LC0 0.31veröffentlicht

By Lothar Jung Date 2024-03-26 09:14 Upvotes 1

After a long gestation, the first release candidate Lc0 v0.31.0-rc1 is now available: <https://github.com/LeelaChessZero/lc0/releases/tag/v0.31.0-rc1>
In this release:
* The blas, cuda, eigen, metal and onnx backends now have support for multihead network architecture and can run BT3/BT4 nets.
* Updated the internal Elo model to better align with regular Elo for human players.
* There is a new XLA backend that uses OpenXLA compiler to produce code to execute the neural network. See <https://github.com/LeelaChessZero/lc0/wiki/XLA-backend> for details. Related are new leela2onnx options to output the HLO format that XLA understands.
* There is a vastly simplified lc0 interface available by renaming the executable to `lc0simple`.
* The backends can now suggest a minibatch size to the search, this is enabled by `--minibatch-size=0` (the new default).
* If the cudnn backend detected an unsupported network architecture it will switch to the cuda backend.
* Two new selfplay options enable value and policy tournaments. A policy tournament is using a single node policy to select the move to play, while a value tournament searches all possible moves at depth 1 to select the one with the best q.
* While it is easy to get a single node policy evaluation (`go nodes 1` using uci), there was no simple way to get the effect of a value only evaluation, so the `--value-only` option was added.
* Button uci options were implemented and a button to clear the tree was added (as hidden option).
* Support for the uci `go mate` option was added.
* The rescorer can now be built from the lc0 code base instead of a separate branch.

By Max Siegfried Date 2024-03-26 10:08 Edited 2024-03-26 10:12

Es enthält Multihead arch metal und läuft somit auf Apple Hardware.
https://github.com/LeelaChessZero/lc0/pull/1916

By Max Siegfried Date 2024-03-26 10:11

v0.31.0-rc1 Pre-release
@borg323 borg323 released this 10 hours ago
v0.31.0-rc1
05172b6

In this version:

The blas, cuda, eigen, metal and onnx backends now have support for multihead network architecture and can run BT3/BT4 nets.
Updated the internal Elo model to better align with regular Elo for human players.
There is a new XLA backend that uses OpenXLA compiler to produce code to execute the neural network. See https://github.com/LeelaChessZero/lc0/wiki/XLA-backend for details. Related are new leela2onnx options to output the HLO format that XLA understands.
There is a vastly simplified lc0 interface available by renaming the executable to lc0simple.
The backends can now suggest a minibatch size to the search, this is enabled by --minibatch-size=0 (the new default).
If the cudnn backend detected an unsupported network architecture it will switch to the cuda backend.
Two new selfplay options enable value and policy tournaments. A policy tournament is using a single node policy to select the move to play, while a value tournament searches all possible moves at depth 1 to select the one with the best q.
While it is easy to get a single node policy evaluation (go nodes 1 using uci), there was no simple way to get the effect of a value only evaluation, so the --value-only option was added.
Button uci options were implemented and a button to clear the tree was added (as hidden option).
Support for the uci go mate option was added.
The rescorer can now be built from the lc0 code base instead of a separate branch.
A dicrete onnx layernorm implementation was added to get around a onnxruntime bug with directml - this has some overhead so it is only enabled for onnx-dml and can be switched off with the alt_layernorm=false backend option.
The --onnx2pytoch option was added to leela2onnx to generate pytorch compatible models.
There is a cuda min_batch backend option to reduce non-determinism with small batches.
New options were added to onnx2leela to fix tf exported onnx models.
The onnx backend can now be built for amd's rocm.
Fixed a bug where the Contempt effect on eval was too low for nets with natively higher draw rates.
Made the WDL Rescale sharpness limit configurable via the --wdl-max-s hidden option.
The search task workers can be set automatically, to either 0 for cpu backends or up to 4 depending on the number of cpu cores. This is enabled by --task-workers=-1 (the new default).
Several assorted fixes and code cleanups.

By Frank Rahde (Mod.) Date 2024-03-26 19:22

Auf Chessify nutze ich schon seit Wochen LC0 v0.31: https://chessify.me/news/lc0-v031-engine-update

By Peter Martan Date 2024-03-26 19:27

Es geistern ja auch schon ewig lang verschiedene 0.31- Compiles durchs Netz, dafür finde ich ein jetzt endlich "offizielles Prerelease" (auch rc1 genannt) eigentlich mal wieder irgendwie typisch

By Lothar Jung Date 2024-03-26 19:43

Die 0.31 Version fasst alle bisherigen vorhandenen Abzweigungen zusammen und konstruiert diese.

By Peter Martan Date 2024-03-26 19:55 Edited 2024-03-26 19:58 Upvotes 2

Ja, so wie das die vorherigen vielen Versionen auch gemacht haben seit dem letzten tatsächlichen offiziellen Release, das immer noch das 0.30 ist:

https://lczero.org/play/download/

Beim jetzigen 0.31 handelt es sich nach wie vor nur um einen ersten release candidate.
Ob ich auf diese Weise wohl noch eine offizielle 1.0- Version erleben werde?

Von den verschiedenen Netzen, von denen auch keiner weiß, welches für welche Hardware- TC und welche Engine- Version das beste sein mag, ganz zu schweigen...

By Lothar Jung Date 2024-03-29 08:12 Upvotes 1

„konsolidiert“

By Jörg Oster Date 2024-03-29 11:19 Upvotes 1

Zitat:

* There is a vastly simplified lc0 interface available by renaming the executable to `lc0simple`.

* While it is easy to get a single node policy evaluation (`go nodes 1` using uci), there was no simple way to get the effect of a value only evaluation, so the `--value-only` option was added.

* Support for the uci `go mate` option was added.

Diese drei Änderungen finde ich interessant.

Das einfachere Interface, mit deutlich weniger Optionen, sieht bei mir so aus:

Code:

       _
|   _ | |
|_ |_ |_| v0.31.0-rc1+git.unknown built Mar 28 2024
uci
id name Lc0 v0.31.0-rc1+git.unknown
id author The LCZero Authors.
option name WeightsFile type string default <autodiscover>
option name Threads type spin default 0 min 0 max 128
option name MultiPV type spin default 1 min 1 max 500
option name Contempt type string default 
option name SyzygyPath type string default 
option name Ponder type check default true
option name UCI_Chess960 type check default false
option name UCI_ShowWDL type check default false
option name UCI_ShowMovesLeft type check default false
option name RamLimitMb type spin default 0 min 0 max 100000000
option name MoveOverheadMs type spin default 200 min 0 max 100000000
option name Slowmover type string default 1.000000
option name ValueOnly type check default false
option name LogFile type string default 
uciok

'go mate x' scheint auch zu funktionieren.
Hier ein einfaches Matt in 2:

position fen 2brrb2/8/p7/7Q/1p1kpPp1/1P1pN1K1/3P4/8 w - -
go mate 2
info depth 1 seldepth 2 time 5393 nodes 5 score cp 71 nps 3 tbhits 0 pv h5a5 d8d7
info depth 2 seldepth 3 time 6547 nodes 9 score cp 136 nps 3 tbhits 0 pv h5a5 d8d7 e3f5
info depth 3 seldepth 4 time 6644 nodes 22 score cp 1285 nps 8 tbhits 0 pv h5a5 e8e6 a5d8
info depth 3 seldepth 5 time 7047 nodes 31 score cp 1754 nps 10 tbhits 0 pv h5a5 e8e6 a5d8 e6d6 d8c8
info depth 3 seldepth 6 time 8118 nodes 40 score cp 2243 nps 9 tbhits 0 pv h5a5 e8e5 a5d8 d4c5
info depth 3 seldepth 6 time 8546 nodes 60 score mate 2 nps 13 tbhits 0 pv h5a5 d8d7 e3f5
bestmove h5a5 ponder d8d7

Seltsam ist nur, dass die 2. Zeile schon die PV der Lösung hat.
Warum da nicht schon früher das Matt erkannt wird, weiß ich nicht.

By Lothar Jung Date 2024-03-29 18:36 Upvotes 1

Das finde ich auch nützlich, besonders wenn man unterschiedliche Hardware und Netze einsetzt:

The backends can now suggest a minibatch size to the search, this is enabled by `--minibatch-size=0` (the new default).

By Max Siegfried Date 2024-03-29 22:30

Lothar Jung schrieb:

Muss man dazu den backendbench verwenden?
Oder reicht beim kompilieren: ./build.sh

By Lothar Jung Date 2024-03-30 07:49

Das hat doch nichts miteinander zu tun.
beckendbench wird zur Ermittlung der günstigsten mbs benutzt.
Der andere Befehl zur Compilierung.