.

[roy7]

Thanks. I never documented that so it's why some people have failed to get the server working in the past, doh!

[gcp]

Currently there's even extra tricks to "fall back" to leela.online-go.com for older networks.

nginx config:

    location /networks {
        fancyindex on;
        fancyindex_default_sort date_desc;
        alias /home/jroy/code/leela-zero-server/network;
        try_files $uri @storage;
    }

    location @storage {
        expires 30s;
        return 301 https://leela.online-go.com$request_uri;
    }

[roy7]

Thanks, good to have that documented someplace. :) Sadly it's taken over a year for me to fix this by default for new people trying to run local servers.

[roy7]

Ok @Nyquixt check out #227

[Nyquixt]

@roy7 nice!

[Nyquixt]

BTW, why is 11 is on the second line of /best-nerwork-hash response?

[roy7]

Might be an old protocol version number or something? I'm not actually sure any more lol.